\u6279\u5224（\u5bf9\u67b6\u6784\u5e08\u89c6\u89d2）：<\/p>\n
- \n
- \u8fd9\u662f\u7ecf\u9a8c\/\u7cfb\u7edf\u8bba\u6587，\u4e0d\u662f\u7406\u8bba\u8bba\u6587\u2014\u2014L0\/L1\/L2 \u662f\u5b9a\u6027\u7c7b\u6bd4，\u65e0\u6cdb\u5316\u754c；<\/li>\n
- \u5168\u7a0b\u65e0\u7f6e\u4fe1\u533a\u95f4，\u4e3b\u5b9e\u9a8c\u5355\u4e00\u7b56\u7565\u65cf（Claude Opus 4.8），\u8d85\u53c2\u5bf9 domain \u5f88\u654f\u611f；<\/li>\n
- critic \u53ea\u80fd\u62e6"\u5199\u6b7b\u4efb\u52a1\u540d\/\u7b54\u6848"\u8fd9\u7c7b\u663e\u5f0f\u6cc4\u6f0f，\u9690\u5f0f\u8fc7\u62df\u5408\u96be\u6839\u9664；<\/li>\n
- \u590d\u73b0\u95e8\u69db\u662f\u7b97\u529b：\u6bcf\u4e2a\u5019\u9009\u8981\u5728 evolve \u96c6\u4e0a\u8dd1\u591a\u6b21\u771f\u5b9e agent rollout\u3002<\/li>\n<\/ul>\n
\u5bf9\u4f60\u505a Agent \u5e73\u53f0\u7684\u8fc1\u79fb\u4ef7\u503c：\u9009\u62e9\u7aef\u56db\u9053\u95f8 \u2248 agent \u81ea\u6539\u8fdb\u7684\u63a7\u5236\u9762\/\u53d8\u66f4\u6cbb\u7406\u5c42（\u53ef\u5ba1\u8ba1\u3001\u975e\u8865\u507f\u51c6\u5165\u3001\u53ef\u56de\u6eda），\u548c Palantir \u5f0f\u51b3\u7b56\u95e8\u540c\u6784；\u53ef\u76f4\u63a5\u6284\u7684 pattern \u662f：\u4f59\u5f26\u9000\u706b\u7f16\u8f91\u9884\u7b97\u3001\u7ed9\u6bcf\u4e2a diff \u7ed3\u6784\u5316\u8bb0\u8d26 (\u7ec4\u4ef6,\u5047\u8bbe,\u0394S,\u0394C,\u662f\u5426\u91c7\u7eb3)\u3001\u7a97\u53e3\u5316\u65e0\u8d21\u732e\u7ec4\u4ef6\u526a\u679d\u3001\u6210\u672c-\u589e\u76ca\u7ebf\u6027\u95f8\u95e8 \u0394C \u2264 \u03b20+\u03b21\u00b7\u0394S\u3002<\/p>\n
\u5b8c\u6574\u8bba\u8bc1\u94fe\u3001\u6d88\u878d\u6570\u5b57\u3001\u5931\u6548\u6761\u4ef6\u548c\u590d\u73b0\u8def\u5f84\u90fd\u5728\u62a5\u544a\u91cc\u3002<\/p>\n
\nDeepThink \u662f\u4f60\u7684\u5f00\u6e90\u514d\u8d39\u81ea\u7531\u4f7f\u7528\u7684\u79c1\u6709 AI \u64cd\u4f5c\u7cfb\u7edf (AI OS)，\u5728\u5b89\u5168\u9694\u79bb\u7684\u6c99\u7bb1\u73af\u5883\u4e2d，\u81ea\u4e3b\u6267\u884c\u4ee3\u7801\u3001\u7ba1\u7406\u6587\u4ef6\u3001\u5b8c\u6210\u8d85\u590d\u6742\u957f\u7a0b\u4efb\u52a1\u3002\u81ea\u6258\u7ba1\u7684\u591a\u7528\u6237\u672c\u5730 AI Agent Loop Engineering \u7cfb\u7edf (\u652f\u6301\u684c\u9762\u7aef+\u6d4f\u89c8\u5668+\u79fb\u52a8\u7aef) \u2014\u2014 \u8ba9 DeepThink \u6210\u4e3a\u4f60\u7684\u5168\u80fd\u6570\u5b57\u52a9\u624b\u3002 \u2014\u2014 Powered By AI Genius Institute & \u5149\u5251AI<\/p>\n
![\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260930025715-6abc7a8b2cb0f.png\")
<\/p>\nDeepThink \u9879\u76ee\u5f00\u6e90\u4ee3\u7801（\u5982\u679c\u4f60\u89c9\u5f97\u597d\u73a9，\u5c31\u4e00\u8d77\u6765\u73a9， Star \u4e00\u4e0b）： Gitcode: https:\/\/gitcode.com\/AIGeniusInstitute\/deepthink Github: https:\/\/github.com\/AIGeniusInstitute\/deepthink<\/p>\n
- \n
- DeepThink v1.5.0 \u2014 Agent \u7fa4\u7ec4\u534f\u4f5c（Swarm）\u6267\u884c\u94fe\u8def\u8d2f\u901a<\/li>\n<\/ul>\n
\u628a v1.4.0 \u4ea4\u4ed8\u7684 Agent \u7fa4\u7ec4\u534f\u4f5c（Swarm）\u4ece"\u80fd\u770b"\u53d8\u6210"\u80fd\u8dd1"： \u7fa4\u7ec4\u6d88\u606f\u771f\u6b63\u9a71\u52a8\u56fe\u8c31\u6267\u884c，\u5e2d\u4f4d\u56de\u590d\u9010\u5b57\u6d41\u5f0f\u56de\u4f20\u5e76\u5f52\u5c5e\u5230\u5e2d\u4f4d， \u5e2d\u4f4d\u7ee7\u627f\u672c\u8f6e\u7684\u6280\u80fd \/ MCP \u5de5\u5177 \/ \u77e5\u8bc6\u5e93\u3002<\/p>\n
- \n
- \u7fa4\u7ec4\u6d88\u606f\u96f6\u6267\u884c \u2192 swarm \u5373 graph_definition，\u6d88\u606f\u89e6\u53d1\u56fe\u8c31\u8fd0\u884c<\/li>\n
- \u5e2d\u4f4d\u7ea7\u6d41\u5f0f\u8f93\u51fa（\u524d\u7aef\u4ece\u53ea\u8ba4 group_message_created \u6539\u4e3a group_message_delta）<\/li>\n
- \u5e2d\u4f4d\u6302\u8f7d\u80fd\u529b\u4e0e\u666e\u901a\u5bf9\u8bdd\u5bf9\u9f50（\u590d\u7528\u540c\u4e00 workspace \u6302\u8f7d store \/ selectedMounts）<\/li>\n
- \u6267\u884c\u6a21\u5f0f\u7ee7\u627f（\u4e0d\u518d\u9759\u9ed8\u56de\u843d container）+ \u51b7\u542f\u52a8\u4e0d\u518d\u541e\u6389\u65b0\u5de5\u4f5c\u533a\u9996\u6761\u6d88\u606f<\/li>\n
- \u56fe\u8c31\u5f15\u64ce\u4ec5\u65b0\u589e\u4e24\u4e2a\u53ef\u9009 Hook（onNodeSettled \/ onNodeStream）， \u672a\u63a5\u7ebf\u65f6\u4e3a no-op；\u65e0\u65b0\u589e\u5f15\u64ce \/ \u4e2d\u95f4\u4ef6 \/ \u4f9d\u8d56<\/li>\n<\/ul>\n
Schema v70 \u2192 v70（\u65e0\u53d8\u66f4） \u00b7 \u53d1\u5e03\u8bf4\u660e\u89c1 docs\/release_notes\/v1.5.0.md<\/p>\n
- \n
- DeepThink v1.4.0 \u2014 \u5206\u5e03\u5f0f\u80fd\u529b\u843d\u5730\u4e0e\u4f01\u4e1a\u534f\u4f5c\u5de5\u4f5c\u8d1f\u8f7d<\/li>\n<\/ul>\n
\u5c06 v1.3.0 \u6253\u4e0b\u7684 K8s \u4e91\u539f\u751f\u57fa\u5ea7\u771f\u6b63\u843d\u5230\u4e1a\u52a1\u9762：\u914d\u7f6e\u5168\u91cf\u8fc1 PostgreSQL\u3001 \u5de5\u4f5c\u533a\u6587\u4ef6\u5165 MinIO\/S3\u3001\u5206\u5e03\u5f0f agent-runner \u8c03\u5ea6\u95ed\u73af\u3001\u8de8 Pod \u6302\u8f7d\u4e0a\u4e0b\u6587\u4fee\u590d； \u5e76\u5728\u6b64\u4e4b\u4e0a\u4ea4\u4ed8\u56db\u5927\u4f01\u4e1a\u7ea7\u80fd\u529b\u2014\u2014\u6570\u5b57\u5458\u5de5\u534f\u4f5c\u5de5\u4f5c\u53f0\u3001AgentNet \u7f51\u76d8\u3001 \u8bc4\u6d4b\u4e2d\u5fc3\u3001Agent \u7fa4\u7ec4\u534f\u4f5c（Swarm）\u3002<\/p>\n
Schema v60 \u2192 v70 \u00b7 \u53d1\u5e03\u8bf4\u660e\u89c1 docs\/release_notes\/v1.4.0.md<\/p>\n
- \n
- DeepThink v1.3.0 \u2014 \u4e91\u539f\u751f\u4e0e\u534f\u4f5c<\/li>\n<\/ul>\n
v1.3.0 \u662f\u4e00\u4e2a\u4e91\u539f\u751f\u4e0e\u534f\u4f5c\u7248\u672c：\u4ece\u5355\u673a\u591a\u667a\u80fd\u4f53\u7f16\u6392\u8dc3\u8fc1\u5230 K8s \u5168\u91cf\u65e0\u72b6\u6001\u5316 + \u6a2a\u5411\u591a Pod \u6269\u7f29\u5bb9 + \u591a\u4eba\u534f\u4f5c\u5de5\u4f5c + \u5e73\u53f0\u80fd\u529b\u6df1\u5316\u3002<\/p>\n
\u6838\u5fc3\u4ea4\u4ed8：<\/p>\n
- \n
- \u4e91\u539f\u751f：K8s \u4e91\u7aef\u90e8\u7f72 \u2192 \u6a2a\u5411\u591a Pod \u65e0\u72b6\u6001\u5316（Redis \u4e8b\u4ef6\u603b\u7ebf + \u5206\u5e03\u5f0f\u9009\u4e3b + PostgreSQL \u540c\u6b65\u6865）\u2192 Agent IPC Redis + Agent Runner \u72ec\u7acb Service \u2192 Phase 3 \u751f\u4ea7\u5316（pgvector\/\u5bf9\u8c61\u5b58\u50a8\/Litestream \u707e\u5907） \u2192 \u5168\u91cf\u65e0\u72b6\u6001\u5316\u7f3a\u53e3\u5ba1\u8ba1\u95ed\u73af + \u5171\u4eab\u5e76\u53d1\u8ba1\u6570\u5668<\/li>\n
- \u591a\u4eba\u534f\u4f5c\u5de5\u4f5c\u80fd\u529b（\u7f16\u6392\u8005-\u5de5\u4f5c\u8005 \/ \u5bf9\u7b49 \/ \u6279\u8bc4\u5bf9\u6297 3 \u6a21\u5f0f + \u7fa4\u5171\u4eab\u5de5\u4f5c\u533a）<\/li>\n
- \u5e73\u53f0\u80fd\u529b\u6df1\u5316（\u5168\u8fc7\u7a0b Trace \/ \u6821\u9a8c\u8282\u70b9\u4e0e Hooks \/ \u6d4b\u8bd5\u8bc4\u6d4b \/ Skills \u7248\u672c\u7ba1\u7406）<\/li>\n
- \u4f01\u4e1a\u7ea7\u5de5\u5177\u6cbb\u7406\u4e0e\u6700\u5c0f\u6743\u9650\u57fa\u7ebf（\u526f\u4f5c\u7528\u5206\u7ea7 \/ \u5e42\u7b49\u952e \/ \u5ba1\u8ba1 \/ \u9650\u6d41 \/ \u51ed\u636e\u52a0\u5bc6）<\/li>\n
- PG INTEGER\u2192BIGINT \u65f6\u95f4\u6233\u6ea2\u51fa\u4fee\u590d + \u672c\u673a\u6301\u4e45\u5316\u5b58\u50a8\u6808<\/li>\n
- \u4e00\u7ea7\u5bfc\u822a\u7b80\u5316\u81f3 7 \u9879<\/li>\n<\/ul>\n
\u53d1\u5e03\u8bf4\u660e：docs\/release_notes\/v1.3.0.md<\/p>\n<\/p>\n
\u6587\u7ae0\u76ee\u5f55<\/h4>\n
- \n
- \u300aRRSI：\u667a\u80fd\u4f53 Harness \u7684\u6b63\u5219\u5316\u9012\u5f52\u81ea\u6211\u6539\u8fdb\u300b\u6df1\u5ea6\u7cbe\u8bfb<\/li>\n
- \n
- \n
- 0. \u4e00\u53e5\u8bdd\u672c\u8d28<\/li>\n
- 1. \u7ed3\u8bba\u901f\u89c8（\u5148\u7ed9\u5224\u65ad）<\/li>\n
- 2. \u7814\u7a76\u6545\u4e8b：\u5b83\u5230\u5e95\u5728\u89e3\u51b3\u4ec0\u4e48\u95ee\u9898<\/li>\n
- 3. \u65b9\u6cd5\u673a\u5236：\u4e24\u7aef\u516d\u9053\u95f8<\/li>\n
- 4. \u5173\u952e\u8bc1\u636e：\u6570\u5b57\u80cc\u540e\u771f\u6b63\u5728\u8bf4\u4ec0\u4e48<\/li>\n
- \n
- \n
- 4.1 \u6838\u5fc3\u4ea4\u6362：\u4e3b\u52a8\u727a\u7272 evolve \u5206，\u4e70 OOD（\u8868 1）<\/li>\n
- 4.2 \u6d88\u878d：\u6b63\u5219\u5316\u786e\u5b9e\u5728\u505a "\u53cd\u76f4\u89c9" \u7684\u4e8b（\u8868 2）<\/li>\n
- 4.3 \u53ef\u8fc1\u79fb\u6027：\u673a\u5236\u4e0d\u5c5e\u4e8e\u67d0\u4e2a\u6a21\u578b（\u8868 3 \/ \u8868 4）<\/li>\n
- 4.4 \u6210\u672c（\u56fe 4）<\/li>\n<\/ul>\n<\/li>\n
- 5. \u53cd\u76f4\u89c9\u7ed3\u679c\u4e0e\u5931\u6548\u6761\u4ef6（\u4e3b\u52a8\u627e\u95ee\u9898）<\/li>\n
- 6. \u53ef\u4fe1\u8fb9\u754c\u3001\u5c40\u9650\u4e0e\u590d\u73b0\u98ce\u9669<\/li>\n
- 7. \u5bf9\u4f60（AI Agent \u67b6\u6784\u89c6\u89d2）\u7684\u53ef\u8fc1\u79fb\u6d1e\u5bdf<\/li>\n
- 8. \u9605\u8bfb\u8303\u56f4\u4e0e\u4fe1\u606f\u8fb9\u754c<\/li>\n<\/ul>\n<\/li>\n
- RRSI：\u667a\u80fd\u4f53 Harness \u7684\u6b63\u5219\u5316\u9012\u5f52\u81ea\u6211\u6539\u8fdb<\/li>\n
- \n
- \n
- \u6458\u8981<\/li>\n
- 1. \u5f15\u8a00<\/li>\n
- 2. \u9884\u5907\u77e5\u8bc6<\/li>\n
- 3. RRSI<\/li>\n
- \n
- \n
- 3.1\u3000Harness \u8fdb\u5316\u7684\u6b63\u5219\u5316\u89c6\u89d2<\/li>\n
- 3.2\u3000\u6b63\u5219\u5316\u63d0\u51fa\u5206\u5e03<\/li>\n
- 3.3\u3000\u6b63\u5219\u5316\u5019\u9009\u9009\u62e9<\/li>\n<\/ul>\n<\/li>\n
- 4. \u5b9e\u9a8c<\/li>\n
- \n
- \n
- 4.1\u3000\u5b9e\u9a8c\u8bbe\u7f6e<\/li>\n
- 4.2\u3000\u4e3b\u8981\u7ed3\u679c<\/li>\n
- 4.3\u3000\u5206\u6790<\/li>\n<\/ul>\n<\/li>\n
- 5. \u76f8\u5173\u5de5\u4f5c<\/li>\n
- 6. \u7ed3\u8bba<\/li>\n
- \n
- \n
- \u5c40\u9650\u6027<\/li>\n<\/ul>\n<\/li>\n
- \u53c2\u8003\u6587\u732e<\/li>\n<\/ul>\n<\/li>\n
- \u9644\u5f55<\/li>\n
- \n
- \n
- A. \u8bc4\u4f30<\/li>\n
- B. \u57fa\u7ebf\u65b9\u6cd5<\/li>\n
- C. \u65b9\u6cd5\u7ec6\u8282<\/li>\n
- D. \u5b9e\u9a8c<\/li>\n
- E. \u5b9a\u6027\u6848\u4f8b\u7814\u7a76<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n
\u300aRRSI：\u667a\u80fd\u4f53 Harness \u7684\u6b63\u5219\u5316\u9012\u5f52\u81ea\u6211\u6539\u8fdb\u300b\u6df1\u5ea6\u7cbe\u8bfb<\/h2>\n
\u539f\u6587：<\/p>\n
RRSI: Regularized Recursive Self-Improvement of Agent Harnesses<\/p>\n
，Peng Xia \u7b49，Google Cloud AI Research \/ UNC \/ Stanford \/ WUSTL，arXiv:2609.24972v1，2026-09-22\u3002 \u4ee3\u7801：<\/p>\n
github.com\/google-research\/rrsi<\/p>\n
\u30fb\u3000\u4e3b\u9875：<\/p>\n
regularized-rsi.com<\/p>\n
\n0. \u4e00\u53e5\u8bdd\u672c\u8d28<\/h3>\n
RRSI \u628a\u7ecf\u5178\u673a\u5668\u5b66\u4e60\u91cc \u201c\u8bad\u7ec3 \/ \u6d4b\u8bd5\u5206\u79bb + \u6b63\u5219\u5316\u201d \u8fd9\u5957\u65b9\u6cd5\u8bba，\u7b2c\u4e00\u6b21\u7cfb\u7edf\u5730\u642c\u5230\u4e86 \u201charness \u8fdb\u5316\u201d \u8fd9\u4e2a\u5916\u5c42\u4f18\u5316\u5faa\u73af\u4e0a\u3002 \u5b83\u6307\u51fa：harness \u8fdb\u5316\u4e0d\u662f\u5728 \u201c\u9020\u66f4\u597d\u7684\u667a\u80fd\u4f53\u201d，\u800c\u662f\u5728\u4e00\u4e2a\u9ad8\u8868\u8fbe\u529b\u641c\u7d22\u7a7a\u95f4\u91cc\u3001\u5bf9\u4e00\u4e2a\u6709\u9650\u7684 evolve \u96c6\u53cd\u590d\u505a\u7ecf\u9a8c\u4f18\u5316\u2014\u2014 \u8fd9\u4e2a\u52a8\u4f5c\u672c\u8eab\u5c31\u4f1a\u8fc7\u62df\u5408；\u4e8e\u662f\u5b83\u4e0d\u9650\u5236 \u201charness \u80fd\u6539\u4ec0\u4e48\u201d，\u53ea\u5728\u63d0\u51fa\u7aef\u600e\u4e48\u82b1\u9884\u7b97\u548c\u9009\u62e9\u7aef\u4ec0\u4e48\u6539\u52a8\u80fd\u6c38\u4e45\u4fdd\u7559\u4e24\u5904\u52a0\u6b63\u5219，\u7528\u4e3b\u52a8\u727a\u7272\u5c11\u91cf evolve \u96c6\u5206\u6570，\u6362\u6765\u771f\u5b9e\u7684\u5206\u5e03\u5916\u8fc1\u79fb\u548c\u66f4\u4f4e\u7684 token \u6210\u672c\u3002<\/p>\n
\n1. \u7ed3\u8bba\u901f\u89c8（\u5148\u7ed9\u5224\u65ad）<\/h3>\n
\n
\u7ef4\u5ea6\u5224\u65ad<\/tr>\n \n\n \u771f\u6b63\u8d21\u732e<\/td>\n \u4e0d\u662f\u65b0\u6a21\u578b \/ \u65b0\u6570\u636e，\u800c\u662f\u8bca\u65ad + \u4e00\u5957\u8fdb\u5316\u5916\u5c42\u6cbb\u7406\u89c4\u5219：\u628a harness \u8fc7\u62df\u5408\u5f62\u5f0f\u5316，\u5e76\u7ed9\u51fa\u53ef\u5de5\u7a0b\u5316\u7684 propose\/select \u6b63\u5219\u5316\u5668<\/td>\n<\/tr>\n \n \u6700\u53ef\u4fe1\u53d1\u73b0<\/td>\n \u6b63\u5219\u5316\u4e0d\u4ec5\u6ca1\u727a\u7272\u6cdb\u5316，\u53cd\u800c\u4e3b\u52a8\u4e70\u5230\u4e86\u6cdb\u5316：RRSI \u5728 evolve \u96c6\u4e0a\u53ea\u6da8 1.1 \u5206（\u6240\u6709\u88ab\u8fdb\u5316 harness \u91cc\u6700\u5c0f），OOD \u5747\u503c\u5374\u4ece 39.7 \u5347\u5230 43.6（\u552f\u4e00\u663e\u8457\u8d85\u8fc7 𝐻\u2080 \u4e00\u5206\u4ee5\u4e0a\u7684），\u540c\u65f6 token \u4ece 3.80M \u964d\u5230 2.42M<\/td>\n<\/tr>\n \n \u6700\u5f3a\u8bc1\u636e<\/td>\n \u8de8\u9aa8\u5e72\u8fc1\u79fb（\u8868 3\/4）：\u7528 Gemini 3.5 Flash \u8fdb\u5316\u51fa\u7684 harness，\u539f\u5c01\u4e0d\u52a8\u62ff\u5230\u4ece\u672a\u53c2\u4e0e\u641c\u7d22\u7684 Gemini 3.1 Flash Lite \u4e0a，\u4ecd\u4ece 11.2\u219214.6\u3002\u8fd9\u8bc1\u660e\u5b66\u5230\u7684\u662f \u201c\u53ef\u590d\u7528\u673a\u5236\u201d \u800c\u975e \u201c\u5bf9\u67d0\u4e2a\u7b56\u7565\u7684\u62df\u5408\u201d<\/td>\n<\/tr>\n \n \u8bba\u6587\u7c7b\u578b<\/td>\n \u7ecf\u9a8c\u6027 \/ \u7cfb\u7edf\u8bba\u6587（empirical），\u4e0d\u662f\u7406\u8bba\u8bba\u6587\u2014\u2014L0\/L1\/L2 \u662f\u5b9a\u6027\u7c7b\u6bd4，\u4f5c\u8005\u81ea\u5df1\u660e\u786e\u627f\u8ba4\u65e0\u6cdb\u5316\u754c\u3001\u65e0 regret bound<\/td>\n<\/tr>\n \n \u6700\u5927\u8f6f\u808b<\/td>\n \u5168\u7a0b\u65e0\u7f6e\u4fe1\u533a\u95f4 \/ \u663e\u8457\u6027\u68c0\u9a8c；\u4e3b\u5b9e\u9a8c\u5355\u4e00\u7b56\u7565\u65cf（Claude Opus 4.8）；\u8d85\u53c2\u5bf9 domain \u654f\u611f；\u201c\u9690\u5f0f\u6cc4\u6f0f\u201d \u80fd\u5426\u88ab critic \u6839\u9664\u5b58\u7591<\/td>\n<\/tr>\n \n \u503c\u4e0d\u503c\u5f97\u8ddf\u8fdb<\/td>\n \u503c\u5f97\u4f5c\u4e3a baseline \u590d\u73b0（\u4ee3\u7801\u5f00\u6e90），\u4f46\u641c\u7d22\u6210\u672c\u6781\u9ad8（\u6bcf\u4e2a\u5019\u9009\u8981\u5728 evolve \u96c6\u4e0a\u8dd1\u591a\u6b21 agent rollout）<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
\n2. \u7814\u7a76\u6545\u4e8b：\u5b83\u5230\u5e95\u5728\u89e3\u51b3\u4ec0\u4e48\u95ee\u9898<\/h3>\n
\u628a\u8bba\u6587\u4ece \u201c\u9010\u8282\u6458\u8981\u201d \u8fd8\u539f\u6210\u4e00\u6761\u8bba\u8bc1\u94fe：<\/p>\n
\u73b0\u4ee3 LLM \u667a\u80fd\u4f53 = \u51bb\u7ed3\u9aa8\u5e72 + Harness（\u63d0\u793a\u8bcd\/\u63a7\u5236\u6d41\/\u5de5\u5177\/\u8bb0\u5fc6\/\u4e0a\u4e0b\u6587\u7ba1\u7406）<\/p>\n
\u2502<\/p>\n
\u25bc<\/p>\n
\u8fd1\u671f\u8fdb\u5c55\u5927\u591a\u6765\u81ea Harness \u5de5\u7a0b，\u800c\u975e\u65b0\u6743\u91cd；\u4f46\u8fd9\u9760\u4eba\u5de5\u8bfb\u5931\u8d25\u8f68\u8ff9，\u5929\u82b1\u677f\u4f4e<\/p>\n
\u2502<\/p>\n
\u25bc<\/p>\n
\u4e8e\u662f\u5927\u5bb6\u7528 LLM \u81ea\u52a8\u8fdb\u5316 harness：proposer \u6539 \u2192 \u5728 evolve \u96c6\u4e0a\u6253\u5206 \u2192 \u5206\u9ad8\u5c31\u7559\u4e0b<\/p>\n
\u2502<\/p>\n
\u25bc \u3010\u8bba\u6587\u7684\u6838\u5fc3\u89c2\u5bdf\u3011<\/p>\n
\u8fd9\u4e2a\u5faa\u73af\u5728\u7edf\u8ba1\u4e0a\u6709\u4e2a\u81f4\u547d\u6027\u8d28：evolve \u96c6\u88ab\u8de8\u8f6e\u6b21\u53cd\u590d\u590d\u7528\u3002<\/p>\n
\u7b2c t \u8f6e\u63d0\u51fa\u7684\u5019\u9009，\u4f9d\u8d56\u7b2c t-1 \u8f6e\u5728\u540c\u4e00\u6279\u4efb\u52a1\u4e0a\u6d4b\u5230\u7684\u7ed3\u679c\u3002<\/p>\n
\u2500\u2500 \u8fd9\u5c31\u662f Dwork et al. (2015) \u7684 "adaptive data analysis \/ holdout reuse"：<\/p>\n
\u6bcf\u591a\u770b\u4e00\u773c\u6709\u9650\u6570\u636e，\u6709\u6548\u6837\u672c\u91cf\u5c31\u88ab"\u6d88\u8017"\u4e00\u6b21\u3002<\/p>\n
\u2502<\/p>\n
\u25bc<\/p>\n
\u540e\u679c（\u56fe 1a \u5b9e\u8bc1）：evolve \u96c6\u5206\u6570\u865a\u9ad8，\u4f46 OOD \u4e0d\u6da8\u751a\u81f3\u4f4e\u4e8e\u521d\u59cb harness 𝐻\u2080\u3002<\/p>\n
\u56db\u79cd\u6b64\u524d\u65b9\u6cd5\u91cc，\u6700\u5f3a\u7684 Meta-Harness \u5728 OOD \u5747\u503c\u4e0a\u53ea +0.9，TTHE \u53cd\u800c -1.7\u3002<\/p>\n
"gain does not transfer" \u2014\u2014 \u63d0\u5347\u662f\u62df\u5408，\u4e0d\u662f\u80fd\u529b\u3002<\/p>\n
\u2502<\/p>\n
\u25bc<\/p>\n
\u3010\u6d1e\u5bdf\u3011\u8fc7\u62df\u5408\u4e0d\u662f\u6765\u81ea\u67d0\u4e2a\u7ec4\u4ef6\u5199\u9519，\u800c\u662f\u4e09\u79cd\u8026\u5408\u884c\u4e3a：<\/p>\n
\u2460 benchmark-specific fitting（\u7f16\u7801\u4efb\u52a1\/\u5b9e\u4f53\/\u7b54\u6848\u8fdb\u4e86\u63d0\u793a\u8bcd）<\/p>\n
\u2461 noise chasing（\u628a\u968f\u673a\u6ce2\u52a8\u5f53\u4fe1\u53f7）<\/p>\n
\u2462 complexity accumulation（\u5806\u65e0\u7528\u590d\u6742\u5ea6\u6362 evolve \u5206）<\/p>\n
\u2502<\/p>\n
\u25bc<\/p>\n
\u3010\u65b9\u6cd5\u3011RRSI \u4e0d\u9650\u5236\u7f16\u8f91\u7a7a\u95f4 \u03a9(H)（\u63d0\u793a\u8bcd\/\u63a7\u5236\u6d41\/\u5de5\u5177\/\u8bb0\u5fc6\/\u5b50\u667a\u80fd\u4f53\u5168\u5f00），<\/p>\n
\u53ea\u5bf9"\u641c\u7d22\u8f68\u8ff9"\u52a0\u6b63\u5219：\u63d0\u51fa\u7aef\u63a7\u9884\u7b97，\u9009\u62e9\u7aef\u63a7\u51c6\u5165\u3002<\/p>\n
\u5173\u952e\u6d1e\u5bdf（\u6700\u8be5\u5e26\u8d70\u7684\u4e00\u53e5）：harness \u8fdb\u5316\u7684\u8fc7\u62df\u5408，\u548c\u795e\u7ecf\u7f51\u7edc\u5728\u5c0f\u6570\u636e\u96c6\u4e0a\u8fc7\u62df\u5408\u662f\u540c\u6784\u95ee\u9898\u2014\u2014 \u533a\u522b\u53ea\u5728\u4e8e，\u8fd9\u91cc\u7684 \u201c\u53c2\u6570\u201d \u662f\u4e00\u6bb5\u53ef\u7f16\u8f91\u7684\u7a0b\u5e8f（harness），\u201c\u8bad\u7ec3\u96c6\u201d \u662f\u4e00\u4e2a evolve benchmark\u3002\u6240\u4ee5\u7ecf\u5178 ML \u7684\u6b63\u5219\u5316\u6b66\u5668\u5e93\u53ef\u4ee5\u6574\u4f53\u5e73\u79fb\u8fc7\u6765\u3002<\/p>\n
\n3. \u65b9\u6cd5\u673a\u5236：\u4e24\u7aef\u516d\u9053\u95f8<\/h3>\n
RRSI \u7684\u7cbe\u5999\u5728\u4e8e\u628a\u6b63\u5219\u5316\u62c6\u5230\u8fdb\u5316\u5faa\u73af\u7684\u4e24\u7aef，\u800c\u4e0d\u662f\u585e\u8fdb harness \u5185\u90e8\u3002<\/p>\n
\u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500 \u63d0\u51fa\u7aef Proposal（\u600e\u4e48\u82b1\u641c\u7d22\u9884\u7b97）\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510<\/p>\n
\u2502 A \u9000\u706b\u66f4\u65b0\u7a00\u758f\u6027 b\\\\_t \u4f59\u5f26\u9000\u706b：\u65e9\u671f\u5141\u8bb8\u6346\u7ed1\u591a\u6539\u52a8，\u540e\u671f\u5f3a\u5236\u7a00\u758f\u53ef\u5f52\u56e0 (\u2248L0)\u2502<\/p>\n
\u2502 B \u8bc1\u636e\u611f\u77e5\u4fe1\u7528\u5206\u914d：\u6bcf\u4e2a diff \u7ed1\u5b9a \u7ec4\u4ef6+\u5047\u8bbe+\u0394S+\u0394C+\u662f\u5426\u88ab\u63a5\u53d7； \u2502<\/p>\n
\u2502 \u5df2\u8bc1\u4f2a\u7684\u5047\u8bbe\u4e0d\u518d\u91cd\u590d\u70e7\u9884\u7b97 \u2502<\/p>\n
\u2502 C \u7ed3\u6784\u5316\u63a2\u7d22：\u8fdb\u5c55\u505c\u6ede(\u8fde\u7eed w \u8f6e\u843d\u5728\u566a\u58f0\u5e26 \u03b4 \u5185)\u65f6， \u2502<\/p>\n
\u2502 \u628a\u9884\u7b97\u62e8\u7ed9\u4ece\u672a\u88ab\u8bd5\u8fc7\u7684\u7ec4\u4ef6\u7c7b\u578b (\u2248\u71b5\/\u591a\u6837\u6027\u6b63\u5219) \u2502<\/p>\n
\u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518<\/p>\n
\u25bc \u5019\u9009 H'<\/p>\n
\u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500 \u9009\u62e9\u7aef Selection（\u4ec0\u4e48\u80fd\u6210\u4e3a\u6c38\u4e45\u72b6\u6001）\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510<\/p>\n
\u2502 D \u6cc4\u6f0f\u7b5b\u67e5 critic：\u5b8c\u6574\u8bc4\u4f30\u524d，\u62d2\u7edd\u5199\u6b7b\u4efb\u52a1\u540d\/\u5b9e\u4f53\/\u7b54\u6848\/\u7279\u5b9a\u903b\u8f91 \u2502<\/p>\n
\u2502 E \u566a\u58f0\u8c03\u6574\u5e95\u7ebf：\u015c(H') \u2265 S\u2605 \u2212 \u03b4（\u4e0d\u8bb8\u6cbf\u4e00\u4e32\u5c0f\u9000\u6b65\u4e0b\u5761\u8d70） \u2502<\/p>\n
\u2502 F \u590d\u6742\u5ea6\u611f\u77e5\u63a5\u53d7：\u0394C \u2264 \u03b20 + \u03b21\u00b7\u0394S（\u6210\u672c\u5fc5\u987b\u88ab\u589e\u76ca\u80cc\u4e66，\u2248Ridge L2） \u2502<\/p>\n
\u2502 G \u7ed3\u6784\u5316\u526a\u679d：n\\\\_prune \u7a97\u53e3\u5185\u65e0\u6b63\u8d21\u732e\u7684\u7ec4\u4ef6，\u6807\u8bb0\u4e3a\u5220\u9664\u76ee\u6807 (\u2248Lasso L1)\u2502<\/p>\n
\u2502 + \u9886\u57df\u5b88\u536b：\u5de5\u7a0b\u8bbe\u8ba1\u91cc valid-output \u66b4\u8dcc\/\u65e0\u63d0\u4ea4\u7387\u98d9\u5347\u5219\u4e00\u7968\u5426\u51b3 \u2502<\/p>\n
\u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518<\/p>\n
\u4e3a\u4ec0\u4e48\u662f \u201c\u975e\u8865\u507f\u6027 (non-compensatory)\u201d \u56db\u9053\u95f8？ \u8fd9\u662f\u8bbe\u8ba1\u7684\u7075\u9b42：\u4e00\u4e2a\u5019\u9009\u5148\u8fc7 D（\u65e0\u6cc4\u6f0f）\u2192 \u518d\u8fc7 E（\u8fc7\u566a\u58f0\u5e95\u7ebf）\u2192 \u518d\u8fc7 F（\u6210\u672c\u5408\u7406）\u2192 \u518d\u8fc7 G \/ \u9886\u57df\u5b88\u536b，\u4efb\u4f55\u4e00\u9053\u4e0d\u8fc7\u5c31\u76f4\u63a5\u6dd8\u6c70，\u4e0d\u5141\u8bb8\u7528\u4e00\u65b9\u9762\u7684\u9ad8\u5206\u8865\u507f\u53e6\u4e00\u65b9\u9762\u7684\u4e0d\u53ca\u683c\u3002\u8fd9\u4ece\u7ed3\u6784\u4e0a\u675c\u7edd\u4e86 \u201c\u9760\u5806\u590d\u6742\u5ea6\u5728 evolve \u96c6\u591a\u62ff\u51e0\u5206\u3001\u5374\u628a\u6cc4\u6f0f \/ \u9000\u6b65 \/ \u9ad8\u6210\u672c\u85cf\u8d77\u6765\u201d\u3002<\/p>\n
\u6bcf\u4e2a\u6a21\u5757\u89e3\u51b3\u4ec0\u4e48\u56f0\u96be（\u5220\u6389\u4f1a\u600e\u6837）：<\/p>\n
- \n
- \n
A \u9000\u706b\u9884\u7b97：\u4e0d\u505a\u7684\u8bdd，proposer \u4f1a\u628a\u4e00\u5806\u65e0\u5173\u6539\u52a8\u6346\u8fdb\u4e00\u4e2a\u5019\u9009，\u6709\u6548\u5bb9\u91cf\u7206\u70b8，\u4e14\u4efb\u4f55\u6d4b\u5f97\u53d8\u5316\u90fd\u65e0\u6cd5\u5f52\u56e0\u3002\u5220\u6389\u540e = \u8868 2 \u91cc \u201c\u672a\u6b63\u5219\u5316\u8fdb\u5316\u201d\u3002<\/p>\n<\/li>\n
- \n
E \u566a\u58f0\u5e95\u7ebf \u03b4：\u4e0d\u505a\u7684\u8bdd，\u9009\u62e9\u5668\u4f1a\u628a\u968f\u673a\u8d62\u5bb6\u56fa\u5316\u6210\u6c38\u4e45\u72b6\u6001 \u2014\u2014 \u8fd9\u6b63\u662f \u201c\u5728\u540c\u4e00 evolve \u96c6\u4e0a\u53cd\u590d\u8bc4\u4f30\u201d \u7684\u7edf\u8ba1\u4ee3\u4ef7\u3002<\/p>\n<\/li>\n
- \n
F \u6210\u672c – \u589e\u76ca\u6743\u8861 (\u03b20+\u03b21\u0394S)：\u4e0d\u505a\u7684\u8bdd，\u65e0\u7ea6\u675f\u7684 \u201c\u89e3\u7684\u6574\u4f53\u5e45\u5ea6\u201d \u4f1a\u81a8\u80c0（\u5bf9\u5e94 harness \u7684 token \u5360\u7528）\u3002\u8fd9\u5c31\u662f\u4e3a\u4ec0\u4e48 RRSI \u53cd\u800c\u66f4\u4fbf\u5b9c\u3002<\/p>\n<\/li>\n
- \n
G \u7ed3\u6784\u5316\u526a\u679d：\u4e0d\u505a\u7684\u8bdd，\u53ea\u589e\u4e0d\u51cf，harness \u8d8a\u6765\u8d8a\u81c3\u80bf；L1 \u7684\u5bf9\u5e94\u7269\u5c31\u662f \u201c\u65e0\u8d21\u732e\u7ec4\u4ef6\u6574\u4f53\u5220\u9664\u201d\u3002<\/p>\n<\/li>\n<\/ul>\n
\n4. \u5173\u952e\u8bc1\u636e：\u6570\u5b57\u80cc\u540e\u771f\u6b63\u5728\u8bf4\u4ec0\u4e48<\/h3>\n
4.1 \u6838\u5fc3\u4ea4\u6362：\u4e3b\u52a8\u727a\u7272 evolve \u5206，\u4e70 OOD（\u8868 1）<\/h4>\n
\n
\u65b9\u6cd5Harvey LAB EvolveOOD \u5747\u503c (JobBench\/GDPval\/APEX)<\/tr>\n \n\n 𝐻\u2080（\u4e0d\u8fdb\u5316）<\/td>\n 89.4<\/td>\n 39.7<\/td>\n<\/tr>\n \n Meta-Harness（evolve \u5206\u6700\u9ad8）<\/td>\n 93.0<\/td>\n 40.6（\u4ec5 +0.9）<\/td>\n<\/tr>\n \n AHE<\/td>\n 90.7<\/td>\n 38.8（\u22120.9）<\/td>\n<\/tr>\n \n TTHE<\/td>\n 91.1<\/td>\n 37.9（\u22121.7，\u4f4e\u4e8e\u8d77\u70b9）<\/td>\n<\/tr>\n \n HarnessX<\/td>\n 91.8<\/td>\n 39.0（\u2248\u56de\u5230\u8d77\u70b9）<\/td>\n<\/tr>\n \n RRSI<\/td>\n 90.5（evolve \u6da8\u5f97\u6700\u5c11）<\/td>\n 43.6（+3.9，\u552f\u4e00\u771f\u6b63\u8fc1\u79fb）<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n \u89e3\u8bfb<\/p>\n
：\u8fd9\u5f20\u8868\u662f\u5168\u6587\u7684\u9898\u773c\u3002\u6240\u6709 prior \u65b9\u6cd5\u90fd\u5728 evolve \u96c6\u4e0a \u201c\u8003\u5f97\u597d\u201d，\u4f46\u4e00\u5230 OOD \u5c31\u9732\u9985 \u2014\u2014 \u6709\u7684\u539f\u5730\u8e0f\u6b65，\u6709\u7684\u6bd4\u4e0d\u8fdb\u5316\u8fd8\u5dee\u3002RRSI \u6545\u610f\u628a evolve \u5206\u538b\u5230\u6700\u4f4e，\u5374\u6362\u6765\u4e86\u552f\u4e00\u53ef\u8fc1\u79fb\u7684\u63d0\u5347\u3002<\/p>\n
\u8fd9\u6b63\u662f\u6b63\u5219\u5316\u5728\u76d1\u7763\u5b66\u4e60\u91cc\u7684\u6807\u51c6\u6545\u4e8b：\u653e\u5f03\u4e00\u90e8\u5206\u8bad\u7ec3\u62df\u5408，\u6362\u6d4b\u8bd5\u96c6\u6cdb\u5316\u3002<\/p>\n
4.2 \u6d88\u878d：\u6b63\u5219\u5316\u786e\u5b9e\u5728\u505a \u201c\u53cd\u76f4\u89c9\u201d \u7684\u4e8b（\u8868 2）<\/h4>\n
\n
\u53d8\u4f53EvolveOOD \u5747\u503cToken \/ \u8bd5\u9a8c<\/tr>\n \n\n \u672a\u6b63\u5219\u5316\u8fdb\u5316<\/td>\n 92.8（\u6700\u9ad8）<\/td>\n 40.3（\u2248𝐻\u2080）<\/td>\n 3.80M（\u6700\u8d35）<\/td>\n<\/tr>\n \n \u53bb\u6389\u63a5\u53d7\u7aef\u6b63\u5219<\/td>\n 91.5<\/td>\n 41.0<\/td>\n 3.59M<\/td>\n<\/tr>\n \n \u53bb\u6389\u63d0\u51fa\u7aef\u6b63\u5219<\/td>\n 90.7<\/td>\n 41.9<\/td>\n 2.69M<\/td>\n<\/tr>\n \n \u5b8c\u6574 RRSI<\/td>\n 90.5<\/td>\n 43.6<\/td>\n 2.42M（\u6700\u4fbf\u5b9c）<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n \u89e3\u8bfb<\/p>\n
：\u53bb\u6389\u4efb\u4f55\u4e00\u7ec4\u6b63\u5219，evolve \u5206\u90fd\u6da8\u3001OOD \u90fd\u8dcc\u3001token \u90fd\u6da8\u3002\u4e24\u7aef\u90fd\u53bb\u6389\u65f6，evolve \u5206\u51b2\u5230\u5168\u573a\u6700\u9ad8 92.8，OOD \u5374\u51e0\u4e4e\u56de\u5230\u539f\u70b9 \u2014\u2014<\/p>\n
\u94c1\u8bc1：evolve \u5206\u548c\u6cdb\u5316\u5728\u8fd9\u91cc\u662f\u5bf9\u7acb\u6307\u6807<\/p>\n
\u3002<\/p>\n
4.3 \u53ef\u8fc1\u79fb\u6027：\u673a\u5236\u4e0d\u5c5e\u4e8e\u67d0\u4e2a\u6a21\u578b（\u8868 3 \/ \u8868 4）<\/h4>\n
- \n
- \n
\u6362\u7b56\u7565：Claude Opus 4.8 \u4e0e Gemini 3.5 Flash \u5404\u81ea\u72ec\u7acb\u8fdb\u5316，OOD \u90fd\u6da8（+1.8 \/ +2.2）\u3002<\/p>\n<\/li>\n
- \n
\u6362\u4ece\u672a\u53c2\u4e0e\u641c\u7d22\u7684\u66f4\u5f31\u9aa8\u5e72：Gemini 3.5 Flash \u8fdb\u5316\u7684 harness，\u62ff\u5230 Gemini 3.1 Flash Lite \u4e0a，11.2\u219214.6（\u76f8\u5bf9 +30.4%）\u3002<\/p>\n<\/li>\n<\/ul>\n
\u89e3\u8bfb<\/p>\n
：\u8fd9\u662f \u201charness \u662f\u7a0b\u5e8f\u3001\u4e0d\u662f\u6743\u91cd\u201d \u7684\u76f4\u63a5\u5b9e\u9a8c\u8bc1\u636e\u3002\u4e00\u4e2a\u53ea\u5bf9\u641c\u7d22\u65f6\u6240\u7528\u7b56\u7565\u6709\u5e2e\u52a9\u7684\u673a\u5236，\u662f\u8be5\u7b56\u7565\u7684\u8fc7\u62df\u5408\u4ea7\u7269；\u800c RRSI \u7684\u6539\u8fdb\u80fd\u8de8\u9aa8\u5e72\u5b58\u6d3b，\u8bf4\u660e\u5b83\u88c5\u7684\u662f\u771f\u673a\u5236\u3002<\/p>\n
\u5bf9\u505a Agent \u5e73\u53f0\u7684\u4eba，\u8fd9\u6761\u610f\u5473\u7740 harness \u8fdb\u5316\u80fd\u529b\u5e94\u6c89\u6dc0\u4e3a\u72ec\u7acb\u4e8e\u5177\u4f53\u6a21\u578b\u7684\u4e00\u5c42\u8d44\u4ea7\u3002<\/p>\n
4.4 \u6210\u672c（\u56fe 4）<\/h4>\n
RRSI \u6bcf\u6b21\u8bd5\u9a8c 2.42M token \/ 26.3 \u6b65；prior \u65b9\u6cd5 2.69\u20133.82M token \/ 27.3\u201334.6 \u6b65\u3002AHE \u6bd4 RRSI \u591a\u82b1 58% token，OOD \u53cd\u800c\u4f4e 4.4 \u5206\u3002\u6b63\u5219\u5316\u4e0d\u662f\u82b1\u94b1\u4e70\u6cdb\u5316，\u800c\u662f\u9760\u5220\u65e0\u7528\u590d\u6742\u5ea6，\u65e2\u4e70\u4e86\u6cdb\u5316\u53c8\u7701\u4e86\u94b1\u3002<\/p>\n
\n5. \u53cd\u76f4\u89c9\u7ed3\u679c\u4e0e\u5931\u6548\u6761\u4ef6（\u4e3b\u52a8\u627e\u95ee\u9898）<\/h3>\n
- \n
evolve \u5206\u8d8a\u4f4e\u8d8a\u597d：\u8fd9\u4e0e \u201c\u8fdb\u5316\u5206\u6570 = \u8fdb\u6b65\u201d \u7684\u76f4\u89c9\u76f8\u53cd，\u662f\u5168\u6587\u6700\u8be5\u88ab\u8bb0\u4f4f\u7684\u53cd\u76f4\u89c9\u70b9\u3002<\/p>\n<\/li>\n
- \n
\u672a\u8fdb\u5316\u7684 𝐻\u2080 \u5e38\u5e38 OOD \u66f4\u7a33：TTHE\/AHE \u8fdb\u5316\u540e OOD \u53cd\u800c\u8dcc \u2014\u2014 \u8bf4\u660e\u76f2\u76ee\u8fdb\u5316\u662f\u8d1f\u6536\u76ca\u3002<\/p>\n<\/li>\n
- \n
\u8fc1\u79fb\u4e0d\u662f\u88c1\u5224\u6253\u5206\u7684\u5047\u8c61：\u4f5c\u8005\u7279\u610f\u7528\u5de5\u7a0b\u8bbe\u8ba1（\u786e\u5b9a\u6027\u4eff\u771f\u3001\u65e0\u88c1\u5224\u6a21\u578b）\u590d\u73b0\u4e86\u540c\u6837\u7684\u589e\u76ca，\u5835\u6b7b\u4e86 \u201c\u9760\u8ba8\u88c1\u5224\u559c\u6b22\u63d0\u5206\u201d \u8fd9\u6761\u89e3\u91ca\u3002<\/p>\n<\/li>\n
- \n
\u4ecd\u5b58\u5728\u7684\u5931\u6548：evolve \u96c6\u63d0\u5347\u8d8a\u5927，OOD \u7559\u5b58\u7387\u8d8a\u4f4e（\u56fe 1a \u7684 1:1 \u865a\u7ebf）；RRSI \u81ea\u5df1\u4e5f\u53ea\u505a\u5230\u7ea6 4.3% OOD \/ 2.5% evolve，\u8fdc\u6ca1\u5230 \u201c\u5b8c\u5168\u8fc1\u79fb\u201d\u3002<\/p>\n<\/li>\n
- \n
\u6848\u4f8b\u7814\u7a76（\u8868 6）\u66b4\u9732\u8fb9\u754c：\u7f16\u7801 R8-B \u90a3\u4e2a\u5019\u9009\u628a\u4efb\u52a1\u6307\u4ee4\u9489\u8fdb\u5b8c\u6210\u95e8，\u6210\u672c\u964d 13.6%，\u4f46\u56e0\u6027\u80fd \u22122.81 \u4ecd\u88ab\u5e95\u7ebf\u62d2\u7edd \u2014\u2014\u5355\u9760\u7701\u94b1\u4e0d\u80fd\u515c\u5e95；\u800c\u5de5\u7a0b R2 \u90a3\u4e2a\u5c0f\u7684\u63a7\u5236\u6d41\u4fee\u6b63（+1.6% token）\u5374\u88ab\u4fdd\u7559，\u56e0\u4e3a\u5b83\u662f\u5c0f\u800c\u53ef\u590d\u7528\u7684\u673a\u5236\u3002<\/p>\n<\/li>\n
\n6. \u53ef\u4fe1\u8fb9\u754c\u3001\u5c40\u9650\u4e0e\u590d\u73b0\u98ce\u9669<\/h3>\n
\u4f5c\u8005\u660e\u786e\u627f\u8ba4\u7684\u5c40\u9650：<\/p>\n
- \n
- \n
\u53ea\u505a\u51bb\u7ed3\u9aa8\u5e72\u4e0b\u7684 harness \u7ea7 RSI，\u4e0d\u6d89\u53ca\u8fdb\u5316\u4e2d\u66f4\u65b0\u6a21\u578b\u6743\u91cd；<\/p>\n<\/li>\n
- \n
\u4ecd\u4f9d\u8d56\u6709\u9650 evolve \u96c6\u548c\u82e5\u5e72\u6b63\u5219\u8d85\u53c2，\u6548\u679c\u4f9d\u8d56\u53cd\u9988\u8d28\u91cf\u4e0e\u641c\u7d22\u9884\u7b97；<\/p>\n<\/li>\n
- \n
\u8de8 domain\/benchmark\/ \u7b56\u7565\u505a\u4e86\u8fc1\u79fb，\u4f46\u672a\u9a8c\u8bc1\u5dee\u5f02\u5de8\u5927\u7684 agent \u67b6\u6784\u3001\u5de5\u5177\u751f\u6001\u548c\u66f4\u957f\u65f6\u7a0b\u7684\u81ea\u6539\u8fdb\u3002<\/p>\n<\/li>\n<\/ul>\n
\u6211\u6839\u636e\u8bc1\u636e\u63a8\u65ad\u7684\u5c40\u9650（\u66f4\u8981\u7d27）：<\/p>\n
- \n
\u65e0\u7edf\u8ba1\u663e\u8457\u6027：\u6bcf\u4efb\u52a1\u4ec5 k=2~4 \u6b21\u8bd5\u9a8c，\u5168\u6587\u65e0\u7f6e\u4fe1\u533a\u95f4 \/ \u8bef\u5dee\u68d2\u3002OOD +1.8~4.7 \u5206\u662f\u5426\u7a33\u5065，\u65e0\u6cd5\u4ece\u8bba\u6587\u5224\u65ad\u3002<\/p>\n<\/li>\n
- \n
\u8d85\u53c2\u5bf9 domain \u654f\u611f：\u03b21 \u5728\u4e09\u4e2a domain \u53d6 44.5 \/ 35.4 \/ 24.4，\u03b4 \u53d6 0.017 \/ 0.004 \/ 0.020，\u5dee\u5f02\u5f88\u5927\u3002\u6362 domain \u5fc5\u987b\u91cd\u65b0\u6821\u51c6，\u201c\u5f00\u7bb1\u5373\u7528\u201d \u6027\u5b58\u7591\u3002<\/p>\n<\/li>\n
- \n
\u201c\u6b63\u5219\u5316\u201d \u662f\u5de5\u7a0b\u89c4\u5219\u4e0d\u662f\u8bc1\u660e：L0\/L1\/L2 \u662f\u5b9a\u6027\u7c7b\u6bd4，\u65e0\u6cdb\u5316\u754c\u3002\u8fd9\u662f\u4e00\u7bc7\u7ecf\u9a8c\u7cfb\u7edf\u8bba\u6587，\u522b\u5f53\u7406\u8bba\u8bba\u6587\u5f15\u7528\u3002<\/p>\n<\/li>\n
- \n
critic \u53ea\u80fd\u6293\u663e\u5f0f\u6cc4\u6f0f：D \u95f8\u9760\u4e00\u4e2a LLM critic \u8bfb diff，\u53ea\u80fd\u62e6 \u201c\u5199\u6b7b\u4efb\u52a1\u540d \/ \u7b54\u6848\u201d \u8fd9\u7c7b\u663e\u5f0f\u8fc7\u62df\u5408；\u63d0\u793a\u8bcd\u91cc\u85cf\u8fdb benchmark \u98ce\u683c\u7ebf\u7d22\u8fd9\u7c7b\u9690\u5f0f\u8fc7\u62df\u5408，critic \u5f88\u96be\u8bc6\u522b\u3002<\/p>\n<\/li>\n
- \n
\u641c\u7d22\u6210\u672c\u6781\u9ad8：\u6bcf\u4e2a\u5019\u9009\u8981\u5728\u6574\u4e2a evolve \u96c6\u4e0a\u8dd1 k \u6b21\u771f\u5b9e agent rollout，20\u201340 \u8f6e \u00d7 \u591a\u5019\u9009 \u2014\u2014 \u590d\u73b0\u7684\u4e3b\u8981\u95e8\u69db\u662f\u94b1\u548c\u7b97\u529b，\u4e0d\u662f\u7b97\u6cd5\u3002<\/p>\n<\/li>\n
- \n
\u8fdb\u5316\u65f6\u7a0b\u77ed：20\u201340 \u8f6e\u79bb \u201c\u957f\u65f6\u7a0b\u3001\u6210\u767e\u4e0a\u5343\u8f6e\u7684\u9012\u5f52\u81ea\u6211\u6539\u8fdb\u201d \u8fd8\u5f88\u8fdc，\u03b4 \u5728\u957f\u7a0b\u4e0b\u662f\u5426\u4ecd\u7136\u5408\u7406\u672a\u9a8c\u8bc1\u3002<\/p>\n<\/li>\n
\u590d\u73b0\u7b2c\u4e00\u6279\u52a8\u4f5c：\u2460 \u8dd1\u901a\u5f00\u6e90 repo \u7684 coding \u5355\u57df（Terminal-Bench + SWE-bench OOD）；\u2461 \u590d\u73b0\u8868 2 \u6d88\u878d，\u786e\u8ba4 \u201c\u53bb\u6b63\u5219\u2192evolve \u6da8 \/ OOD \u8dcc \/token \u6da8\u201d \u7684\u65b9\u5411；\u2462 \u7ed9\u5173\u952e OOD \u6570\u5b57\u8865\u591a\u6b21\u91cd\u590d，\u4f30\u8ba1\u7f6e\u4fe1\u533a\u95f4\u3002<\/p>\n
\n7. \u5bf9\u4f60（AI Agent \u67b6\u6784\u89c6\u89d2）\u7684\u53ef\u8fc1\u79fb\u6d1e\u5bdf<\/h3>\n
- \n
\u628a evolve \u96c6\u770b\u4f5c \u201c\u8bad\u7ec3\u7f13\u5b58\u201d，\u53cd\u590d\u8bfb\u5c31\u4f1a\u6c61\u67d3\u3002 \u8fd9\u548c\u4f60\u628a Context Window \u7c7b\u6bd4\u7f13\u5b58\u5c42\u7ea7\u662f\u5bf9\u5076\u7684：\u03b4（\u566a\u58f0\u5e95\u7ebf）\u672c\u8d28\u662f \u201c\u8bfb\u7a7f\u7f13\u5b58\u540e，\u5fc5\u987b\u505c\u6b62\u628a\u6ce2\u52a8\u5f53\u4fe1\u53f7\u201d\u3002\u4efb\u4f55 \u201c\u7528\u56fa\u5b9a\u53cd\u9988\u96c6\u53cd\u590d\u6539\u8fdb agent\u201d \u7684\u7cfb\u7edf，\u90fd\u8be5\u5185\u7f6e\u4e00\u4e2a \u201c\u6709\u6548\u6837\u672c\u91cf\u6d88\u8017\u201d \u7684\u6982\u5ff5\u3002<\/p>\n<\/li>\n
- \n
Harness \u662f\u7a0b\u5e8f\u8d44\u4ea7，\u53ef\u8de8\u6a21\u578b\u6c89\u6dc0\u3002 \u8868 4 \u7ed9\u4e86\u5b9e\u9a8c\u652f\u6491：harness \u5c42\u7684\u673a\u5236\u6539\u8fdb\u4e0d\u7ed1\u5b9a\u5177\u4f53\u9aa8\u5e72\u3002\u8fd9\u610f\u5473\u7740\u4f60\u505a Agent OS \/ \u5e73\u53f0\u65f6，harness \u8fdb\u5316\u4e0e\u6cbb\u7406\u5e94\u4f5c\u4e3a\u72ec\u7acb\u4e8e\u5177\u4f53\u6a21\u578b\u7248\u672c\u7684\u4e00\u5c42，\u800c\u4e0d\u662f\u6bcf\u6b21\u53d1\u6a21\u578b\u91cd\u5199\u4e00\u904d\u3002<\/p>\n<\/li>\n
- \n
\u9009\u62e9\u7aef\u56db\u9053\u95f8 = Agent \u7684 \u201c\u63a7\u5236\u9762 \/ \u53d8\u66f4\u6cbb\u7406\u5c42\u201d\u3002 critic（\u6cc4\u6f0f\u5ba1\u67e5）\u2192 noise floor（\u7a33\u5b9a\u6027）\u2192 cost-gate（\u6027\u4ef7\u6bd4）\u2192 domain guard（\u9886\u57df\u7ea2\u7ebf），\u548c\u4f01\u4e1a\u7ea7 AI OS \u9700\u8981\u7684\u53ef\u5ba1\u8ba1\u53d8\u66f4\u6cbb\u7406\u3001\u975e\u8865\u507f\u6027\u51c6\u5165\u540c\u6784\u3002\u8fd9\u6b63\u662f Palantir \u5f0f \u201c\u51b3\u7b56\u95e8 \/ Ontology\u201d \u5728 agent \u81ea\u6539\u8fdb\u573a\u666f\u7684\u5bf9\u5e94\u7269：\u6bcf\u4e2a\u5bf9\u751f\u4ea7 harness \u7684\u6539\u52a8\u90fd\u8981\u7559\u75d5\u3001\u8fc7\u95f8\u3001\u53ef\u56de\u6eda\u3002<\/p>\n<\/li>\n
- \n
\u53ef\u76f4\u63a5\u5de5\u7a0b\u5316\u7684 pattern：<\/p>\n<\/li>\n
- \n
- \n
\u4f59\u5f26\u9000\u706b\u7f16\u8f91\u9884\u7b97 b_t = b_min + (b_max\u2212b_min)\u00b7\u00bd(1+cos(\u03c0t\/T))；<\/p>\n<\/li>\n
- \n
\u7ed9\u6bcf\u4e2a diff \u7ed3\u6784\u5316\u8bb0\u8d26：(\u7ec4\u4ef6\u2113, \u5047\u8bbeh, diff, \u0394S, \u0394C, \u662f\u5426\u88ab\u63a5\u53d7)，\u4f5c\u4e3a proposer \u7684\u6761\u4ef6\u4e0a\u4e0b\u6587；<\/p>\n<\/li>\n
- \n
\u56fa\u5b9a\u7a97\u53e3 n_prune \u5185\u65e0\u6b63\u8d21\u732e\u7684\u7ec4\u4ef6\u2192\u5220\u9664\u76ee\u6807；<\/p>\n<\/li>\n
- \n
\u6210\u672c – \u589e\u76ca\u7ebf\u6027\u95f8\u95e8 \u0394C \u2264 \u03b20 + \u03b21\u00b7\u0394S，\u9632\u6b62 token \u5360\u7528\u65e0\u7ea6\u675f\u81a8\u80c0\u3002<\/p>\n<\/li>\n<\/ul>\n
- \u4e00\u53e5\u5de5\u7a0b\u7ed3\u8bba：\u5728\u4f60\u81ea\u5df1\u7684 agent loop \u91cc，\u5982\u679c\u5df2\u7ecf\u5728 \u201c\u7528\u53cd\u9988\u81ea\u52a8\u6539 harness\/prompt\/ \u5de5\u5177\u201d，\u90a3\u4f60\u5df2\u7ecf\u5728\u505a RSI\u2014\u2014 \u7f3a\u7684\u5f80\u5f80\u4e0d\u662f\u66f4\u5f3a\u7684 proposer，\u800c\u662f\u4e00\u4e2a\u6297\u8fc7\u62df\u5408\u7684\u9009\u62e9\u7aef\u6cbb\u7406\u5c42\u3002<\/li>\n
\n8. \u9605\u8bfb\u8303\u56f4\u4e0e\u4fe1\u606f\u8fb9\u754c<\/h3>\n
- \n
- \n
\u672c\u89e3\u8bfb\u57fa\u4e8e\u7528\u6237\u63d0\u4f9b\u7684\u5b8c\u6574 24 \u9875 PDF（\u6b63\u6587 + \u9644\u5f55 A\u2013E + \u7b97\u6cd5 1\/2 + \u8868 1\u20136），\u975e\u4ec5\u51ed\u6458\u8981\u3002<\/p>\n<\/li>\n
- \n
\u6570\u5b57\u5747\u51fa\u81ea\u539f\u6587\u8868 \/ \u56fe；\u7b2c 6 \u8282 \u201c\u63a8\u65ad\u5c40\u9650\u201d \u4e3a\u5206\u6790\u8005\u5224\u65ad，\u975e\u8bba\u6587\u539f\u6587\u3002<\/p>\n<\/li>\n
- \n
\u672a\u8054\u7f51\u6838\u5bf9 repo \u5b9e\u9645\u5b9e\u73b0\u4e0e\u8d85\u53c2\u4ee3\u7801；\u590d\u73b0\u53ef\u884c\u6027\u4ee5\u4f5c\u8005\u516c\u5f00\u6750\u6599\u4e3a\u51c6\u3002<\/p>\n<\/li>\n<\/ul>\n
RRSI：\u667a\u80fd\u4f53 Harness \u7684\u6b63\u5219\u5316\u9012\u5f52\u81ea\u6211\u6539\u8fdb<\/h2>\n
RRSI: Regularized Recursive Self-Improvement of Agent Harnesses<\/p>\n
Peng Xia\u00b9\u00b7\u00b2*, Rujun Han\u00b9, Zifeng Wang\u00b9, Yanfei Chen\u00b9, Yufan Zhang\u00b9, Yoonho Lee\u00b3, Chengsong Huang\u2074, Han Yu\u00b9, Zhongying CuiZhu\u00b9, Yifei Ming\u00b9, Huaxiu Yao\u00b2, Burak Gokturk\u00b9, Tomas Pfister\u00b9, Chen-Yu Lee\u00b9<\/p>\n
\u00b9 Google Cloud AI Research（\u8c37\u6b4c\u4e91 AI \u7814\u7a76\u9662）；\u00b2 UNC-Chapel Hill（\u5317\u5361\u7f57\u6765\u7eb3\u5927\u5b66\u6559\u5802\u5c71\u5206\u6821）；\u00b3 Stanford University（\u65af\u5766\u798f\u5927\u5b66）；\u2074 Washington University in St. Louis（\u5723\u8def\u6613\u65af\u534e\u76db\u987f\u5927\u5b66）<\/p>\n
\u8bba\u6587\u65e5\u671f：2026-09-22\u3000|\u3000arXiv:2609.24972v1 [cs.LG] \u4ee3\u7801\u4e0e\u4e3b\u9875：github.com\/google-research\/rrsi\u3000\u00b7\u3000regularized-rsi.com<\/p>\n
\n\u6458\u8981<\/h3>\n
LLM \u667a\u80fd\u4f53（agent）\u7684\u80fd\u529b，\u5f88\u5927\u7a0b\u5ea6\u4e0a\u662f\u7531\u5176 harness（\u5916\u58f3\/\u5de5\u4f5c\u6846\u67b6） \u653e\u5927\u7684\u2014\u2014\u4e5f\u5c31\u662f\u5305\u88f9\u5728\u51bb\u7ed3\u9aa8\u5e72\u6a21\u578b\u4e4b\u5916\u7684\u63d0\u793a\u8bcd（prompts）\u3001\u63a7\u5236\u6d41（control flow）\u3001\u5de5\u5177（tooling）\u3001\u8bb0\u5fc6（memory）\u4e0e\u4e0a\u4e0b\u6587\u7ba1\u7406（context management）\u3002\u8fd1\u671f\u65b9\u6cd5\u8d8a\u6765\u8d8a\u591a\u5730\u901a\u8fc7\u8fed\u4ee3\u5730\u63d0\u51fa\u5e76\u9009\u62e9\u5bf9 harness \u7684\u9010\u7ec4\u4ef6\u7f16\u8f91，\u6765\u81ea\u52a8\u5316\u8fd9\u4e00\u8fc7\u7a0b，\u5b9e\u9645\u4e0a\u5728"\u667a\u80fd\u4f53\u7cfb\u7edf"\u5c42\u9762\u5efa\u7acb\u4e86\u4e00\u79cd\u9012\u5f52\u81ea\u6211\u6539\u8fdb（recursive self-improvement, RSI）\u3002\u7136\u800c，\u8fd9\u79cd\u9012\u5f52\u5f0f\u8fdb\u5316\u53ef\u80fd\u901a\u8fc7\u6b7b\u8bb0\u8bad\u7ec3\u4efb\u52a1\u800c\u8fc7\u62df\u5408：\u5728\u5206\u5e03\u5185（in-distribution）\u57fa\u51c6\u4e0a\u53d6\u5f97\u5927\u5e45\u63d0\u5347，\u5374\u5728\u5206\u5e03\u5916（out-of-distribution, OOD）\u57fa\u51c6\u4e0a\u7f29\u6c34\u751a\u81f3\u6d88\u5931\u3002<\/p>\n
\u6211\u4eec\u63d0\u51fa RRSI（Regularized Recursive Self-Improvement of Agent Harnesses，\u667a\u80fd\u4f53 Harness \u7684\u6b63\u5219\u5316\u9012\u5f52\u81ea\u6211\u6539\u8fdb），\u901a\u8fc7\u7ea6\u675f\u8fdb\u5316\u5019\u9009\u7684\u63d0\u51fa（proposal）\u4e0e\u9009\u62e9（selection），\u628a\u6b63\u5219\u5316\u539f\u5219\u5f15\u5165 harness \u7684\u81ea\u6211\u6539\u8fdb\u3002proposer（\u63d0\u51fa\u5668）\u5728\u4e00\u4e2a\u968f\u65f6\u95f4\u9000\u706b（temporally annealed）\u7684\u9884\u7b97\u4e0b\u5de5\u4f5c，\u9650\u5236\u5355\u4e2a\u5019\u9009\u80fd\u6346\u7ed1\u591a\u5c11\u5904\u7f16\u8f91，\u5e76\u57fa\u4e8e\u8fdb\u5316\u5386\u53f2\u9f13\u52b1\u5c1a\u672a\u63a2\u7d22\u8fc7\u7684\u8def\u5f84；selector（\u9009\u62e9\u5668）\u5219\u914d\u5907\u4e00\u4e2a critic（\u6279\u8bc4\u5668）\u548c\u4e00\u4e2a pruner（\u526a\u679d\u5668）：critic \u8fc7\u6ee4\u6389\u9488\u5bf9\u7279\u5b9a benchmark \u7684\u63d0\u6848，pruner \u79fb\u9664\u90a3\u4e9b\u6539\u52a8\u8fc7\u5c0f\u3001\u6210\u672c\u8fc7\u9ad8\u6216\u4e0d\u518d\u6709\u7528\u7684\u53d8\u66f4\u3002\u8fd9\u4e9b\u7ea6\u675f\u5171\u540c\u503e\u5411\u4e8e\u53ef\u590d\u7528\u7684\u667a\u80fd\u4f53\u673a\u5236，\u800c\u975e\u9488\u5bf9\u7279\u5b9a benchmark \u7684\u6539\u52a8\u3001\u751a\u81f3\u566a\u58f0\u3002<\/p>\n
\u5728\u6a2a\u8de8\u7f16\u7801\u3001\u667a\u80fd\u4f53\u5de5\u4f5c\u533a（agentic workspace）\u4e0e\u5de5\u7a0b\u8bbe\u8ba1\u4efb\u52a1\u7684\u516b\u4e2a benchmark \u4e0a，RRSI \u5728\u5176\u8d56\u4ee5\u8fdb\u5316\u7684（evolve）\u6570\u636e\u96c6\u4e0a\u6700\u9ad8\u63d0\u5347 14.1 \u5206，\u5728\u4e94\u4e2a\u5206\u5e03\u5916 benchmark \u4e0a\u6700\u9ad8\u63d0\u5347 4.7 \u5206；\u4e0e\u6b64\u540c\u65f6，\u5176\u6240\u4ea7\u51fa harness \u6d88\u8017\u7684\u7b56\u7565 token \u6570\u91cf\u6bd4\u672a\u6b63\u5219\u5316\u7684\u8fdb\u5316\u5c11 30%\u3002<\/p>\n
\n1. \u5f15\u8a00<\/h3>\n
\u73b0\u4ee3 LLM \u667a\u80fd\u4f53\u662f"\u7cfb\u7edf"\u800c\u975e\u5b64\u7acb\u7684\u6a21\u578b（Lopopolo, 2026; Rajasekaran, 2026）\u3002\u4e00\u4e2a\u51bb\u7ed3\u7684\u9aa8\u5e72\u6a21\u578b\u88ab\u5305\u88f9\u5728\u7531\u63d0\u793a\u8bcd\u3001\u63a7\u5236\u6d41\u3001\u5de5\u5177\u63a5\u53e3\u3001\u8bb0\u5fc6\u4e0e\u4e0a\u4e0b\u6587\u7ba1\u7406\u6784\u6210\u7684 harness \u4e4b\u4e2d\u3002harness \u51b3\u5b9a\u4e86\u540c\u4e00\u4e2a\u6a21\u578b：\u5728\u7f16\u8f91\u6587\u4ef6\u524d\u662f\u5426\u8bfb\u5bf9\u4e86\u6587\u4ef6\u3001\u80fd\u5426\u4ece\u5931\u8d25\u7684\u547d\u4ee4\u4e2d\u6062\u590d\u3001\u80fd\u5426\u9ad8\u6548\u5730\u7ef4\u62a4\u5de5\u4f5c\u4e0a\u4e0b\u6587\u3001\u4ee5\u53ca\u80fd\u5426\u628a\u53d1\u73b0\u5199\u8fdb\u4ea4\u4ed8\u7269\u3002\u8fd1\u671f\u667a\u80fd\u4f53\u4ea7\u54c1\u7684\u8bb8\u591a\u8fdb\u5c55\u6765\u81ea harness \u5de5\u7a0b，\u800c\u975e\u65b0\u7684\u6a21\u578b\u6743\u91cd（Karten et al., 2026a; Weng, 2026; Zhang and Khattab, 2026）\u3002\u7136\u800c\u8fd9\u79cd\u5de5\u7a0b\u4f9d\u8d56\u4eba\u5de5：\u4eba\u53bb\u68c0\u89c6\u5931\u8d25\u7684\u8f68\u8ff9\u3001\u624b\u5de5\u8c03\u6574\u811a\u624b\u67b6（scaffold），\u8fdb\u5c55\u53d7\u9650\u4e8e\u4e00\u4e2a\u5de5\u7a0b\u5e08\u80fd\u8bfb\u591a\u5c11\u6761\u8f68\u8ff9\u3002<\/p>\n
\u8fd1\u671f\u65b9\u6cd5\u7528 LLM \u4ece\u4efb\u52a1\u53cd\u9988\u51fa\u53d1\u4f18\u5316 harness \u7ec4\u4ef6，\u4ece\u800c\u81ea\u52a8\u5316\u8fd9\u4e2a\u5faa\u73af（Chen et al., 2026; Karten et al., 2026b; Lee et al., 2026a,b; Lin et al., 2026a; Lou et al., 2026; Nie et al., 2026; Niklaus, 2026; Zhang et al., 2026a,e）\u3002\u8fd9\u79cd\u8fed\u4ee3\u5f0f\u7684 harness \u8fdb\u5316，\u5728\u667a\u80fd\u4f53\u7cfb\u7edf\u5c42\u9762\u63d0\u4f9b\u4e86\u4e00\u79cd\u5b9e\u7528\u5f62\u5f0f\u7684\u9012\u5f52\u81ea\u6211\u6539\u8fdb（RSI）（RSI-Exam Team, 2026; Team et al., 2026; Wang et al., 2025; Zhang et al., 2026b）：\u7528\u5f53\u524d\u7cfb\u7edf\u7684\u53cd\u9988\u53bb\u6539\u8fdb\u5851\u9020\u5176\u540e\u7eed\u884c\u4e3a\u7684 harness\u3002\u7136\u800c，\u5982\u56fe 1(a) \u6240\u793a，\u6d4b\u8bd5\u671f（test-time）harness \u8fdb\u5316\u53cd\u590d\u7528\u4e00\u4e2a\u6709\u9650\u7684 evolve \u6570\u636e\u96c6\u7684\u53cd\u9988\u6765\u63d0\u51fa\u5e76\u9009\u62e9\u7f16\u8f91，\u5e26\u6765\u4e86"\u81ea\u9002\u5e94\u8fc7\u62df\u5408"\u98ce\u9669：evolve \u96c6\u4e0a\u7684\u8868\u73b0\u53ef\u80fd\u63d0\u5347，\u5374\u5e76\u672a\u5728\u672a\u89c1\u4efb\u52a1\u4e0a\u5e26\u6765\u76f8\u5e94\u6536\u76ca\u3002\u8fd1\u671f\u7814\u7a76\u89c2\u5bdf\u5230\u8fdb\u5316\u8868\u73b0\u4e0e\u7559\u51fa（held-out）\u8868\u73b0\u4e4b\u95f4\u5b58\u5728\u663e\u8457\u5dee\u8ddd，\u5e76\u8868\u660e\u8868\u9762\u4e0a\u7684\u63d0\u5347\u53ef\u80fd\u6765\u81ea\u9488\u5bf9\u7279\u5b9a\u4efb\u52a1\u7684\u62df\u5408，\u6216\u6765\u81ea\u589e\u52a0\u7684\u6d4b\u8bd5\u671f\u8ba1\u7b97，\u800c\u975e\u53ef\u590d\u7528\u7684\u673a\u5236（Ding et al., 2026; Lin et al., 2026b; Wang et al., 2026b）\u3002\u76f8\u5e94\u5730，\u8fd1\u671f\u5de5\u4f5c\u663e\u5f0f\u5730\u628a\u8fdb\u5316\u4efb\u52a1\u4e0e\u8bc4\u4f30\u4efb\u52a1\u5206\u5f00，\u4ee5\u5ea6\u91cf\u6cdb\u5316\u80fd\u529b（Huang et al., 2026d; Ke et al., 2026; Zhang et al., 2026d）\u3002\u56e0\u6b64\u6211\u4eec\u7814\u7a76\u9012\u5f52\u81ea\u6211\u6539\u8fdb\u4e2d\u7684\u6cdb\u5316\u95ee\u9898，\u5176\u5b9a\u4e49\u4e3a：\u8fdb\u5316\u540e\u7684 harness \u80fd\u5426\u8fc1\u79fb\u5230\u4efb\u52a1\u63cf\u8ff0\u3001\u5de5\u5177\u63a5\u53e3\u6216\u9a8c\u8bc1\u5668\u90fd\u4e0d\u540c\u7684\u672a\u89c1 benchmark \u4e0a\u3002<\/p>\n
\u6211\u4eec\u7684\u7814\u7a76\u8868\u660e，\u8fc7\u62df\u5408\u4f1a\u901a\u8fc7\u82e5\u5e72\u76f8\u4e92\u8026\u5408\u7684\u884c\u4e3a\u4ea7\u751f（Yang et al., 2026a; Zhang et al., 2026d）：\u8fdb\u5316\u641c\u7d22\u53ef\u80fd\u7f16\u7801\u7279\u5b9a benchmark \u7684\u6a21\u5f0f\u3001\u63a8\u4e3e\u88ab\u8bc4\u4f30\u566a\u58f0\u504f\u7231\u7684\u5019\u9009\u3001\u6216\u79ef\u7d2f\u90a3\u4e9b\u63d0\u9ad8\u4e86 evolve \u96c6\u5206\u6570\u5374\u5e76\u672a\u6539\u8fdb\u5e95\u5c42\u667a\u80fd\u4f53\u673a\u5236\u7684\u590d\u6742\u5ea6\u3002\u8fd9\u4e9b"\u9488\u5bf9\u7279\u5b9a benchmark \u7684\u62df\u5408"\u201c\u8ffd\u9010\u566a\u58f0\u201d"\u79ef\u7d2f\u590d\u6742\u5ea6"\u90fd\u4f1a\u62c9\u5927"\u8fdb\u5316\u2014\u8fc1\u79fb"\u5dee\u8ddd\u3002\u53d7\u8fd9\u4e9b\u89c2\u5bdf\u542f\u53d1，\u6211\u4eec\u7684\u89e3\u51b3\u65b9\u6848\u5bf9"\u6709\u9650\u4e14\u5e26\u566a\u7684\u53cd\u9988\u5982\u4f55\u88ab\u7528\u4e8e\u9012\u5f52\u5f0f harness \u6539\u8fdb"\u52a0\u4ee5\u6b63\u5219\u5316\u3002<\/p>\n
\u6211\u4eec\u63d0\u51fa RRSI，\u4e00\u4e2a\u5bf9\u667a\u80fd\u4f53 harness \u7684 RSI \u8fdb\u884c\u6b63\u5219\u5316\u7684\u6846\u67b6：\u5b83\u4fdd\u6301 harness \u5b8c\u5168\u53ef\u7f16\u8f91，\u540c\u65f6\u7ea6\u675f\u6709\u9650\u7684 evolve \u96c6\u53cd\u9988\u5982\u4f55\u5f15\u5bfc\u641c\u7d22\u3002\u5982\u56fe 2 \u6240\u793a，RRSI \u5728\u8fdb\u5316\u5faa\u73af\u7684\u4e24\u7aef\u90fd\u65bd\u52a0\u6b63\u5219\u5316：\u5728\u63d0\u51fa\u5019\u9009\u65f6\u9f13\u52b1\u66f4\u7b80\u5355\u3001\u66f4\u53ef\u590d\u7528\u7684\u7f16\u8f91；\u5728\u9009\u62e9\u65f6\u65bd\u52a0\u7a33\u5065\u7684\u7b5b\u9009\u6807\u51c6，\u907f\u514d\u4fdd\u7559\u90a3\u4e9b\u7531\u7279\u5b9a benchmark \u4fe1\u53f7\u3001\u8bc4\u4f30\u566a\u58f0\u6216\u4e0d\u5fc5\u8981\u590d\u6742\u5ea6\u9a71\u52a8\u7684"\u6539\u8fdb"\u3002\u8fd9\u6837，RRSI \u503e\u5411\u4e8e\u80fd\u8fc1\u79fb\u5230 evolve \u96c6\u4e4b\u5916\u7684\u7f16\u8f91，\u540c\u65f6\u5e76\u4e0d\u9650\u5236\u54ea\u4e9b harness \u7ec4\u4ef6\u53ef\u4ee5\u88ab\u66f4\u65b0\u3002<\/p>\n
\u6211\u4eec\u5728\u6a2a\u8de8\u4e09\u4e2a\u9886\u57df\u3001\u4efb\u52a1\u7c7b\u578b\/\u5de5\u5177\/\u9a8c\u8bc1\u5668\u5404\u5f02\u7684\u516b\u4e2a benchmark \u4e0a\u8bc4\u4f30 RRSI\u3002\u6bcf\u4e2a\u9886\u57df\u4e2d，harness \u53ea\u5728\u4e00\u4e2a\u5957\u4ef6（suite）\u4e0a\u8fdb\u5316，\u7136\u540e\u539f\u5c01\u4e0d\u52a8\u5730\u5728\u7559\u51fa benchmark \u4e0a\u8fd0\u884c\u3002\u5982\u56fe 1(b\u2013d) \u6240\u793a，\u5b83\u5728\u8fdb\u5316\u6570\u636e\u96c6\u4e0a\u6700\u9ad8\u63d0\u5347 14.1 \u5206，\u5e76\u6539\u8fdb\u5168\u90e8\u516d\u4e2a\u7559\u51fa\u6570\u636e\u96c6，\u5206\u5e03\u5916\u6700\u9ad8\u63d0\u5347 4.7 \u5206，\u4e14\u6d88\u8017\u7684\u7b56\u7565 token \u6bd4\u672a\u6b63\u5219\u5316\u8fdb\u5316\u66f4\u5c11\u3002\u66f4\u91cd\u8981\u7684\u662f，\u8fd9\u4e9b\u63d0\u5347\u80fd\u6cdb\u5316\u5230\u8fdb\u5316\u6240\u7528\u73af\u5883\u4e4b\u5916：RRSI \u5728\u5dee\u5f02\u663e\u8457\u7684\u4efb\u52a1\u4e0e\u8bc4\u4f30\u8bbe\u7f6e\u4e0a\u90fd\u4fdd\u7559\u4e86\u6539\u8fdb，\u8868\u660e\u5b83\u5b66\u5230\u7684\u662f\u5e7f\u6cdb\u6709\u7528\u7684 harness \u6539\u52a8\u3002\u66f4\u91cd\u8981\u7684\u662f，RRSI \u5728\u5404\u7559\u51fa\u73af\u5883\u4e0a\u7684\u6cdb\u5316\u8d85\u8fc7\u6b64\u524d\u6700\u4f73\u57fa\u7ebf\u5e73\u5747\u503c\u8fbe 22.9%\u3002<\/p>\n
![\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260930025716-6abc7a8c48cbd.png\")
<\/p>\n\u56fe 1｜\u5728"\u88ab\u6253\u5206\u7684\u90a3\u4e2a split"\u4e0a\u8fc7\u62df\u5408，\u800c RRSI \u8ba9\u6539\u8fdb\u5f97\u4ee5\u6cdb\u5316\u3002 (a) \u667a\u80fd\u4f53\u5de5\u4f5c\u533a benchmark \u4e0a，evolve split \u4e0a\u7684\u63d0\u5347 vs. \u5206\u5e03\u5916\u63d0\u5347\u3002\u6b64\u524d\u65b9\u6cd5\u51e0\u4e4e\u7559\u4e0d\u4f4f evolve \u96c6\u7684\u63d0\u5347，\u82e5\u5e72\u65b9\u6cd5\u6700\u7ec8\u8fd8\u4f4e\u4e8e\u521d\u59cb harness 𝐻\u2080\u3002 (b\u2013d) 𝐻\u2080\u3001\u56db\u79cd\u57fa\u7ebf\u65b9\u6cd5\u7684\u5e73\u5747\u3001\u4ee5\u53ca\u6211\u4eec\u7684 RRSI \u5728 SWE-bench Verified（\u7f16\u7801）\u3001JobBench\/GDPval\/APEX-Agents \u4e09\u8005\u5747\u503c（\u667a\u80fd\u4f53\u5de5\u4f5c\u533a）\u3001Frontier-Eng（\u5de5\u7a0b\u8bbe\u8ba1）\u4e0a\u7684\u5206\u5e03\u5916\u7559\u51fa\u5206\u6570\u3002<\/p>\n
\u6838\u5fc3\u4fe1\u606f：\u5728 evolve split \u4e0a\u4e70\u6765\u5f97\u8d8a\u9ad8\u7684\u63d0\u5347（\u6a2a\u8f74），\u5206\u5e03\u5916 held-out \u63d0\u5347（\u7eb5\u8f74）\u5374\u51e0\u4e4e\u4f20\u4e0d\u8fc7\u53bb\u2014\u2014\u201cgain does not transfer\u201d\u3002RRSI（\u84dd\u8272\u661f）\u5728\u7ea6 2.5% evolve \u63d0\u5347\u5904\u6362\u5230\u4e86\u7ea6 4.3% \u7684 OOD \u63d0\u5347，\u662f\u552f\u4e00\u663e\u8457\u8d8a\u8fc7 1:1 transfer \u865a\u7ebf\u7684\u65b9\u6cd5\u3002<\/p>\n
\u6211\u4eec\u7684\u8d21\u732e\u6709\u4e09：(1) \u6211\u4eec\u8bc6\u522b\u51fa \u8fc7\u62df\u5408\u662f harness \u5f0f\u9012\u5f52\u81ea\u6211\u6539\u8fdb\u7684\u5173\u952e\u6311\u6218；(2) \u6211\u4eec\u63d0\u51fa RRSI，\u5728 harness \u8fdb\u5316\u671f\u95f4\u540c\u65f6\u5bf9"\u63d0\u51fa"\u4e0e"\u9009\u62e9"\u65bd\u52a0\u6b63\u5219\u5316，\u540c\u65f6\u4fdd\u6301\u6bcf\u4e2a harness \u7ec4\u4ef6\u53ef\u7f16\u8f91；(3) \u5728\u4e09\u4e2a\u9886\u57df\u516b\u4e2a benchmark \u4e0a，RRSI \u540c\u65f6\u63d0\u5347\u4e86\u8fc1\u79fb\u80fd\u529b\u4e0e\u6548\u7387，\u8bc1\u660e\u4e86\u6240\u63d0\u65b9\u6cd5\u7684\u6709\u6548\u6027\u3002<\/p>\n
\n2. \u9884\u5907\u77e5\u8bc6<\/h3>\n
\u667a\u80fd\u4f53\u4e0e Harness\u3002 \u6211\u4eec\u8003\u8651\u4e00\u4e2a\u7531\u9aa8\u5e72\u7b56\u7565 𝜋 \u4e0e harness 𝐻 \u6784\u6210\u7684\u667a\u80fd\u4f53 𝐴 = (𝜋, 𝐻)\u3002harness \u5c31\u662f\u6743\u91cd\u4e4b\u5916\u7684\u4e00\u5207（Lopopolo, 2026; Rajasekaran, 2026）：\u7cfb\u7edf\u63d0\u793a\u8bcd\u4e0e\u4efb\u52a1\u63d0\u793a\u8bcd\u3001\u51b3\u5b9a\u667a\u80fd\u4f53\u4f55\u65f6\u89c4\u5212\/\u884c\u52a8\/\u53cd\u601d\/\u505c\u6b62\u7684\u63a7\u5236\u6d41\u3001\u5de5\u5177\u63a5\u53e3\u53ca\u5176\u63cf\u8ff0\u3001\u667a\u80fd\u4f53\u53ef\u67e5\u9605\u7684\u8bb0\u5fc6\u4e0e\u6280\u80fd\u6587\u4ef6\u3001\u4ee5\u53ca\u51b3\u5b9a\u7b56\u7565\u6bcf\u4e00\u6b65\u80fd\u770b\u5230\u4ec0\u4e48\u7684\u4e0a\u4e0b\u6587\u7ba1\u7406\u3002\u7ed9\u5b9a\u4efb\u52a1 𝑥 \u53ca\u5176\u73af\u5883，\u667a\u80fd\u4f53\u4ea7\u751f\u4e00\u6761\u8f68\u8ff9 𝜏 \u223c 𝐴(\u00b7 | 𝑥) \u4e0e\u4e00\u4efd\u4ea4\u4ed8\u7269，\u7531\u9a8c\u8bc1\u5668\u6253\u5206 𝑟(𝑥, 𝜏) \u2208 [0, 1]\u3002\u9a8c\u8bc1\u5668\u53ef\u4ee5\u662f\u7f16\u7801\u73af\u5883\u4e2d\u7684\u5355\u5143\u6d4b\u8bd5\u5957\u4ef6，\u4e5f\u53ef\u4ee5\u662f\u667a\u80fd\u4f53\u5de5\u4f5c\u533a\u73af\u5883\u4e2d\u7684"LLM \u5f53\u88c1\u5224"\u7a0b\u5e8f\u3002\u5bf9\u4efb\u52a1\u96c6 D，\u6211\u4eec\u5ea6\u91cf\u4efb\u52a1\u8868\u73b0\u4e0e\u7b56\u7565 token \u6210\u672c\u4e3a：<\/p>\n
𝑆(𝐻; D) = 𝔼_{𝑥\u223cD} 𝔼_{𝜏\u223c𝐴(\u00b7|𝑥)} [ 𝑟(𝑥, 𝜏) ]，\u3000\u3000𝐶(𝐻; D) = 𝔼_{𝑥\u223cD} 𝔼_{𝜏\u223c𝐴(\u00b7|𝑥)} [ 𝑐(𝜏) ]\u3000\u3000(1)<\/p>\n
\u5176\u4e2d 𝑐(𝜏) \u662f\u8be5\u8f68\u8ff9\u6d88\u8017\u7684\u7b56\u7565 token \u6570\u3002<\/p>\n
Harness \u8fdb\u5316\u3002 harness \u8fdb\u5316\u628a 𝐻 \u5f53\u4f5c\u4f18\u5316\u53d8\u91cf，\u540c\u65f6\u56fa\u5b9a\u9aa8\u5e72\u7b56\u7565（Lee et al., 2026b）\u3002\u591a\u6570\u65b9\u6cd5\u5b9e\u4f8b\u5316\u540c\u4e00\u4e2a\u901a\u7528\u5faa\u73af：\u5728\u7b2c 𝑡 \u8f6e，\u7528\u5f53\u524d harness 𝐻\u209c \u5728 evolve \u96c6 D_evolve \u4e0a\u6267\u884c\u4ee5\u83b7\u5f97\u8f68\u8ff9；\u628a\u8fd9\u4e9b\u8f68\u8ff9\u6c47\u603b\u6210\u53cd\u9988 F\u209c；\u7531\u4e00\u4e2a proposer LLM \u751f\u6210\u5019\u9009 harness；\u5728\u540c\u4e00\u4e2a evolve \u96c6\u4e0a\u8bc4\u4f30\u5019\u9009；\u9009\u51fa\u6700\u4f73\u5019\u9009\u4f5c\u4e3a\u4e0b\u4e00\u4efb incumbent（\u73b0\u4efb\u6700\u4f73）\u3002\u62bd\u8c61\u5730：<\/p>\n
𝐇\u209c = { 𝐻\u207d\u00b9\u207e\u209c, \u2026, 𝐻\u207d\u1d50\u1d57\u207e\u209c } \u223c 𝑃\u2080(\u00b7 | 𝐻\u209c, F\u209c)，\u3000\u3000𝐻\u209c\u208a\u2081 = argmax_{𝐻\u2032 \u2208 𝐇\u209c \u222a {𝐻\u209c}} \u015c(𝐻\u2032; D_evolve)\u3000\u3000(2)<\/p>\n
\u5176\u4e2d 𝑃\u2080 \u8868\u793a\u65e0\u7ea6\u675f\u7684\u63d0\u51fa\u8fc7\u7a0b，\u015c \u662f\u7531\u6709\u9650\u6b21\u968f\u673a\u667a\u80fd\u4f53\u8fd0\u884c\u5f97\u5230\u7684\u7ecf\u9a8c\u5206\u6570\u3002\u6bcf\u4e2a\u4efb\u52a1\u8dd1 𝑘 \u6b21\u8bd5\u9a8c\u65f6：<\/p>\n
\u015c(𝐻) = 1\/(𝑘|D_evolve|) \u03a3_{𝑥\u2208D_evolve} \u03a3_{𝑗=1\u2026𝑘} 𝑟(𝑥, 𝜏\u207d\u02e3\u207e,\u2c7c)，\u3000\u3000\u0108(𝐻) = 1\/(𝑘|D_evolve|) \u03a3_{𝑥\u2208D_evolve} \u03a3_{𝑗=1\u2026𝑘} 𝑐(𝜏\u207d\u02e3\u207e.\u2c7c)\u3000\u3000(3)<\/p>\n
\u4e0e\u666e\u901a\u8bc4\u4f30\u4e0d\u540c，\u8fd9\u79cd\u5bf9 D_evolve \u7684\u590d\u7528\u662f\u81ea\u9002\u5e94\u7684：\u7b2c 𝑡 \u8f6e\u63d0\u51fa\u7684\u5019\u9009\u4f9d\u8d56\u4e8e\u66f4\u65e9\u8f6e\u6b21\u5728\u540c\u4e00\u6279\u4efb\u52a1\u4e0a\u6d4b\u5f97\u7684\u7ed3\u679c\u3002\u56e0\u6b64 harness \u8fdb\u5316\u53ef\u89c6\u4e3a\u5728\u4e00\u4e2a\u5f02\u5e38\u9ad8\u8868\u8fbe\u529b\u7684\u641c\u7d22\u7a7a\u95f4\u4e0a\u505a\u81ea\u9002\u5e94\u7ecf\u9a8c\u4f18\u5316\u3002<\/p>\n
\n3. RRSI<\/h3>\n
\u6211\u4eec\u901a\u8fc7\u8fed\u4ee3\u5f0f harness \u8fdb\u5316\u6765\u7814\u7a76 RSI：\u53cd\u590d\u7528\u5f53\u524d\u667a\u80fd\u4f53\u7cfb\u7edf\u7684\u53cd\u9988\u6765\u63d0\u51fa\u5e76\u9009\u62e9\u5bf9 harness \u7684\u4fee\u6539\u3002RRSI \u6cbf\u7528\u8fd9\u4e00\u9012\u5f52\u6539\u8fdb\u8fc7\u7a0b\u3001\u4fdd\u6301 harness \u7f16\u8f91\u7a7a\u95f4\u5f00\u653e，\u4f46\u5bf9"\u641c\u7d22\u5982\u4f55\u5728\u8be5\u7a7a\u95f4\u4e2d\u79fb\u52a8"\u52a0\u4ee5\u6b63\u5219\u5316\u3002\u6838\u5fc3\u601d\u60f3\u662f\u628a\u673a\u5668\u5b66\u4e60\u4e2d\u7684\u6b63\u5219\u5316\u539f\u5219\u8fc1\u79fb\u5230\u81ea\u9002\u5e94 harness \u641c\u7d22\u4e2d：\u7a00\u758f\u66f4\u65b0\u9650\u5236\u4e00\u8f6e\u53cd\u9988\u80fd\u6539\u52a8\u591a\u5c11\u673a\u5236；\u8bc1\u636e\u611f\u77e5\u7684\u4fe1\u7528\u5206\u914d\u9632\u6b62\u641c\u7d22\u53cd\u590d\u628a\u5bb9\u91cf\u82b1\u5728\u5df2\u88ab\u8bc1\u4f2a\u7684\u5047\u8bbe\u4e0a；\u4fdd\u5b88\u9009\u62e9\u9632\u6b62\u6cc4\u6f0f（leakage）\u3001\u8bc4\u4f30\u566a\u58f0\u6216\u65e0\u4f9d\u636e\u7684\u8d44\u6e90\u81a8\u80c0\u6210\u4e3a\u6c38\u4e45\u7684 harness \u72b6\u6001\u3002<\/p>\n
![\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260930025717-6abc7a8d30d8d.png\")
<\/p>\n\u56fe 2｜RRSI \u603b\u89c8\u3002RRSI \u6b63\u5219\u5316\u7684\u662f"\u641c\u7d22\u8f68\u8ff9"，\u800c\u975e\u9650\u5236\u6f5c\u5728\u7684 harness \u7f16\u8f91\u7a7a\u95f4：\u63d0\u51fa\u4fa7\u7ea6\u675f\u63a7\u5236\u641c\u7d22\u5bb9\u91cf\u5982\u4f55\u88ab\u4f7f\u7528；\u9009\u62e9\u4fa7\u7ea6\u675f\u63a7\u5236\u54ea\u4e9b"\u6d4b\u5f97\u7684\u6539\u8fdb"\u88ab\u5141\u8bb8\u6210\u4e3a\u6c38\u4e45\u72b6\u6001\u3002<\/p>\n
- \n
- \u63d0\u51fa\u4fa7\u6b63\u5219\u5316（Proposal-side）\n
- \n
- A \u9000\u706b\u66f4\u65b0\u7a00\u758f\u6027（Annealed update sparsity）：\u66f4\u65b0\u5bb9\u91cf\u968f\u8f6e\u6b21 𝑡 \u4ece\u9ad8\u964d\u5230\u4f4e\u3002\u65e9\u671f\u8f6e\u6b21\u5141\u8bb8\u66f4\u5bbd\u7684\u7f16\u8f91，\u540e\u671f\u8f6e\u6b21\u504f\u597d\u7a00\u758f\u3001\u53ef\u5f52\u56e0\u7684\u6539\u52a8\u3002<\/li>\n
- B \u8bc1\u636e\u611f\u77e5\u4fe1\u7528\u5206\u914d（Evidence-aware credit assignment）：\u4f7f\u7528\u589e\u76ca\u4e0e\u9000\u6b65\u7684\u5b8c\u6574\u5386\u53f2，\u800c\u4e0d\u53ea\u662f\u6700\u8fd1\u4e00\u6b21\u80dc\u5229\u3002<\/li>\n
- C \u7ed3\u6784\u5316\u63a2\u7d22（Structured exploration）：\u5f53\u8fdb\u5c55\u505c\u6ede\u65f6，\u628a\u641c\u7d22\u91cd\u5b9a\u5411\u5230\u672a\u5145\u5206\u63a2\u7d22\u7684\u7ec4\u4ef6（\u5df2\u8bd5 𝑇\u209c \/ \u672a\u8bd5 𝑈\u209c \/ \u505c\u6ede）\u3002<\/li>\n<\/ul>\n<\/li>\n
- \u6b63\u5219\u5316\u8f6c\u79fb\u89c4\u5219（Regularized transition rule）\n
- \n
- \u63d0\u51fa：𝐻\u2032 \u223c P_reg(\u00b7 | 𝐻\u209c, feedback)\u2014\u2014\u4ece\u5f53\u524d harness \u4e0e\u53cd\u9988\u5386\u53f2\u91c7\u6837\u4e00\u4e2a\u6b63\u5219\u5316\u5019\u9009\u3002<\/li>\n
- \u9009\u62e9：𝐻\u209c\u208a\u2081 = \u6700\u4f18\u53ef\u63a5\u53d7\u7684 𝐻\u2032 \u6216 𝐻\u209c\u2014\u2014\u82e5\u67d0\u5019\u9009\u7a33\u5065\u5730\u66f4\u597d\u5219\u53d6\u4e4b，\u5426\u5219\u4fdd\u7559 𝐻\u209c\u3002<\/li>\n<\/ul>\n<\/li>\n
- \u9009\u62e9\u4fa7\u6b63\u5219\u5316（Selection-side）\n
- \n
- D \u6cc4\u6f0f\u7b5b\u67e5（Leakage screening）：\u62d2\u7edd\u9488\u5bf9\u7279\u5b9a benchmark \u6216\u7279\u5b9a\u4efb\u52a1\u7684\u903b\u8f91\u3002<\/li>\n
- E \u566a\u58f0\u8c03\u6574\u540e\u7684\u6027\u80fd\u5e95\u7ebf（Noise-adjusted performance floor）：\u9632\u6b62\u63a5\u53d7\u90a3\u4e9b\u53ef\u7531\u8bc4\u4f30\u566a\u58f0\u89e3\u91ca\u7684\u63d0\u5347\u3002<\/li>\n
- F \u590d\u6742\u5ea6\u611f\u77e5\u63a5\u53d7（L1 \u5f0f）：\u989d\u5916\u590d\u6742\u5ea6\u5fc5\u987b\u6709\u53ef\u6d4b\u589e\u76ca\u6765\u80cc\u4e66\u3002<\/li>\n
- G \u7ed3\u6784\u5316\u526a\u679d（L0 \u5f0f）：\u526a\u9664\u8fd1\u671f\u65e0\u6b63\u8d21\u732e\u7684\u7ec4\u4ef6\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n
3.1\u3000Harness \u8fdb\u5316\u7684\u6b63\u5219\u5316\u89c6\u89d2<\/h4>\n
\u8bb0 \u03a9(𝐻) \u4e3a\u4ece 𝐻 \u7ecf\u4efb\u610f\u6e90\u7801\u7f16\u8f91\u53ef\u8fbe\u7684 harness \u96c6\u5408\u3002RRSI \u6709\u610f\u4fdd\u6301 \u03a9(𝐻) \u5f00\u653e：\u63d0\u793a\u8bcd\u3001\u63a7\u5236\u6d41\u3001\u914d\u7f6e\u3001\u4e0a\u4e0b\u6587\u7ba1\u7406\u3001\u5de5\u5177\u3001\u6280\u80fd\u3001\u8bb0\u5fc6\u3001\u5b50\u667a\u80fd\u4f53\u90fd\u53ef\u4ee5\u88ab\u4fee\u6539\u3001\u6dfb\u52a0\u6216\u5220\u9664\u3002\u6211\u4eec\u5e76\u4e0d\u76f4\u63a5\u9650\u5236\u8fd9\u4e2a\u5047\u8bbe\u7a7a\u95f4，\u800c\u662f\u6b63\u5219\u5316\u7a7f\u8fc7\u5b83\u7684\u641c\u7d22\u8f68\u8ff9\u3002\u6bcf\u4e00\u8f6e 𝑡，proposer \u7528\u6709\u9650 evolve \u96c6\u7684\u53cd\u9988\u751f\u6210\u5bf9\u5f53\u524d harness 𝐻\u209c \u7684\u5019\u9009\u7f16\u8f91，selector \u51b3\u5b9a\u5176\u4e2d\u662f\u5426\u6709（\u82e5\u6709，\u54ea\u4e2a）\u5e94\u66ff\u6362 incumbent\u3002<\/p>\n
\u8fd9\u4e2a\u89c6\u89d2\u533a\u5206\u4e24\u79cd\u4e92\u8865\u7684\u6b63\u5219\u5316\u3002\u5728\u63d0\u51fa\u4fa7，\u6211\u4eec\u7ea6\u675f\u5355\u8f6e\u80fd\u52a8\u7528\u591a\u5c11\u81ea\u9002\u5e94\u5bb9\u91cf\u3001\u4ee5\u53ca\u8fd9\u5bb9\u91cf\u82b1\u5728\u54ea\u91cc；\u5728\u9009\u62e9\u4fa7，\u6211\u4eec\u7ea6\u675f\u54ea\u4e9b\u7ecf\u9a8c\u6539\u8fdb\u8db3\u591f\u5f3a\u3001\u8db3\u591f\u7701\u3001\u8db3\u591f\u65e0\u6cc4\u6f0f，\u4ece\u800c\u80fd\u5b58\u6d3b\u4e0b\u6765\u3002\u6211\u4eec\u7684\u590d\u6742\u5ea6\u63a7\u5236\u6846\u67b6\u501f\u9274\u4e09\u79cd\u7ecf\u5178\u6b63\u5219\u5316\u65b9\u6cd5（Goodfellow et al., 2016; Hastie et al., 2009; Louizos et al., 2018），\u7c7b\u6bd4\u5982\u4e0b：\u7f16\u8f91\u9884\u7b97\u6700\u63a5\u8fd1 𝐿\u2080 \u5f0f\u57fa\u6570\u7ea6\u675f，\u56e0\u4e3a\u5b83\u76f4\u63a5\u9650\u5236\u4e00\u6b21\u66f4\u65b0\u4e2d\u72ec\u7acb\u751f\u6548\u7684\u7f16\u8f91\u6570；\u7ed3\u6784\u5316\u526a\u679d\u7c7b\u6bd4 Lasso\/𝐿\u2081 \u5f0f\u7a00\u758f\u5316，\u56e0\u4e3a\u6301\u7eed\u65e0\u4ea7\u51fa\u7684\u7ec4\u4ef6\u4f1a\u4ece\u4fdd\u7559\u7684 harness \u4e2d\u88ab\u79fb\u9664，\u5f97\u5230\u66f4\u7a00\u758f\u7684\u7ed3\u6784；\u590d\u6742\u5ea6\u611f\u77e5\u63a5\u53d7\u7c7b\u6bd4 Ridge\/𝐿\u2082 \u5f0f\u6536\u7f29，\u56e0\u4e3a\u5b83\u6291\u5236\u8d44\u6e90\u5360\u7528\u603b\u91cf\u7684\u65e0\u7ea6\u675f\u589e\u957f，\u800c\u4e0d\u8981\u6c42\u5254\u9664\u4efb\u4f55\u7279\u5b9a\u7ec4\u4ef6\u3002\u8be6\u7ec6\u7b97\u6cd5\u89c1\u9644\u5f55 C\u3002<\/p>\n
3.2\u3000\u6b63\u5219\u5316\u63d0\u51fa\u5206\u5e03<\/h4>\n
\u63d0\u51fa\u5206\u5e03\u51b3\u5b9a\u4e86\u641c\u7d22\u80fd\u591a\u6fc0\u8fdb\u5730\u54cd\u5e94 evolve \u96c6\u53cd\u9988\u3002RRSI \u4ece\u4e09\u65b9\u9762\u6b63\u5219\u5316\u5b83：\u9000\u706b\u5355\u8f6e\u53ef\u52a8\u7528\u7684\u66f4\u65b0\u5bb9\u91cf；\u8ba9\u4fe1\u7528\u5206\u914d\u5728\u6574\u6b21\u8fd0\u884c\u4e2d\u8bc1\u636e\u611f\u77e5；\u4ee5\u53ca\u7ed3\u6784\u5316\u5730\u5b89\u6392\u8fd9\u5bb9\u91cf\u82b1\u5728\u54ea\u91cc\u3002<\/p>\n
𝐿\u2080 \u5f0f\u9000\u706b\u66f4\u65b0\u7a00\u758f\u6027\u3002 \u65e0\u7ea6\u675f\u7684 proposer \u4f1a\u628a\u8bb8\u591a\u4e92\u4e0d\u76f8\u5173\u7684\u4fee\u6539\u6346\u8fdb\u4e00\u4e2a\u5019\u9009\u3002\u8fd9\u7c7b\u5019\u9009\u6709\u6548\u5bb9\u91cf\u5f88\u9ad8：\u80fd\u62df\u5408\u5f53\u524d\u53cd\u9988\u66f4\u591a\u7684\u602a\u7656，\u800c\u4efb\u4f55\u6d4b\u5f97\u7684\u53d8\u5316\u90fd\u96be\u4ee5\u5f52\u56e0\u5230\u67d0\u4e2a\u5177\u4f53\u673a\u5236\u3002\u56e0\u6b64\u6211\u4eec\u5bf9\u4e00\u4e2a\u63d0\u6848\u4e2d\u53ef\u72ec\u7acb\u5f52\u56e0\u7684\u7f16\u8f91\u6570\u91cf\u8bbe\u4e0a\u9650\u3002\u5728 𝑇 \u8f6e\u8fd0\u884c\u7684\u7b2c 𝑡 \u8f6e，\u8be5\u9884\u7b97\u4e3a：<\/p>\n
𝑏\u209c = \u2308 𝑏_min + (𝑏_max \u2212 𝑏_min) \u00b7 \u00bd(1 + cos(𝜋𝑡\/𝑇)) \u2309\u3000\u3000(4)<\/p>\n
\u8be5\u8c03\u5ea6\u4ece 𝑏_max \u964d\u5230 𝑏_min：\u65e9\u671f\u8f6e\u6b21\u53ef\u4ee5\u7ec4\u5408\u82e5\u5e72\u534f\u540c\u6539\u52a8\u4ee5\u53d1\u73b0\u65b0\u673a\u5236，\u540e\u671f\u8f6e\u6b21\u5219\u8d8a\u6765\u8d8a\u7a00\u758f\u3001\u8d8a\u6765\u8d8a\u53ef\u5f52\u56e0\u3002\u8fd9\u662f\u6211\u4eec\u6700\u76f4\u63a5\u7684\u7ecf\u5178\u7c7b\u6bd4：\u82e5\u628a\u5019\u9009\u4e2d"\u53ef\u72ec\u7acb\u5f52\u56e0\u7684\u7f16\u8f91"\u7528\u4e8c\u503c\u6d3b\u52a8\u6307\u793a\u8868\u793a，\u9884\u7b97\u5c31\u662f\u5bf9\u5176\u57fa\u6570\u7684\u7ea6\u675f，\u5373\u5bf9\u66f4\u65b0\u65bd\u52a0\u7684 𝐿\u2080 \u5f0f\u7ea6\u675f\u3002\u8be5\u7c7b\u6bd4\u9488\u5bf9\u7684\u662f\u66f4\u65b0\u7a00\u758f\u6027，\u800c\u975e\u56fa\u5b9a\u7684\u6a21\u578b\u53c2\u6570\u5411\u91cf：\u7f16\u8f91\u6c60\u8de8\u8f6e\u6b21\u4f1a\u53d8\u5316，\u6211\u4eec\u4e5f\u4e0d\u4f18\u5316\u4e00\u4e2a\u5e26 𝐿\u2080 \u60e9\u7f5a\u7684\u76ee\u6807\u3002<\/p>\n
\u8bc1\u636e\u611f\u77e5\u4fe1\u7528\u5206\u914d\u3002 \u53ea\u7ea6\u675f\u66f4\u65b0\u5927\u5c0f，\u524d\u63d0\u662f\u641c\u7d22\u77e5\u9053\u6b64\u524d\u7684\u66f4\u65b0\u786e\u7acb\u4e86\u4ec0\u4e48\u3002\u6bcf\u6b21\u8bc4\u4f30\u90fd\u662f\u5bf9\u540c\u4e00\u4e2a\u6709\u9650 evolve \u96c6\u7684\u53c8\u4e00\u6b21\u81ea\u9002\u5e94\u89c2\u5bdf；\u53cd\u590d\u68c0\u9a8c\u65e9\u671f\u8f6e\u6b21\u5df2\u7ecf\u8bc1\u4f2a\u8fc7\u7684\u5047\u8bbe，\u53ea\u4f1a\u6d88\u8017\u641c\u7d22\u5bb9\u91cf\u800c\u4e0d\u589e\u52a0\u6709\u7528\u8bc1\u636e（Dwork et al., 2015）\u3002\u56e0\u6b64 RRSI \u4e3a\u6bcf\u4e2a\u88ab\u8bc4\u4f30\u7684\u5019\u9009\u8bb0\u5f55：\u5b83\u4fee\u6539\u7684\u7ec4\u4ef6\u3001\u5b83\u68c0\u9a8c\u7684\u5047\u8bbe\u3001\u6e90\u7801 diff\u3001\u7531\u6b64\u4ea7\u751f\u7684\u5206\u6570\u4e0e\u6210\u672c\u53d8\u5316\u3001\u4ee5\u53ca\u8be5\u5019\u9009\u662f\u5426\u88ab\u63a5\u53d7\u3002proposer \u5728\u540e\u7eed\u8f6e\u6b21\u4ee5\u8fd9\u6bb5\u5386\u53f2\u4e3a\u6761\u4ef6：\u88ab\u62d2\u7edd\u7684\u673a\u5236\u4ecd\u662f\u8d1f\u9762\u8bc1\u636e，\u6210\u529f\u7684\u673a\u5236\u4fdd\u7559\u660e\u786e\u4fe1\u7528\u3002\u968f\u7740\u540e\u671f\u8f6e\u6b21\u5141\u8bb8\u6bcf\u4e2a\u5019\u9009\u5305\u542b\u66f4\u5c11\u7f16\u8f91，\u8bc6\u522b"\u662f\u54ea\u4e00\u5904\u6539\u52a8\u5bfc\u81f4\u4e86\u89c2\u6d4b\u5230\u7684\u63d0\u5347"\u4e5f\u53d8\u5f97\u66f4\u5bb9\u6613\u3002<\/p>\n
\u7ed3\u6784\u5316\u63a2\u7d22\u3002 \u540c\u4e00\u6bb5\u5386\u53f2\u4e5f\u63ed\u793a proposer \u662f\u5426\u574d\u7f29\u5230\u4e86\u4e00\u4e2a\u72ed\u7a84\u7684\u7f16\u8f91\u5bb6\u65cf\u2014\u2014\u4f8b\u5982\u53cd\u590d\u91cd\u5199\u63d0\u793a\u8bcd，\u5374\u4ece\u4e0d\u89e6\u78b0\u667a\u80fd\u4f53\u7684\u7ed3\u6784\u673a\u5236\u3002\u5f53\u8fc7\u53bb 𝑤 \u8f6e\u7684\u8fdb\u5c55\u4ecd\u843d\u5728\u7ecf\u9a8c\u566a\u58f0\u5e26 𝛿 \u4e4b\u5185\u65f6，\u6211\u4eec\u8ba4\u4e3a\u641c\u7d22\u505c\u6ede\u3002\u505c\u6ede\u671f\u95f4，\u4e00\u5c0f\u90e8\u5206\u63d0\u51fa\u9884\u7b97\u88ab\u9884\u7559\u7ed9\u672c\u6b21\u8fd0\u884c\u4e2d\u5c1a\u672a\u52a8\u7528\u8fc7\u7684\u7ec4\u4ef6\u3002\u8fd9\u8d77\u5230\u7c7b\u4f3c\u591a\u6837\u6027\/\u71b5\u6b63\u5219\u5316\u7684\u4f5c\u7528：\u5728\u4e0d\u6539\u53d8 harness \u88ab\u5141\u8bb8\u5305\u542b\u54ea\u4e9b\u673a\u5236\u7684\u524d\u63d0\u4e0b，\u628a\u6709\u9650\u7684\u63d0\u51fa\u5bb9\u91cf\u91cd\u5b9a\u5411\u5230\u63a2\u7d22\u4e0d\u8db3\u7684\u673a\u5236（Haarnoja et al., 2018）\u3002<\/p>\n
3.3\u3000\u6b63\u5219\u5316\u5019\u9009\u9009\u62e9<\/h4>\n
\u6807\u51c6 harness \u8fdb\u5316\u4f1a\u63a8\u4e3e\u6d4b\u5f97\u5206\u6570\u6700\u9ad8\u7684\u5019\u9009，\u54ea\u6015\u8be5\u5206\u6570\u53cd\u6620\u7684\u662f\u663e\u5f0f\u6cc4\u6f0f\u3001\u968f\u673a\u6ce2\u52a8\u6216\u6602\u8d35\u7684\u81a8\u80c0\u3002RRSI \u4fdd\u7559\u540c\u4e00\u4e2a\u7ecf\u9a8c\u76ee\u6807，\u4f46\u6b63\u5219\u5316"\u54ea\u4e9b\u5019\u9009\u88ab\u5141\u8bb8\u6210\u4e3a\u6c38\u4e45\u72b6\u6001"\u3002\u4e00\u4e2a\u5019\u9009\u5fc5\u987b\u5148\u6ee1\u8db3\u82e5\u5e72**\u975e\u8865\u507f\u6027（non-compensatory）**\u51c6\u5219，\u5b83\u7684\u5206\u6570\u624d\u8db3\u4ee5\u8bc1\u660e\u66ff\u6362 incumbent \u662f\u5408\u7406\u7684\u3002<\/p>\n
\u6cc4\u6f0f\u7b5b\u67e5（Leakage Screening）\u3002 \u5728\u5b8c\u6574\u8bc4\u4f30\u4e4b\u524d，\u4e00\u4e2a critic \u9605\u8bfb\u6bcf\u4e2a\u5019\u9009 diff，\u62d2\u7edd\u90a3\u4e9b\u663e\u5f0f\u7f16\u7801\u4efb\u52a1\u540d\u3001\u5b9e\u4f53\u540d\u3001\u4efb\u52a1\u7279\u5b9a\u53d6\u503c\u3001\u7b54\u6848\u6216\u5176\u4ed6 evolve benchmark \u4e13\u5c5e\u903b\u8f91\u7684\u7f16\u8f91，\u4ee5\u53ca\u90a3\u4e9b\u6dfb\u52a0\u65e0\u7528\u673a\u68b0\u7ed3\u6784（inert machinery）\u7684\u7f16\u8f91\u3002\u7b5b\u67e5\u9488\u5bf9\u7684\u662f"\u7279\u5b9a benchmark \u7684\u5185\u5bb9"，\u800c\u975e\u7279\u5b9a harness \u7ec4\u4ef6：\u901a\u7528\u7684\u63d0\u793a\u8bcd\u6216\u5de5\u5177\u63cf\u8ff0\u6539\u8fdb\u4ecd\u662f\u5408\u6cd5\u5019\u9009\u3002\u5728\u8bc4\u4f30\u524d\u7b5b\u67e5\u5f88\u91cd\u8981，\u56e0\u4e3a\u4e00\u4e2a\u6cc4\u6f0f\u7684\u5019\u9009\u6c38\u8fdc\u62ff\u4e0d\u5230\u90a3\u79cd\u4f1a\u8ba9\u5b83\u5728\u540e\u7eed\u8f6e\u6b21\u663e\u5f97\u6709\u5438\u5f15\u529b\u7684\u3001\u88ab\u5439\u9ad8\u7684 evolve \u96c6\u5206\u6570\u3002<\/p>\n
\u7a33\u5b9a\u6027\u611f\u77e5\u63a5\u53d7（Stability-Aware Acceptance）\u3002 \u5728\u5e26\u566a\u8bc4\u4f30\u4e2d\u53cd\u590d\u9009\u62e9，\u4f1a\u628a\u968f\u673a\u8d62\u5bb6\u56fa\u5316\u4e3a\u6c38\u4e45\u641c\u7d22\u72b6\u6001\u3002\u8fdb\u5316\u5f00\u59cb\u524d，\u6211\u4eec\u53cd\u590d\u8bc4\u4f30\u672a\u6539\u52a8\u7684\u57fa\u5ea7 harness，\u4f30\u8ba1\u4e00\u4e2a\u7ecf\u9a8c\u566a\u58f0\u5e26 𝛿\u3002\u8bb0 𝑆\u2605 \u4e3a\u8fc4\u4eca\u89c2\u5bdf\u5230\u7684\u6700\u4f73 evolve \u96c6\u5206\u6570\u3002\u5019\u9009\u5fc5\u987b\u6ee1\u8db3\u566a\u58f0\u8c03\u6574\u540e\u7684\u5e95\u7ebf：<\/p>\n
\u015c(𝐻\u2032) \u2265 𝑆\u2605 \u2212 𝛿\u3000\u3000(5)<\/p>\n
\u8be5\u5e95\u7ebf\u9632\u6b62\u641c\u7d22\u6cbf\u7740\u4e00\u4e32\u5404\u81ea\u90fd\u5c0f\u5230\u53ef\u88ab\u8bef\u8ba4\u4e3a\u566a\u58f0\u7684\u9000\u6b65\u4e00\u8def\u5411\u4e0b\u8d70\u3002\u66f4\u5e7f\u4e49\u5730，\u5b83\u8ba9\u9009\u62e9\u5bf9"\u5728\u540c\u4e00 evolve \u96c6\u4e0a\u53cd\u590d\u968f\u673a\u8bc4\u4f30\u6240\u5f15\u8d77\u7684\u6ce2\u52a8"\u4fdd\u6301\u4fdd\u5b88（Dwork et al., 2015）\u3002<\/p>\n
Ridge\/𝐿\u2082 \u5f0f\u590d\u6742\u5ea6\u611f\u77e5\u63a5\u53d7\u3002 \u5bf9\u76f8\u5bf9\u5f53\u524d harness 𝐻\u209c \u7684\u5019\u9009 𝐻\u2032，\u4ee4：<\/p>\n
\u0394𝑆 = \u015c(𝐻\u2032) \u2212 \u015c(𝐻\u209c)，\u3000\u3000\u0394𝐶 = ( \u0108(𝐻\u2032) \u2212 \u0108(𝐻\u209c) ) \/ \u0108(𝐻\u209c)\u3000\u3000(6)<\/p>\n
\u5bf9\u63d0\u5347\u8d85\u51fa\u566a\u58f0\u5e26（\u0394𝑆 > 𝛿）\u7684\u5019\u9009，\u6211\u4eec\u8981\u6c42：<\/p>\n
\u0394𝐶 \u2264 𝛽\u2080 + 𝛽\u2081\u00b7\u0394𝑆\u3000\u3000(7)<\/p>\n
\u8fd9\u91cc 𝛽\u2080 \u8bbe\u5b9a\u4e86"\u53ef\u5ffd\u7565\u7684\u5206\u6570\u63d0\u5347"\u6240\u5bb9\u5fcd\u7684\u6210\u672c\u589e\u52a0；𝛽\u2081 \u5219\u63a7\u5236\u968f\u7740\u6d4b\u5f97\u6539\u8fdb\u589e\u5927，\u5141\u8bb8\u8ffd\u52a0\u591a\u5c11\u989d\u5916\u6210\u672c\u3002\u8fd9\u4e9b\u503c\u5728 evolve \u96c6\u4e0a\u9009\u5b9a，\u7136\u540e\u5bf9\u6240\u6709\u8fc1\u79fb\u8bc4\u4f30\u56fa\u5b9a\u4e0d\u53d8\u3002\u4e8e\u662f，\u989d\u5916\u7684\u63a8\u7406\u6210\u672c\u5fc5\u987b\u7531\u53ef\u6d4b\u7684\u6027\u80fd\u6539\u8fdb\u6765\u80cc\u4e66\u3002\u8fd9\u4e2a\u8fc7\u7a0b\u7c7b\u6bd4 Ridge\/𝐿\u2082 \u5f0f\u6536\u7f29：\u5b83\u6291\u5236\u89e3\u6574\u4f53\u5e45\u5ea6\u7684\u65e0\u7ea6\u675f\u589e\u957f\u2014\u2014\u5728\u6211\u4eec\u8fd9\u91cc，\u89e3\u7684\u5e45\u5ea6\u7531 harness \u7684\u805a\u5408\u8d44\u6e90\u5360\u7528\u8868\u5f81\u3002\u4f5c\u4e3a\u4e00\u79cd\u6536\u7f29\u65b9\u6cd5，\u5b83\u4e0d\u8981\u6c42\u4e3a\u4e86\u7a00\u758f\u800c\u5254\u9664\u4efb\u4f55\u7279\u5b9a\u7ec4\u4ef6\u3002\u6211\u4eec\u7528\u7b56\u7565 token \u6210\u672c\u4f5c\u4e3a\u8be5\u8d44\u6e90\u5360\u7528\u7684\u7edf\u4e00\u53ef\u6d4b\u4ee3\u7406\u3002\u8fd9\u662f\u5bf9 Ridge"\u975e\u7a00\u758f\u5316\u590d\u6742\u5ea6\u63a7\u5236"\u7684\u7c7b\u6bd4\u3002\u5173\u4e8e"\u6d4b\u5f97\u53d8\u5316\u843d\u5728\u566a\u58f0\u5e26\u5185"\u7684\u5019\u9009，\u5176\u8be6\u7ec6\u89c4\u5219\u89c1\u9644\u5f55 C.3\u3002<\/p>\n
Lasso\/𝐿\u2081 \u5f0f\u7ed3\u6784\u5316\u526a\u679d\u3002 \u5f0f (4) \u7684\u9000\u706b\u9884\u7b97\u7a00\u758f\u5316\u4e86\u6bcf\u4e00\u6b21\u66f4\u65b0；\u526a\u679d\u5219\u7a00\u758f\u5316\u88ab\u4fdd\u7559\u4e0b\u6765\u7684 harness\u3002RRSI \u8ffd\u8e2a\u8fd1\u671f\u52a8\u7528\u8fc7\u7684\u7ec4\u4ef6\u5728\u4e00\u4e2a\u56fa\u5b9a\u526a\u679d\u7a97\u53e3\u5185\u662f\u5426\u4ea7\u751f\u4e86\u4e25\u683c\u4e3a\u6b63\u7684\u6d4b\u5f97\u589e\u76ca\u3002\u59cb\u7ec8\u65e0\u4ea7\u51fa\u7684\u7ec4\u4ef6\u4f1a\u5728\u540e\u7eed\u8f6e\u6b21\u88ab\u62a5\u544a\u7ed9 proposer \u4f5c\u4e3a\u5220\u9664\u76ee\u6807\u3002\u8fd9\u4e2a\u8fc7\u7a0b\u6a21\u4eff Lasso\/𝐿\u2081 \u5f0f\u7a00\u758f\u5316：\u7f3a\u4e4f\u6548\u7528\u8bc1\u636e\u7684\u673a\u5236\u88ab\u6574\u4f53\u79fb\u9664，\u4e8e\u662f\u4fdd\u7559\u7684 harness \u5728\u7ed3\u6784\u4e0a\u53d8\u5f97\u66f4\u7a00\u758f，\u800c\u4e0d\u53ea\u662f\u805a\u5408\u4e0a\u66f4\u4fbf\u5b9c\u3002\u8fd9\u79cd\u5bf9\u5e94\u540c\u6837\u662f\u5b9a\u6027\u7684：Lasso \u901a\u8fc7 𝐿\u2081 \u6b63\u5219\u5316\u51cf\u5c11\u53c2\u6570\u6570\u91cf，\u800c\u6211\u4eec\u7684\u526a\u679d\u89c4\u5219\u6839\u636e\u89c2\u5bdf\u5230\u7684\u8d21\u732e\u5220\u9664\u79bb\u6563\u7684 harness \u7ec4\u4ef6\u3002\u5171\u540c\u76f4\u89c9\u662f"\u9009\u62e9\u6027\u7a00\u758f\u5316"：\u4e00\u4e2a\u673a\u5236\u5fc5\u987b\u6301\u7eed\u6323\u5f97\u81ea\u5df1\u7684\u4f4d\u7f6e，\u800c\u4e0d\u80fd\u56e0\u4e3a"\u53ea\u770b\u5206\u6570\u7684\u8fdb\u5316\u6ca1\u6709\u52a8\u673a\u5220\u9664\u5b83"\u5c31\u4e00\u76f4\u8d56\u7740\u4e0d\u8d70（Hastie et al., 2009）\u3002<\/p>\n
\n4. \u5b9e\u9a8c<\/h3>\n
\u6211\u4eec\u5728\u6a2a\u8de8\u4e09\u4e2a\u9886\u57df\u7684\u516b\u4e2a benchmark \u4e0a\u8bc4\u4f30 RRSI：\u7f16\u7801\u9886\u57df\u4e3a Terminal-Bench 2.1 \u4e0e SWE-bench Verified；\u667a\u80fd\u4f53\u5de5\u4f5c\u533a\u4efb\u52a1\u4e3a Harvey LAB\u3001JobBench\u3001GDPval\u3001APEX-Agents；\u5de5\u7a0b\u8bbe\u8ba1\u4e3a EngDesign \u4e0e Frontier-Eng\u3002\u5b9e\u9a8c\u56de\u7b54\u4ee5\u4e0b\u95ee\u9898：1) RRSI \u4e0e\u6700\u5148\u8fdb\u7684 harness \u8fdb\u5316\u65b9\u6cd5\u76f8\u6bd4\u5982\u4f55？2) \u63d0\u5347\u80fd\u5426\u8fc1\u79fb\u5230\u5206\u5e03\u5185\u7559\u51fa\u4efb\u52a1，\u4ee5\u53ca\u641c\u7d22\u4ece\u672a\u89c1\u8fc7\u7684\u5206\u5e03\u5916 benchmark？3) \u5404\u7ec4\u4ef6\u7684\u8d21\u732e\u5982\u4f55？4) harness \u5728\u4e00\u6b21\u8fd0\u884c\u4e2d\u5982\u4f55\u6f14\u53d8？\u4ee3\u4ef7\u662f\u591a\u5c11 token？<\/p>\n
4.1\u3000\u5b9e\u9a8c\u8bbe\u7f6e<\/h4>\n
\u73af\u5883\u3002 \u6211\u4eec\u5728\u4e09\u7c7b\u4efb\u52a1\u4e0a\u8fdb\u5316 harness：\u7f16\u7801\u4efb\u52a1\u3001\u667a\u80fd\u4f53\u5de5\u4f5c\u533a\u4efb\u52a1\u3001\u5de5\u7a0b\u8bbe\u8ba1\u4efb\u52a1\u3002<\/p>\n
- \n
- \u7f16\u7801：Terminal-Bench 2.1（Merrill et al., 2026）\u662f\u4e00\u5957 89 \u4e2a\u5bb9\u5668\u5316\u7ec8\u7aef\u4efb\u52a1，\u667a\u80fd\u4f53\u9a71\u52a8\u771f\u5b9e shell，\u7531\u4efb\u52a1\u81ea\u5e26\u7684\u5355\u5143\u6d4b\u8bd5\u9a8c\u8bc1\u3002<\/li>\n
- \u667a\u80fd\u4f53\u5de5\u4f5c\u533a：Harvey LAB（Harvey AI, 2026）\u662f\u6a2a\u8de8 25 \u4e2a\u6cd5\u5f8b\u6267\u4e1a\u9886\u57df\u7684\u6cd5\u5f8b\u5de5\u4f5c benchmark，\u5212\u5206\u4e3a 120 \u4e2a\u4efb\u52a1\u7684\u56fa\u5b9a evolve \u96c6\u4e0e 40 \u4e2a\u4efb\u52a1\u7684\u5e72\u51c0\u5206\u5e03\u5185\u7559\u51fa\u96c6\u3002<\/li>\n
- \u5de5\u7a0b\u8bbe\u8ba1：EngDesign（Guo et al., 2025）\u63d0\u4f9b 61 \u4e2a\u8bbe\u8ba1\u4efb\u52a1，\u6bcf\u4e2a\u7531\u5176\u81ea\u5e26\u7684\u51bb\u7ed3\u6a21\u62df\u5668（\u800c\u975e\u88c1\u5224\u6a21\u578b）\u6253\u5206\u3002<\/li>\n<\/ul>\n
\u4e3a\u6d4b\u8bd5\u6cdb\u5316，\u6211\u4eec\u989d\u5916\u5728\u5206\u5e03\u5916（OOD）\u7559\u51fa benchmark \u4e0a\u8bc4\u4f30：\u7f16\u7801\u4efb\u52a1\u4e0a\u7528 SWE-bench Verified（Jimenez et al., 2024）\u505a\u4ed3\u5e93\u7ea7 bug \u4fee\u590d；\u667a\u80fd\u4f53\u5de5\u4f5c\u533a\u4e0a\u7528 JobBench（Li et al., 2026）\u3001GDPval（Patwardhan et al., 2026）\u3001APEX-Agents（Vidgen et al., 2026）；\u5de5\u7a0b\u8bbe\u8ba1\u4e0a\u7528 Frontier-Eng（Chi et al., 2026）\u3002<\/p>\n
\u57fa\u7ebf\u3002 \u6211\u4eec\u5bf9\u6bd4\u6bcf\u4e2a\u8fd0\u884c\u90fd\u4ece\u4e2d\u51fa\u53d1\u7684\u672a\u8fdb\u5316\u57fa\u5ea7 harness 𝐻\u2080，\u4ee5\u53ca\u56db\u79cd\u8fd1\u671f harness \u8fdb\u5316\u65b9\u6cd5：Meta-Harness（Lee et al., 2026b）\u3001AHE（Lin et al., 2026a）\u3001TTHE（Nie et al., 2026）\u3001HarnessX（Chen et al., 2026）\u3002\u6240\u6709\u57fa\u7ebf\u4ece\u540c\u4e00\u4e2a 𝐻\u2080 \u51fa\u53d1，\u5171\u4eab\u51bb\u7ed3\u7b56\u7565\u3001evolve \u96c6\u4e0e\u5019\u9009\u9884\u7b97\u3002\u57fa\u7ebf\u7ec6\u8282\u89c1\u9644\u5f55 B\u3002<\/p>\n
\u5b9e\u73b0\u7ec6\u8282\u3002 \u4e09\u4e2a\u9886\u57df\u5168\u7a0b\u51bb\u7ed3\u7684\u7b56\u7565\u5747\u4e3a Claude Opus 4.8（Anthropic, 2026a）\u3002proposer\u3001\u64b0\u5199\u8de8\u8f6e\u5931\u8d25\u53cd\u9988\u7684 analyst\u3001\u4ee5\u53ca\u6cc4\u6f0f critic \u4e5f\u90fd\u662f Claude Opus 4.8\u3002\u6240\u7528\u57fa\u5ea7 harness：\u7f16\u7801\u4e3a Terminus-2（Merrill et al., 2026）；\u5728 MCP \u5de5\u5177\u7f51\u5173\u4e0a\u7684 ReAct \u5faa\u73af（Yao et al., 2022）；\u52a8\u6001\u5de5\u5177\u5e26（dynamic toolbelt, Vidgen et al., 2026）；\u4ee5\u53ca ReSum \u5f0f\u4e0a\u4e0b\u6587\u7ba1\u7406（\u7528\u4e8e Harvey LAB \u4e0e EngDesign）\u3002\u5176\u4f59\u8d85\u53c2\u89c1\u9644\u5f55 D.1\u3002<\/p>\n
4.2\u3000\u4e3b\u8981\u7ed3\u679c<\/h4>\n
RRSI \u5728\u4e09\u4e2a\u9886\u57df\u4e2d\u3001\u5bf9 evolve \u96c6\u4e4b\u5916\u7684\u6bcf\u4e00\u4e2a split \u90fd\u6709\u63d0\u5347\u3002\u56fe 3 \u628a\u6bcf\u4e2a\u6570\u5b57\u90fd\u5bf9\u7167"\u540c\u4e00\u65f6\u95f4\u7a97\u5185\u6d4b\u5f97\u7684\u672a\u8fdb\u5316\u57fa\u5ea7 harness 𝐻\u2080"，\u56e0\u6b64\u6ca1\u6709\u4efb\u4f55\u63d0\u5347\u53ef\u5f52\u56e0\u4e8e\u8bc4\u4f30\u57fa\u7840\u8bbe\u65bd\u7684\u6f02\u79fb\u3002evolve \u96c6\u63d0\u5347\u4e3a：Terminal-Bench 2.1 \u4e0a +6.0\u3001EngDesign \u4e0a +4.9\u3001Harvey LAB \u4e0a +1.1\u3002\u771f\u6b63\u8981\u7d27\u7684\u662f harness \u79bb\u5f00\u8fd9\u4e9b split \u4e4b\u540e\u8fd8\u5269\u4ec0\u4e48：SWE-bench Verified \u63d0\u5347 1.8 \u5206（\u5c3d\u7ba1\u4ed3\u5e93\u7ea7 bug \u4fee\u590d\u4ece\u672a\u88ab\u6253\u5206\u8fc7）；Harvey LAB \u7684\u5206\u5e03\u5185\u7559\u51fa split \u63d0\u5347 2.3；\u4e09\u4e2a\u5206\u5e03\u5916\u667a\u80fd\u4f53 benchmark \u63d0\u5347 3.5～4.7 \u5206（7.2%～13.1%）；Frontier-Eng \u63d0\u5347 4.3 \u4e2a Medal \u5206，\u76f8\u5bf9\u63d0\u5347 24.3%\u3002\u6ca1\u6709\u4efb\u4f55\u7559\u51fa split \u5728\u4efb\u4f55\u5730\u65b9\u9000\u6b65\u2014\u2014\u800c\u9000\u6b65\u6b63\u662f"\u6b7b\u8bb0 harness"\u4f1a\u4ea7\u51fa\u7684\u5931\u8d25\u3002<\/p>\n
RRSI \u5728\u6240\u6709\u7559\u51fa\u6570\u636e\u96c6\u4e0a\u4e00\u81f4\u4f18\u4e8e\u57fa\u7ebf\u3002\u8868 1 \u8ba9\u56db\u79cd\u6b64\u524d\u65b9\u6cd5\u4ece\u540c\u4e00\u4e2a 𝐻\u2080 \u51fa\u53d1\u3001\u5728\u540c\u4e00\u4e2a evolve split \u4e0a\u3001\u7528\u76f8\u540c\u5019\u9009\u9884\u7b97\u8fd0\u884c\u3002\u5b83\u4eec\u6bcf\u4e00\u4e2a\u5728 evolve \u96c6\u4e0a\u90fd\u8868\u73b0\u826f\u597d\u3002\u5206\u5e03\u5185\u7559\u51fa split \u4e0a\u7684\u8868\u73b0\u4e5f\u76f8\u5f53\u63a5\u8fd1\u3002\u5206\u91ce\u51fa\u73b0\u5728\u5206\u5e03\u5916，\u800c\u90a3\u91cc\u7684\u6392\u540d\u53cd\u8f6c\u4e86：\u5728 evolve split \u4e0a\u6700\u5f3a\u7684\u57fa\u7ebf Meta-Harness，\u5728\u5206\u5e03\u5916\u5747\u503c\u4e0a\u53ea\u52a0\u4e86 0.9 \u5206；HarnessX \u56de\u5230\u57fa\u5ea7\u6c34\u5e73；AHE \u4e0e TTHE \u5219\u4f4e\u4e8e\u5b83\u4eec\u51fa\u53d1\u65f6\u7684 harness，TTHE \u4f4e 1.7 \u5206\u3002RRSI \u5728\u6240\u6709\u88ab\u8fdb\u5316\u8fc7\u7684 harness \u4e2d evolve \u96c6\u63d0\u5347\u6700\u5c0f，\u5374\u662f\u552f\u4e00\u4e00\u4e2a\u628a\u5206\u5e03\u5916\u5747\u503c\u62c9\u9ad8\u5230\u8d85\u51fa 𝐻\u2080 \u4e00\u5206\u4ee5\u4e0a\u7684\u2014\u201443.6 \u5bf9 39.7，\u8fd9\u6b63\u662f\u6b63\u5219\u5316\u5668\u88ab\u8bbe\u8ba1\u6765\u8fbe\u6210\u7684\u4ea4\u6362\u3002<\/p>\n
![\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260930025718-6abc7a8e37040.png\")
<\/p>\n\u56fe 3｜\u4e09\u4e2a\u9886\u57df\u7684\u4e3b\u8981\u7ed3\u679c\u3002\u7070\u8272\u4e3a 𝐻\u2080（\u672a\u8fdb\u5316 harness），\u6d45\u84dd\u4e3a RRSI"\u88ab\u6253\u5206\u7684 split"，\u6df1\u84dd\u4e3a RRSI"\u4ece\u672a\u89c1\u8fc7\u7684 split"\u3002<\/p>\n
\n
Terminal-Bench 2.1 EvolveSWE-bench Verified OODHarvey LAB EvolveHarvey LAB ID \u7559\u51faJobBench OODGDPval OODAPEX-Agents OODEngDesign EvolveFrontier-Eng OOD<\/tr>\n \n\n 𝐻\u2080<\/td>\n 74.2<\/td>\n 82.0<\/td>\n 89.4<\/td>\n 86.9<\/td>\n 36.0<\/td>\n 48.8<\/td>\n 34.2<\/td>\n 50.0<\/td>\n 17.7<\/td>\n<\/tr>\n \n RRSI<\/td>\n 80.2 (+6.0)<\/td>\n 83.8 (+1.8)<\/td>\n 90.5 (+1.1)<\/td>\n 89.2 (+2.3)<\/td>\n 40.7 (+4.7)<\/td>\n 52.3 (+3.5)<\/td>\n 37.9 (+3.7)<\/td>\n 54.9 (+4.9)<\/td>\n 22.0 (+4.3)<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n \u8868 1｜\u5728\u667a\u80fd\u4f53\u5de5\u4f5c\u533a\u4efb\u52a1\u4e0a\u4e0e\u6b64\u524d harness \u8fdb\u5316\u65b9\u6cd5\u7684\u5bf9\u6bd4\u3002<\/p>\n
\n
\u65b9\u6cd5Harvey LAB (Evolve)Harvey LAB (ID \u7559\u51fa)JobBenchGDPvalAPEX-Agents<\/tr>\n \n\n 𝐻\u2080（\u4e0d\u8fdb\u5316）<\/td>\n 89.4<\/td>\n 86.9<\/td>\n 36.0<\/td>\n 48.8<\/td>\n 34.2<\/td>\n<\/tr>\n \n Meta-Harness (Lee et al., 2026b)<\/td>\n 93.0<\/td>\n 89.2<\/td>\n 37.1<\/td>\n 49.1<\/td>\n 35.7<\/td>\n<\/tr>\n \n AHE (Lin et al., 2026a)<\/td>\n 90.7<\/td>\n 88.7<\/td>\n 37.2<\/td>\n 47.2<\/td>\n 33.1<\/td>\n<\/tr>\n \n TTHE (Nie et al., 2026)<\/td>\n 91.1<\/td>\n 88.5<\/td>\n 35.2<\/td>\n 47.0<\/td>\n 31.7<\/td>\n<\/tr>\n \n HarnessX (Chen et al., 2026)<\/td>\n 91.8<\/td>\n 89.1<\/td>\n 36.3<\/td>\n 48.5<\/td>\n 34.3<\/td>\n<\/tr>\n \n RRSI（\u672c\u6587）<\/td>\n 90.5<\/td>\n 89.2<\/td>\n 40.7<\/td>\n 52.3<\/td>\n 37.9<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n \u8fd9\u79cd\u8fc1\u79fb\u4e0d\u662f"\u88c1\u5224\u6253\u5206"\u6216"\u5171\u4eab\u4efb\u52a1\u683c\u5f0f"\u9020\u6210\u7684\u5047\u8c61\u3002Harvey LAB\u3001JobBench\u3001GDPval \u90fd\u7531\u88c1\u5224\u6a21\u578b\u6253\u5206，\u56e0\u6b64\u539f\u5219\u4e0a harness \u53ef\u4ee5\u9760"\u5199\u5f97\u8ba8\u88c1\u5224\u559c\u6b22"\u800c\u975e"\u505a\u51fa\u66f4\u597d\u5de5\u4f5c"\u6765\u63d0\u5206\u3002\u5de5\u7a0b\u8bbe\u8ba1\u5b9e\u4f8b\u5835\u6b7b\u4e86\u8fd9\u6761\u8def：\u6bcf\u4e2a EngDesign \u4e0e Frontier-Eng \u4efb\u52a1\u90fd\u7531\u5176\u81ea\u5e26\u7684\u4eff\u771f\u5668\/\u6d4b\u8bd5\u53f0\u6253\u5206，\u6253\u5206\u662f\u786e\u5b9a\u6027\u7684\u2014\u2014\u4e00\u4e2a\u8bbe\u8ba1\u8981\u4e48\u6ee1\u8db3\u65e2\u5b9a\u7ea6\u675f，\u8981\u4e48\u4e0d\u6ee1\u8db3\u3002\u63d0\u5347\u5728\u90a3\u91cc\u539f\u6837\u4fdd\u7559，\u800c\u786e\u5b9a\u6027\u6253\u5206\u4e5f\u628a\u88c1\u5224\u65b9\u5dee\u4ece\u6d4b\u91cf\u4e2d\u5254\u9664\u4e86\u3002<\/p>\n
4.3\u3000\u5206\u6790<\/h4>\n
\u4e3b\u8981\u7ed3\u679c\u786e\u7acb\u4e86"\u8fdb\u5316\u540e\u7684 harness \u80fd\u8fc1\u79fb"；\u672c\u8282\u8ffd\u95ee：\u662f\u4ec0\u4e48\u9020\u5c31\u4e86\u8fd9\u79cd\u6027\u8d28？\u5b83\u662f\u5426\u4f9d\u8d56\u641c\u7d22\u6240\u7528\u7684\u9aa8\u5e72？\u4ee3\u4ef7\u662f\u4ec0\u4e48？\u9664\u975e\u53e6\u4f5c\u8bf4\u660e，\u4e0b\u9762\u6bcf\u6b21\u8fd0\u884c\u90fd\u4f7f\u7528\u667a\u80fd\u4f53\u5de5\u4f5c\u533a\u5b9e\u4f8b，\u5e76\u5171\u4eab\u4e3b\u5b9e\u9a8c\u7684\u57fa\u5ea7 harness\u3001\u7b56\u7565\u3001evolve split\u3001\u8f6e\u6b21\u6570\u4e0e\u5019\u9009\u9884\u7b97，\u4f7f\u5404\u81c2\u53ea\u5728\u88ab\u7814\u7a76\u7684\u90a3\u4e2a\u56e0\u5b50\u4e0a\u4e0d\u540c\u3002<\/p>\n
\u6d88\u878d\u5206\u6790\u3002 \u6211\u4eec\u6d88\u878d\u4e24\u7ec4\u6b63\u5219\u5316\u5668：(i) \u63d0\u51fa\u4fa7\u7ea6\u675f，(ii) \u63a5\u53d7\u4fa7\u7ea6\u675f\u3002\u5982\u8868 2，\u53bb\u6389\u4efb\u4e00\u7ec4\u90fd\u4f1a\u62ac\u9ad8 evolve \u96c6\u5206\u6570\u3001\u964d\u4f4e\u8fc1\u79fb\u3002\u53bb\u6389\u63a5\u53d7\u7ea6\u675f\u65f6，evolve \u96c6\u5206\u6570\u4ece 90.5 \u5347\u5230 91.5，\u800c\u5206\u5e03\u5916\u5747\u503c\u4ece 43.6 \u8dcc\u5230 41.0，token \u6210\u672c\u4e0a\u6da8\u4e00\u534a\u2014\u2014\u8bf4\u660e\u65e0\u7ea6\u675f\u7684\u9009\u62e9\u89c4\u5219\u628a\u5b83\u63a5\u53d7\u7684\u5927\u591a\u6570\u7f16\u8f91\u82b1\u5728\u4e86\u566a\u58f0\u548c\u4e0a\u4e0b\u6587\u4e0a，\u800c\u975e\u673a\u5236\u4e0a\u3002\u53bb\u6389\u63d0\u51fa\u7ea6\u675f，evolve split \u53ea\u635f\u5931 0.2 \u5206，\u5206\u5e03\u5916\u5374\u635f\u5931 1.7 \u5206\u2014\u2014\u8bf4\u660e\u5373\u4f7f\u4ec0\u4e48\u90fd\u4e0d\u62d2\u7edd，\u5f15\u5bfc"\u641c\u7d22\u5f80\u54ea\u91cc\u770b"\u4e5f\u5f88\u8981\u7d27\u3002\u6700\u4e25\u91cd\u7684\u9000\u5316\u6765\u81ea\u4e24\u8005\u90fd\u53bb\u6389：evolve \u96c6\u5206\u6570\u5347\u5230\u6240\u6709\u81c2\u4e2d\u6700\u9ad8\u7684 92.8，\u800c\u5206\u5e03\u5916\u5747\u503c\u505c\u5728 40.3（\u53ea\u6bd4\u672a\u8fdb\u5316 harness \u9ad8\u4e0d\u5230\u4e00\u5206），\u6bcf\u6b21\u8bd5\u9a8c\u8017 380 \u4e07 token，\u800c\u6211\u4eec\u662f 242 \u4e07\u3002<\/p>\n
\u8868 2｜\u5728\u667a\u80fd\u4f53\u5de5\u4f5c\u533a\u4efb\u52a1\u4e0a\u5bf9\u6b63\u5219\u5316\u5668\u7684\u6d88\u878d\u3002OOD Avg. \u4e3a JobBench\u3001GDPval\u3001APEX-Agents \u7684\u5747\u503c\u3002<\/p>\n
\n
\u53d8\u4f53Harvey LAB (Evolve)Harvey LAB (ID \u7559\u51fa)OOD \u5747\u503cToken\/\u8bd5\u9a8c (\u767e\u4e07) \u2193<\/tr>\n \n\n 𝐻\u2080（\u4e0d\u8fdb\u5316）<\/td>\n 89.4<\/td>\n 86.9<\/td>\n 39.7<\/td>\n 1.56<\/td>\n<\/tr>\n \n \u672a\u6b63\u5219\u5316\u8fdb\u5316<\/td>\n 92.8<\/td>\n 88.9<\/td>\n 40.3<\/td>\n 3.80<\/td>\n<\/tr>\n \n \u53bb\u6389\u63d0\u51fa\u4fa7\u6b63\u5219\u5316<\/td>\n 90.7<\/td>\n 88.8<\/td>\n 41.9<\/td>\n 2.69<\/td>\n<\/tr>\n \n \u53bb\u6389\u63a5\u53d7\u4fa7\u6b63\u5219\u5316<\/td>\n 91.5<\/td>\n 88.7<\/td>\n 41.0<\/td>\n 3.59<\/td>\n<\/tr>\n \n RRSI<\/td>\n 90.5<\/td>\n 89.2<\/td>\n 43.6<\/td>\n 2.42<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n RRSI \u4e0d\u7ed1\u5b9a\u67d0\u4e00\u4e2a\u7b56\u7565\u5bb6\u65cf\u3002 \u4e3a\u68c0\u9a8c\u6b63\u5219\u5316 harness \u8fdb\u5316\u7684\u63d0\u5347\u662f\u5426\u4f9d\u8d56\u641c\u7d22\u65f6\u6240\u7528\u7684\u7b56\u7565，\u6211\u4eec\u7528\u4e24\u4e2a\u4e0d\u540c\u5bb6\u65cf\u7684\u7b56\u7565\u6a21\u578b\u72ec\u7acb\u8dd1\u7f16\u7801\u8fdb\u5316：Claude Opus 4.8 \u4e0e Gemini 3.5 Flash（Google, 2026b）\u3002\u5bf9\u6bcf\u4e2a\u7b56\u7565，\u4ece\u540c\u4e00\u4e2a\u7f16\u7801 harness \u51fa\u53d1，\u53ea\u5728 Terminal-Bench 2.1 \u4e0a\u8fdb\u5316，\u7136\u540e\u5728 evolve benchmark \u4e0e SWE-bench Verified \u4e0a\u8bc4\u4f30\u6240\u5f97 harness\u3002\u5982\u8868 3，\u5728 Gemini 3.5 Flash \u4e0b，RRSI \u628a Terminal-Bench 2.1 \u4ece 64.6 \u63d0\u5230 78.7，\u5e76\u628a 2.2 \u5206\u7684\u63d0\u5347\u8fc1\u79fb\u5230 SWE-bench Verified\u3002\u5728 Claude Opus 4.8 \u4e0b\u6a21\u5f0f\u76f8\u540c：Terminal-Bench 2.1 \u4ece 74.2 \u5230 80.2，SWE-bench Verified \u4ece 82.0 \u5230 83.8\u2014\u2014\u5c3d\u7ba1\u66f4\u5f3a\u7684\u7b56\u7565\u4e00\u5f00\u59cb\u5c31\u66f4\u63a5\u8fd1\u4e24\u4e2a\u5957\u4ef6\u7684\u5929\u82b1\u677f，\u63d0\u5347\u7a7a\u95f4\u66f4\u5c0f\u3002\u4e24\u79cd\u60c5\u51b5\u4e0b，harness \u90fd\u5728\u4ece\u672a\u88ab\u6253\u5206\u8fc7\u7684\u672a\u89c1 benchmark \u4e0a\u6539\u8fdb\u4e86，\u8bf4\u660e RRSI \u7684\u6536\u76ca\u5e76\u975e\u7279\u5b9a\u4e8e\u67d0\u4e2a\u9aa8\u5e72\u3002<\/p>\n
\u8868 3｜\u7f16\u7801\u9886\u57df\u7684\u7b56\u7565\u9c81\u68d2\u6027\u3002\u7528\u6bcf\u4e2a\u51bb\u7ed3\u7b56\u7565\u72ec\u7acb\u5728 Terminal-Bench \u4e0a\u8dd1 harness \u8fdb\u5316，\u7136\u540e\u539f\u5c01\u4e0d\u52a8\u5730\u5728 SWE-bench Verified \u4e0a\u8bc4\u4f30\u3002<\/p>\n
\n
\u7b56\u7565Benchmark𝐻\u2080RRSI\u0394<\/tr>\n \n\n Claude Opus 4.8<\/td>\n Terminal-Bench 2.1 (Evolve)<\/td>\n 74.2<\/td>\n 80.2<\/td>\n +6.0<\/td>\n<\/tr>\n \n Claude Opus 4.8<\/td>\n SWE-bench Verified (OOD)<\/td>\n 82.0<\/td>\n 83.8<\/td>\n +1.8<\/td>\n<\/tr>\n \n Gemini 3.5 Flash<\/td>\n Terminal-Bench 2.1 (Evolve)<\/td>\n 64.6<\/td>\n 78.7<\/td>\n +14.1<\/td>\n<\/tr>\n \n Gemini 3.5 Flash<\/td>\n SWE-bench Verified (OOD)<\/td>\n 76.8<\/td>\n 79.0<\/td>\n +2.2<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n \u8fdb\u5316\u540e\u7684 harness \u5728\u641c\u7d22\u4ece\u672a\u7528\u8fc7\u7684\u9aa8\u5e72\u4e0b\u4f9d\u7136\u6709\u5e2e\u52a9\u3002 harness \u662f\u4e00\u4e2a\u7a0b\u5e8f，\u4e0d\u662f\u4e00\u7ec4\u6743\u91cd；\u4e00\u4e2a\u53ea\u5bf9\u5b83\u6240\u88ab\u641c\u7d22\u65f6\u6240\u7528\u7b56\u7565\u6709\u5e2e\u52a9\u7684\u673a\u5236，\u662f\u8be5\u7b56\u7565\u7684\u4ea7\u7269\u800c\u975e\u53ef\u590d\u7528\u673a\u5236\u3002\u6211\u4eec\u628a\u7528 Gemini 3.5 Flash \u8fdb\u5316\u51fa\u7684\u7f16\u7801\u6700\u7ec8 harness，\u539f\u5c01\u4e0d\u52a8\u5730\u7528 Gemini 3.1 Flash Lite（\u4e00\u4e2a\u4ece\u672a\u53c2\u4e0e\u641c\u7d22\u7684\u66f4\u5c0f\u6a21\u578b）\u8bc4\u4f30\u3002\u5982\u8868 4，Terminal-Bench 2.1 \u51c6\u786e\u7387\u4ece 11.2 \u5347\u5230 14.6\u2014\u2014\u76f8\u5bf9\u63d0\u5347 30.4%，\u800c\u5176\u57fa\u5ea7\u5206\u6570\u8fd8\u4e0d\u5230\u641c\u7d22\u7b56\u7565\u7684\u4e94\u5206\u4e4b\u4e00\u3002\u56e0\u6b64\u8fd9\u4e9b\u673a\u5236\u5e76\u4e0d\u4f9d\u8d56\u5b83\u4eec\u88ab\u641c\u7d22\u65f6\u6240\u5904\u7684\u80fd\u529b\u6c34\u5e73，\u5c3d\u7ba1\u7edd\u5bf9\u63d0\u5347\u66f4\u5c0f，\u56e0\u4e3a\u66f4\u5f31\u7684\u9aa8\u5e72\u628a\u66f4\u5c11\u4efb\u52a1\u7559\u5728\u4e86\u4efb\u4f55 harness \u591f\u5f97\u7740\u7684\u8303\u56f4\u5185\u3002<\/p>\n
\u8868 4｜Terminal-Bench 2.1 \u4e0a\u7684\u8de8\u6a21\u578b\u8fc1\u79fb\u3002\u4ee5 Gemini 3.5 Flash \u4e3a\u51bb\u7ed3\u7b56\u7565\u8fdb\u5316\u51fa\u7684 harness，\u539f\u5c01\u4e0d\u52a8\u5730\u7528\u4e00\u4e2a\u4ece\u672a\u53c2\u4e0e\u641c\u7d22\u7684\u66f4\u5f31\u9aa8\u5e72\u8fd0\u884c\u3002<\/p>\n
\n
\u8bc4\u4f30\u7b56\u7565𝐻\u2080RRSI\u0394<\/tr>\n \n\n Gemini 3.5 Flash（\u641c\u7d22\u7b56\u7565）<\/td>\n 64.6<\/td>\n 78.7<\/td>\n +14.1<\/td>\n<\/tr>\n \n Gemini 3.1 Flash Lite（\u672a\u89c1）<\/td>\n 11.2<\/td>\n 14.6<\/td>\n +3.4<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n RRSI \u4ea7\u51fa\u4e86\u6240\u6709\u88ab\u8fdb\u5316 harness \u4e2d\u6700\u8f7b\u91cf\u7684\u4e00\u4e2a\u3002 \u4e24\u4e2a\u6b63\u5219\u5316\u5668\u76f4\u63a5\u4f5c\u7528\u4e8e\u6210\u672c：𝐿\u2081 \u5f0f\u9884\u7b97\u5728"\u672a\u88ab\u4e70\u5355\u7684\u589e\u957f"\u521a\u88ab\u63d0\u51fa\u65f6\u5c31\u62d2\u7edd\u5b83；\u526a\u679d\u89c4\u5219\u5219\u79fb\u9664\u90a3\u4e9b\u6b64\u540e\u4e0d\u518d\u88ab\u4e70\u5355\u7684\u589e\u957f\u3002\u6b64\u524d\u6ca1\u6709\u4efb\u4f55\u65b9\u6cd5\u5e26\u8fd9\u4e24\u4e2a\u7ea6\u675f，\u56fe 4(a) \u663e\u793a\u4e86\u540e\u679c：\u56db\u79cd\u65b9\u6cd5\u5168\u90fd\u843d\u5728 RRSI \u652f\u914d\u7684\u533a\u57df\u2014\u2014\u6bcf\u6b21\u8bd5\u9a8c\u82b1\u66f4\u591a\u7b56\u7565 token，\u5374\u6362\u5f97\u66f4\u4f4e\u7684\u5206\u5e03\u5916\u5747\u503c\u3002AHE \u662f\u6781\u7aef：\u6bcf\u6b21\u8bd5\u9a8c 382 \u4e07 token，\u6bd4\u6211\u4eec\u591a 58%，\u5206\u5e03\u5916\u5374\u4f4e 4.4 \u5206\u3002\u8fd9\u4e00\u6392\u5e8f\u5ef6\u7eed\u5230\u56fe 4(b) \u7684\u8f68\u8ff9\u957f\u5ea6：RRSI \u6bcf\u6b21\u8bd5\u9a8c 26.3 \u6b65，\u6b64\u524d\u65b9\u6cd5\u4e3a 27.3～34.6 \u6b65\u3002\u6ca1\u6709\u54ea\u4e2a\u88ab\u8fdb\u5316 harness \u50cf 𝐻\u2080 \u90a3\u6837\u4fbf\u5b9c（156 \u4e07 token\u300121.2 \u6b65），\u6240\u4ee5\u8fdb\u5316\u786e\u5b9e\u7528\u6d4b\u8bd5\u671f\u8ba1\u7b97\u4e70\u6765\u4e86\u4e00\u90e8\u5206\u63d0\u5347；\u800c\u9884\u7b97\u51b3\u5b9a\u4e86\u4e70\u591a\u5c11\u3002<\/p>\n
![\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260930025719-6abc7a8f48245.png\")
<\/p>\n\u56fe 4｜\u5404\u81c2\u6700\u7ec8 harness \u7684\u6210\u672c，\u5728\u667a\u80fd\u4f53\u5de5\u4f5c\u533a\u5b9e\u4f8b\u7684 evolve split \u4e0a\u6d4b\u5f97\u3002OOD Avg. \u4e3a JobBench\u3001GDPval\u3001APEX-Agents \u5747\u503c\u3002(a) \u9634\u5f71\u533a\u662f RRSI \u652f\u914d\u7684\u4e00\u5207：\u6bcf\u6b21\u8bd5\u9a8c\u82b1\u66f4\u591a\u7b56\u7565 token，\u5206\u5e03\u5916\u5747\u503c\u5374\u66f4\u4f4e\u3002(b) \u8f68\u8ff9\u957f\u5ea6（\u6b65\/\u8bd5\u9a8c）：AHE 34.6\u3001HarnessX 29.5\u3001Meta-Harness 28.7\u3001TTHE 27.3\u3001RRSI 26.3\u3001𝐻\u2080 21.2\u3002<\/p>\n
\n5. \u76f8\u5173\u5de5\u4f5c<\/h3>\n
\u667a\u80fd\u4f53 harness\u3002 \u51b3\u5b9a\u4e00\u4e2a\u667a\u80fd\u4f53\u80fd\u505a\u6210\u4ec0\u4e48\u7684，\u4e0d\u53ea\u662f\u9aa8\u5e72\u6a21\u578b，\u8fd8\u6709 harness：\u524d\u6cbf\u5b9e\u9a8c\u5ba4\u7684\u5de5\u7a0b\u62a5\u544a\u63cf\u8ff0\u4e86\u63d0\u793a\u8bcd\u7ed3\u6784\u3001\u5de5\u5177\u63a5\u53e3\u3001\u4e0a\u4e0b\u6587\u538b\u7f29\u4e0e\u6062\u590d\u903b\u8f91\u5982\u4f55\u51b3\u5b9a\u4e00\u4e2a\u957f\u65f6\u95f4\u8fd0\u884c\u7684\u667a\u80fd\u4f53\u7a76\u7adf\u80fd\u5426\u5b8c\u6210\u4efb\u52a1（Lopopolo, 2026; Rajasekaran, 2026）；\u8fd1\u671f\u5206\u6790\u4e3b\u5f20 harness \u81ea\u8eab\u5c31\u5177\u5907\u7ec4\u5408\u6027\u4e0e\u6cdb\u5316\u80fd\u529b（Wang et al., 2026a; Weng, 2026; Zhang and Khattab, 2026）\u3002\u8bbe\u8ba1\u826f\u597d\u7684 harness \u751a\u81f3\u80fd\u66ff\u4ee3\u89c4\u6a21，\u4ee5\u5c0f\u5f97\u591a\u7684\u6210\u672c\u6062\u590d\u5927\u9aa8\u5e72\u6a21\u578b\u7684\u5927\u90e8\u5206\u80fd\u529b（Yang et al., 2026a）\u3002\u8fd9\u79cd\u5de5\u7a0b overwhelmingly \u662f\u624b\u5de5\u7684，\u800c\u56e0\u4e3a\u6700\u4f73 harness \u7ed1\u5b9a\u4e8e\u7279\u5b9a\u9aa8\u5e72，\u6bcf\u53d1\u4e00\u4e2a\u65b0\u6a21\u578b\u90fd\u8981\u518d\u4ed8\u4e00\u904d\u8fd9\u4efd\u6210\u672c（Huang et al., 2026b）\u3002<\/p>\n
Harness \u8fdb\u5316\u3002 \u6700\u63a5\u8fd1\u7684\u4e00\u6761\u5de5\u4f5c\u7ebf\u628a\u8fd9\u4e2a\u5faa\u73af\u81ea\u52a8\u5316：\u7531 LLM proposer \u91cd\u5199 harness，\u82e5\u7f16\u8f91\u63d0\u5347\u4e86 benchmark \u5206\u6570\u5c31\u4fdd\u7559\u5b83（Chen et al., 2026; Karten et al., 2026b; Lee et al., 2026a,b; Lin et al., 2026a; Liu et al., 2026b; Nie et al., 2026; Zhang et al., 2026a）；\u6216\u662f\u53ea\u8fdb\u5316\u5355\u4e2a\u7ec4\u4ef6，\u5982\u6280\u80fd（Xia et al., 2026a; Yang et al., 2026b）\u3001\u8bb0\u5fc6（Liu et al., 2026a; Ouyang et al., 2026; Tang et al., 2025; Wu et al., 2026）\u6216\u5bf9 rollout \u7684\u504f\u597d\u4fe1\u53f7（Pan et al., 2026）\u3002\u8fd9\u7ee7\u627f\u4e86"\u5728\u7ecf\u9a8c\u9002\u5e94\u5ea6\u4fe1\u53f7\u4e0b\u641c\u7d22\u81ea\u8eab\u4ee3\u7801\u7684\u81ea\u6539\u8fdb\u667a\u80fd\u4f53"\u7684\u673a\u5236\u4e0e\u98ce\u9669（Huang et al., 2026a,c; Wang et al., 2025; Xia et al., 2026b,c; Zhang et al., 2026b,c）\u3002\u8d2f\u7a7f\u59cb\u7ec8，\u641c\u7d22\u90fd\u7531\u5b83\u6240\u9488\u5bf9\u7684\u5957\u4ef6\u5206\u6570\u9a71\u52a8，\u6ca1\u6709\u4efb\u4f55\u9879\u5bf9\u5e94\u6cdb\u5316；\u800c\u4ee3\u4ef7\u5e76\u975e\u5047\u8bbe\u6027\u7684：\u62a5\u544a\u7684\u63d0\u5347\u5e38\u5e38\u5728\u6362\u5957\u4ef6\u540e\u5c31\u4e0d\u6210\u7acb（Huang et al., 2026d; Wang et al., 2026b），delta \u5f52\u56e0\u80fd\u628a"\u5b89\u88c5\u4e86\u53ef\u590d\u7528\u673a\u5236"\u7684\u7f16\u8f91\u4e0e"\u53ea\u662f\u62df\u5408\u4e86\u8fdb\u5316\u4efb\u52a1"\u7684\u7f16\u8f91\u533a\u5206\u5f00（Ding et al., 2026）\u3002\u5e76\u53d1\u5de5\u4f5c\u4e5f\u76f4\u63a5\u9488\u5bf9\u6cdb\u5316，\u8981\u4e48\u628a\u5b83\u4f5c\u4e3a\u641c\u7d22\u7684\u663e\u5f0f\u76ee\u6807（Zhang et al., 2026d），\u8981\u4e48\u7528\u4e00\u4e2a\u4fdd\u6301\u591a\u6837\u6027\u7684\u5019\u9009 harness \u6863\u6848（archive）\u66ff\u4ee3\u8d2a\u5fc3\u9009\u62e9（Luo et al., 2026）\u3002\u6211\u4eec\u7684\u8d21\u732e\u4e0e"\u8fd9\u4e9b\u65b9\u6cd5\u7f16\u8f91\u4ec0\u4e48"\u662f\u6b63\u4ea4\u7684：\u6211\u4eec\u4fdd\u6301\u540c\u6837\u5f00\u653e\u7684\u7f16\u8f91\u7a7a\u95f4，\u8f6c\u800c\u6b63\u5219\u5316\u641c\u7d22\u52a8\u6001\u2014\u2014\u5728\u5b8c\u6574\u8fdb\u5316\u5386\u53f2\u4e0a\u5206\u914d\u4fe1\u7528\u3001\u5728\u6253\u5206\u524d\u8fc7\u6ee4\u4efb\u52a1\u7279\u5b9a\u903b\u8f91\u3001\u6309\u566a\u58f0\u8c03\u6574\u540e\u7684\u57fa\u7ebf\u6765\u51b3\u5b9a\u63a5\u53d7\u2014\u2014\u4e8e\u662f\u6d3b\u4e0b\u6765\u7684\u662f\u673a\u5236，\u800c\u4e0d\u662f\u5bf9\u8fdb\u5316\u5957\u4ef6\u7684\u62df\u5408\u3002<\/p>\n
\n6. \u7ed3\u8bba<\/h3>\n
\u6211\u4eec\u628a\u8fed\u4ee3\u5f0f harness \u8fdb\u5316\u89c6\u4e3a\u667a\u80fd\u4f53\u7cfb\u7edf\u5c42\u9762\u4e00\u79cd\u5b9e\u7528\u7684\u9012\u5f52\u81ea\u6211\u6539\u8fdb，\u5e76\u8868\u660e：\u8fd9\u4e2a\u9012\u5f52\u8fc7\u7a0b\u672c\u8eab\u5c31\u9700\u8981\u6b63\u5219\u5316\u3002\u56e0\u4e3a\u4e00\u4e2a\u6709\u9650 evolve \u96c6\u88ab\u8de8\u8f6e\u6b21\u81ea\u9002\u5e94\u5730\u590d\u7528，\u8868\u9762\u4e0a\u7684\u81ea\u6211\u6539\u8fdb\u53ef\u80fd\u53cd\u6620\u7684\u662f\u9488\u5bf9\u7279\u5b9a benchmark \u7684\u62df\u5408\u3001\u8bc4\u4f30\u566a\u58f0\u6216\u4e0d\u5fc5\u8981\u7684\u590d\u6742\u5ea6，\u800c\u975e\u53ef\u8fc1\u79fb\u7684\u8fdb\u6b65\u3002RRSI \u901a\u8fc7\u5728\u4fdd\u6301 harness \u7f16\u8f91\u7a7a\u95f4\u5f00\u653e\u7684\u540c\u65f6\u3001\u5bf9\u63d0\u51fa\u4e0e\u9009\u62e9\u4e24\u7aef\u90fd\u52a0\u4ee5\u6b63\u5219\u5316\u6765\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\u3002\u5728\u7f16\u7801\u3001\u667a\u80fd\u4f53\u5de5\u4f5c\u533a\u4e0e\u5de5\u7a0b\u8bbe\u8ba1\u4efb\u52a1\u4e0a，\u6240\u5f97 harness \u5728\u7559\u51fa\u4e0e\u8de8 benchmark \u8868\u73b0\u4e0a\u90fd\u6709\u63d0\u5347，\u540c\u65f6\u6bd4\u672a\u6b63\u5219\u5316\u8fdb\u5316\u6d88\u8017\u66f4\u5c11\u63a8\u7406\u6210\u672c\u3002\u8fd9\u4e9b\u7ed3\u679c\u63d0\u793a：\u8981\u901a\u8fc7\u9012\u5f52\u81ea\u6211\u6539\u8fdb\u8ba9\u667a\u80fd\u4f53\u7cfb\u7edf\u8d8a\u6765\u8d8a\u5f3a，\u4e0d\u4ec5\u8981\u63a7\u5236"\u4ec0\u4e48\u53ef\u4ee5\u6539"，\u8fd8\u8981\u63a7\u5236"\u53cd\u590d\u53cd\u9988\u5982\u4f55\u88ab\u8f6c\u5316\u4e3a\u6301\u4e45\u6539\u52a8"\u3002<\/p>\n
\u5c40\u9650\u6027<\/h4>\n
\u6211\u4eec\u7684\u7814\u7a76\u805a\u7126\u4e8e\u51bb\u7ed3\u9aa8\u5e72\u6a21\u578b\u4e0b\u7684 harness \u7ea7\u9012\u5f52\u81ea\u6211\u6539\u8fdb，\u56e0\u6b64\u4e0d\u8ba8\u8bba\u8fdb\u5316\u8fc7\u7a0b\u4e2d\u66f4\u65b0\u6a21\u578b\u6743\u91cd\u7684\u8bbe\u7f6e\u3002\u6b64\u5916，RRSI \u4ecd\u4f9d\u8d56\u4e00\u4e2a\u6709\u9650 evolve \u96c6\u4e0e\u82e5\u5e72\u6b63\u5219\u5316\u8d85\u53c2，\u5176\u6709\u6548\u6027\u53ef\u80fd\u4f9d\u8d56\u53cd\u9988\u4fe1\u53f7\u8d28\u91cf\u4e0e\u6240\u9009\u641c\u7d22\u9884\u7b97\u3002\u6700\u540e，\u5c3d\u7ba1\u6211\u4eec\u8de8\u591a\u4e2a\u9886\u57df\u3001benchmark \u4e0e\u7b56\u7565\u6a21\u578b\u8bc4\u4f30\u4e86\u8fc1\u79fb，\u4ecd\u9700\u8981\u66f4\u5e7f\u6cdb\u7684\u9a8c\u8bc1，\u624d\u80fd\u786e\u5b9a\u8be5\u65b9\u6cd5\u5bf9\u5dee\u5f02\u663e\u8457\u7684\u667a\u80fd\u4f53\u67b6\u6784\u3001\u5de5\u5177\u751f\u6001\u7cfb\u7edf\u4ee5\u53ca\u66f4\u957f\u65f6\u7a0b\u7684\u81ea\u6211\u6539\u8fdb\u8fc7\u7a0b\u6cdb\u5316\u5f97\u5982\u4f55\u3002<\/p>\n
\n\u53c2\u8003\u6587\u732e<\/h3>\n
\u53c2\u8003\u6587\u732e\u6309\u5b66\u672f\u60ef\u4f8b\u4fdd\u7559\u539f\u6587（\u4f5c\u8005\u3001\u673a\u6784\u3001\u53d1\u8868 venue\u3001arXiv \u7f16\u53f7\u4e0e\u94fe\u63a5\u4e0d\u8bd1）\u3002<\/p>\n
- \n
- Anthropic. Introducing Claude Opus 4.8, 2026a. https:\/\/www.anthropic.com\/news\/claude-opus-4-8.<\/li>\n
- Anthropic. Introducing Claude Sonnet 4.6, 2026b. https:\/\/www.anthropic.com\/news\/claude-sonnet-4-6.<\/li>\n
- T. Chen et al. HarnessX: A composable, adaptive, and evolvable agent harness foundry. arXiv:2606.14249, 2026.<\/li>\n
- Y. Chi et al. Frontier-Eng: Benchmarking self-evolving agents on real-world engineering tasks with generative optimization. arXiv:2604.12290, 2026.<\/li>\n
- W. Ding et al. What evolves when we talk about harness evolution? wenwen-d.github.io, 2026.<\/li>\n
- C. Dwork et al. Generalization in adaptive data analysis and holdout reuse. NeurIPS 28, 2015.<\/li>\n
- I. Goodfellow, Y. Bengio, A. Courville. Deep Learning, vol. 1. MIT Press, 2016.<\/li>\n
- Google. Gemini 3.1 Pro, 2026a. https:\/\/deepmind.google\/models\/gemini\/pro\/.<\/li>\n
- Google. Gemini 3.5: frontier intelligence with action, 2026b.<\/li>\n
- X. Guo et al. Toward engineering AGI: Benchmarking the engineering design capabilities of LLMs. NeurIPS, 2025.<\/li>\n
- T. Haarnoja et al. Soft Actor-Critic. ICML, 2018.<\/li>\n
- Harvey AI. Harvey LAB: The legal agent benchmark, 2026.<\/li>\n
- T. Hastie, R. Tibshirani, J. Friedman. The Elements of Statistical Learning, vol. 2. Springer, 2009.<\/li>\n
- C. Huang et al. G-Zero: Self-play for open-ended generation from zero data. arXiv:2605.09959, 2026a.<\/li>\n
- C. Huang et al. EnvHarness: Awakening static worlds for agent learning. arXiv:2608.19880, 2026b.<\/li>\n
- C. Huang et al. R-Zero: Self-evolving reasoning LLM from zero data. ICLR, 2026c.<\/li>\n
- L. Huang et al. Evo-Bench: Can language models improve agent harness? arXiv:2608.09096, 2026d.<\/li>\n
- C. E. Jimenez et al. SWE-bench: Can language models resolve real-world GitHub issues? ICLR, 2024.<\/li>\n
- S. Karten et al. Prime Agent: A self-improving RLM harness. Prime Intellect Blog, 2026a.<\/li>\n
- S. Karten et al. Continual Harness: Online adaptation for self-improving foundation agents. arXiv:2605.09998, 2026b.<\/li>\n
- Z. Ke et al. EvoHarnessBench: Can your agents keep pace with an evolving harness? arXiv:2609.04280, 2026.<\/li>\n
- H. Lee et al. Recursive harness self-improvement. arXiv:2607.15524, 2026a.<\/li>\n
- Y. Lee et al. Meta-Harness: End-to-end optimization of model harnesses. CoLM, 2026b.<\/li>\n
- Y. Li et al. JobBench: Aligning agent work with human will. arXiv:2605.26329, 2026.<\/li>\n
- J. Lin et al. Agentic harness engineering: Observability-driven automatic evolution of coding-agent harnesses. arXiv:2604.25850, 2026a.<\/li>\n
- M. Lin et al. Harness updating is not harness benefit: Disentangling evolution capabilities in self-evolving LLM agents. arXiv:2605.30621, 2026b.<\/li>\n
- J. Liu et al. EvolveMem: Self-evolving memory architecture via auto-research for LLM agents. arXiv:2605.13941, 2026a.<\/li>\n
- Z. Liu et al. Adaptive auto-harness: Sustained self-improvement for agentic system deployment on open-ended task streams. arXiv:2606.01770, 2026b.<\/li>\n
- R. Lopopolo. Harness engineering: leveraging Codex in an agent-first world, 2026.<\/li>\n
- X. Lou et al. AutoHarness: improving LLM agents by automatically synthesizing a code harness. arXiv:2603.03329, 2026.<\/li>\n
- C. Louizos, M. Welling, D. P. Kingma. Learning sparse neural networks through L0 regularization. ICLR, 2018.<\/li>\n
- X. Luo et al. Self-evolving agent harnesses via gated semantic quality-diversity. arXiv:2607.13683, 2026.<\/li>\n
- M. Merrill et al. Terminal-Bench: Benchmarking agents on hard, realistic tasks in command line interfaces. ICLR, 2026.<\/li>\n
- J. Nie et al. TTHE: Test-time harness evolution. arXiv:2607.08124, 2026.<\/li>\n
- J. Niklaus. Don\u2019t train the model, evolve the harness, 2026.<\/li>\n
- S. Ouyang et al. ReasoningBank: Scaling agent self-evolving with reasoning memory. ICLR, 2026.<\/li>\n
- W. Pan et al. Retrospective harness optimization: Improving LLM agents via self-preference over trajectory rollouts. arXiv:2606.05922, 2026.<\/li>\n
- T. Patwardhan et al. GDPval: Evaluating AI model performance on real-world economically valuable tasks. ICLR, 2026.<\/li>\n
- Qwen Team. Qwen3.6-Plus: Towards real world agents, 2026.<\/li>\n
- P. Rajasekaran. Harness design for long-running application development, 2026.<\/li>\n
- RSI-Exam Team. RSI-Exam: Benchmarking recursive self-improvement through executable research, 2026.<\/li>\n
- X. Tang et al. Agent KB: Leveraging cross-domain experience for agentic problem solving. arXiv:2507.06229, 2025.<\/li>\n
- N. Team et al. NeoHorses-1: Towards recursive self-improvement via agentic post-training with routing harness. arXiv:2609.08183, 2026.<\/li>\n
- B. Vidgen et al. APEX-Agents. arXiv:2601.14242, 2026.<\/li>\n
- R. Wang et al. Harness handbook: Making evolving agent harnesses readable, navigable, and editable. arXiv:2607.13285, 2026a.<\/li>\n
- W. Wang et al. Huxley-G\u00f6del machine: Human-level coding agent development by an approximation of the optimal self-improving machine. arXiv:2510.21614, 2025.<\/li>\n
- Y. Wang et al. Rethinking the evaluation of harness evolution for agents. COLM 2026 Lifelong Agents Workshop, 2026b.<\/li>\n
- L. Weng. Harness engineering for self-improvement. lilianweng.github.io, 2026.<\/li>\n
- S. Wu et al. AutoMem: Automated learning of memory as a cognitive skill. arXiv:2607.01224, 2026.<\/li>\n
- P. Xia et al. SkillRL: Evolving agents via recursive skill-augmented reinforcement learning. arXiv:2602.08234, 2026a.<\/li>\n
- P. Xia et al. MetaClaw: Just talk\u2013an agent that meta-learns and evolves in the wild. arXiv:2603.17187, 2026b.<\/li>\n
- P. Xia et al. Agent0: Unleashing self-evolving agents from zero data via tool-integrated reasoning. CoLM, 2026c.<\/li>\n
- C. Yang et al. Better harnesses, smaller models: Building 90% cheaper agents via automated harness adaptation. arXiv:2607.08938, 2026a.<\/li>\n
- Y. Yang et al. SkillOpt: Executive strategy for self-evolving agent skills. arXiv:2605.23904, 2026b.<\/li>\n
- S. Yao et al. ReAct: Synergizing reasoning and acting in language models. arXiv:2210.03629, 2022.<\/li>\n
- A. Zhang, O. Khattab. Language model harnesses are compositional generalizers, 2026.<\/li>\n
- H. Zhang et al. Self-Harness: Harnesses that improve themselves. arXiv:2606.09498, 2026a.<\/li>\n
- J. Zhang et al. Darwin G\u00f6del machine: open-ended evolution of self-improving agents. ICLR, 2026b.<\/li>\n
- J. Zhang et al. HyperAgents. arXiv:2603.19461, 2026c.<\/li>\n
- L. Zhang et al. HarnessCompass: Guiding automatic harness evolution toward generalizable and effective agent harnesses. arXiv:2608.01918, 2026d.<\/li>\n
- Y. Zhang et al. DarwinX: Evolving agent harnesses through natural selection. arXiv:2608.07545, 2026e.<\/li>\n<\/ul>\n
\n\u9644\u5f55<\/h2>\n
A. \u8bc4\u4f30<\/h3>\n
\u6b64\u5904\u8bf4\u660e\u6bcf\u4e2a\u73af\u5883\u5982\u4f55\u8fd0\u884c\u4e0e\u6253\u5206\u3002\u4e00\u4e2a harness \u4e0e\u5176\u57fa\u7ebf\u59cb\u7ec8\u5728\u540c\u4e00\u65f6\u95f4\u7a97\u3001\u540c\u4e00\u5de5\u5177\u73af\u5883\u3001\u540c\u4e00\u88c1\u5224\u3001\u76f8\u540c\u8bd5\u9a8c\u6b21\u6570\u4e0b\u8bc4\u4f30\u3002<\/p>\n
A.1\u3000Terminal-Bench 2.1\u3000\u6bcf\u4e2a\u4efb\u52a1\u662f\u4e00\u4e2a\u5bb9\u5668\u955c\u50cf，\u542b\u4efb\u52a1\u63cf\u8ff0\u3001\u5de5\u4f5c\u76ee\u5f55\u4e0e\u4e00\u7ec4\u5bf9\u667a\u80fd\u4f53\u9690\u85cf\u7684\u5355\u5143\u6d4b\u8bd5（Merrill et al., 2026）\u3002\u667a\u80fd\u4f53\u901a\u8fc7 harness \u9a71\u52a8\u771f\u5b9e shell；\u53ea\u6709\u5728\u667a\u80fd\u4f53\u505c\u6b62\u540e\u8be5\u4efb\u52a1\u81ea\u5e26\u6d4b\u8bd5\u5957\u4ef6\u5168\u90e8\u901a\u8fc7，\u4efb\u52a1\u624d\u7b97\u89e3\u51b3\u2014\u2014\u5956\u52b1\u662f\u7cbe\u786e\u7684，\u65e0\u6cd5\u7528\u4e00\u4e2a"\u770b\u8d77\u6765\u5408\u7406"\u7684\u7b54\u6848\u8499\u6df7\u3002\u62a5\u544a\u7684\u51c6\u786e\u7387\u5373\u4ee5\u6b64\u65b9\u5f0f\u89e3\u51b3\u7684 89 \u4e2a\u4efb\u52a1\u4e2d\u7684\u5360\u6bd4\u3002\u5404\u81c2\u4e4b\u95f4\u5bb9\u5668\u4f1a\u88ab\u62c6\u9664\u91cd\u5efa，\u786e\u4fdd\u72b6\u6001\u4e0d\u4ece\u4e00\u6b21\u8bc4\u4f30\u5e26\u5230\u4e0b\u4e00\u6b21\u3002<\/p>\n
A.2\u3000SWE-bench Verified\u3000\u6bcf\u4e2a\u5b9e\u4f8b\u662f\u4e00\u4e2a\u771f\u5b9e GitHub issue，\u8fde\u540c\u62a5\u544a\u65f6\u7684\u4ed3\u5e93\u5feb\u7167（Jimenez et al., 2024）\u3002\u667a\u80fd\u4f53\u5fc5\u987b\u4ea7\u51fa\u4e00\u4e2a patch，\u5e94\u7528\u5230\u5feb\u7167\u4e0a\u540e\u68c0\u67e5：fail-to-pass \u6d4b\u8bd5\u5fc5\u987b\u4ece\u5931\u8d25\u53d8\u901a\u8fc7，pass-to-pass \u6d4b\u8bd5\u5fc5\u987b\u4fdd\u6301\u901a\u8fc7\u3002\u62a5\u544a\u7684\u89e3\u51b3\u7387\u5373\u540c\u65f6\u6ee1\u8db3\u4e24\u6761\u4ef6\u7684\u5b9e\u4f8b\u5360\u6bd4\u3002<\/p>\n
A.3\u3000Harvey LAB\u3000\u6bcf\u4e2a\u4efb\u52a1\u63d0\u4f9b\u4e00\u4e2a\u542b Word\/Excel\/PDF \u6e90\u6587\u6863\u7684\u6587\u4ef6\u5939，\u8981\u6c42\u667a\u80fd\u4f53\u6309\u7cbe\u786e\u6307\u5b9a\u7684\u6587\u4ef6\u540d\u4ea7\u51fa\u4ea4\u4ed8\u6587\u4ef6（Harvey AI, 2026），\u518d\u6309\u4e25\u683c\u7684\u9010\u6761\u6807\u51c6（rubric）\u6253\u5206\u2014\u2014\u6bcf\u4efb\u52a1 20～100 \u4e2a\u72ec\u7acb\u8bc4\u5224\u7684\u6807\u51c6，\u5b8c\u6574\u8bc4\u4f30\u7ea6 14000 \u6761\u6807\u51c6\u5224\u5b9a\u3002\u6bcf\u6761\u6807\u51c6\u7531\u4e00\u4e2a LLM \u88c1\u5224（Gemini-3.5-Flash）\u5355\u72ec\u8bc4\u5224：\u5b83\u540c\u65f6\u8bfb\u5230\u6240\u4ea7\u51fa\u4ea4\u4ed8\u7269\u4e0e\u90a3\u4e00\u6761\u6807\u51c6\u3002\u4e00\u6b21\u8fd0\u884c\u7684\u5206\u6570\u5373\u6240\u6709\u4efb\u52a1\u4e2d\u901a\u8fc7\u6807\u51c6\u7684\u6bd4\u4f8b，\u56e0\u6b64 rubric \u957f\u7684\u4efb\u52a1\u8d21\u732e\u8bc1\u636e\u66f4\u591a；\u7f3a\u5931\u4ea4\u4ed8\u7269\u4f1a\u4ee4\u5176\u672c\u5e94\u6ee1\u8db3\u7684\u6bcf\u6761\u6807\u51c6\u90fd\u4e0d\u901a\u8fc7，\u800c\u4e0d\u662f\u88ab\u4e22\u5f03\u3002160 \u4e2a\u4efb\u52a1\u4e00\u6b21\u6027\u5212\u5206\u4e3a 120 \u4e2a evolve \u4efb\u52a1\u4e0e 40 \u4e2a\u7559\u51fa\u4efb\u52a1，\u5212\u5206\u5728\u5b9e\u9a8c\u4e2d\u56fa\u5b9a\u3002<\/p>\n
A.4\u3000JobBench\u3000\u4efb\u52a1\u53d6\u81ea\u771f\u5b9e\u4e13\u4e1a\u5de5\u4f5c\u6d41（Li et al., 2026），\u6bcf\u4e2a\u9644\u5e26\u8f93\u5165\u6587\u4ef6\u6587\u4ef6\u5939\u4e0e\u4e00\u4e2a\u5305\u88c5\u63d0\u793a\u8bcd，\u667a\u80fd\u4f53\u672c\u5e94\u67e5\u9605\u7684\u53c2\u8003\u8d44\u6599\u88ab\u6709\u610f\u6263\u7559，\u4f7f\u4e00\u90e8\u5206\u5de5\u4f5c\u6210\u4e3a\u771f\u6b63\u7684\u68c0\u7d22\u3002harness \u66b4\u9732\u4e00\u4e2a\u6587\u4ef6\u7cfb\u7edf\u3001\u4e00\u4e2a\u7528\u4e8e\u4ea7\u51fa Office\/PDF \u4ea4\u4ed8\u7269\u7684\u4ee3\u7801\u6267\u884c\u5de5\u5177\u3001\u4ee5\u53ca\u4e00\u4e2a\u63a5\u5730（grounded）\u7f51\u9875\u641c\u7d22\u5de5\u5177\u3002\u4ea4\u4ed8\u7269\u6309 benchmark \u81ea\u5e26\u7684\u52a0\u6743 rubric \u6253\u5206，\u62a5\u544a\u6570\u5b57\u5373\u88ab\u8bc4\u4f30 split \u4e0a\u7684\u52a0\u6743 rubric \u5206\u3002\u6211\u4eec\u4f7f\u7528\u4e00\u4e2a LLM \u88c1\u5224（Gemini-3.5-Flash \u4e0e Claude Opus 4.8 \u7684\u5e73\u5747\u5206）\u3002<\/p>\n
A.5\u3000GDPval\u3000\u5bf9\u6bcf\u4e2a\u4efb\u52a1，\u628a harness \u4ea7\u51fa\u7684\u4ea4\u4ed8\u7269\u4e0e benchmark \u81ea\u5e26\u7684\u4eba\u7c7b\u4e13\u5bb6\u4ea4\u4ed8\u7269\u5e76\u6392\u653e\u5728\u4e00\u8d77（Patwardhan et al., 2026），\u4e00\u4e2a\u7531\u4e09\u4f4d\u4e0d\u540c\u51fa\u8eab\u88c1\u5224\u7ec4\u6210\u7684\u8bc4\u5ba1\u56e2\u9009\u51fa\u4e24\u8005\u4e2d\u8f83\u597d\u7684\u4e00\u4e2a，\u62a5\u544a\u6570\u5b57\u5373\u5bf9 185 \u4e2a\u4efb\u52a1\u76f8\u5bf9\u4e13\u5bb6\u7684\u80dc\u7387\u3002\u8bc4\u5ba1\u56e2\u7531\u4e00\u4e2a\u672c\u5730\u90e8\u7f72\u7684\u5f00\u6e90\u6a21\u578b（Qwen3.6-35B-A3B）\u4e0e\u4e24\u4e2a\u6765\u81ea\u4e0d\u540c\u5382\u5546\u7684\u4e13\u6709\u6a21\u578b（Claude Sonnet 4.6\u3001Gemini-3.1 Pro）\u7ec4\u6210；\u6bcf\u4e00\u5bf9\u90fd\u6309\u4e24\u79cd\u5448\u73b0\u987a\u5e8f\u5404\u5224\u4e00\u6b21\u4ee5\u6d88\u9664\u4f4d\u7f6e\u504f\u5dee，\u4e00\u4e2a\u4efb\u52a1\u7684\u5224\u5b9a\u4e3a\u4e09\u4eba\u591a\u6570\u7968\u3002\u56e0\u6b64\u6bcf\u4f4d\u88c1\u5224\u5bf9\u6bcf\u4e2a harness \u51fa\u5177 204 \u6b21\u6bd4\u8f83；\u80dc\u7387\u9ad8\u4e8e 50% \u610f\u5473\u7740\u8be5 harness \u6bd4\u4e0e\u4e4b\u5bf9\u6bd4\u7684\u4eba\u7c7b\u4e13\u5bb6\u66f4\u5e38\u4ea7\u51fa\u53d7\u504f\u597d\u7684\u4ea4\u4ed8\u7269\u3002<\/p>\n
A.6\u3000APEX-Agents\u3000\u6bcf\u4e2a\u4efb\u52a1\u628a\u667a\u80fd\u4f53\u653e\u8fdb\u4e00\u4e2a\u5e26\u81ea\u6709 MCP \u5de5\u5177\u9762\u7684\u6c99\u76d2\u4e16\u754c，\u8986\u76d6\u6587\u4ef6\u7cfb\u7edf\u3001PDF \u9605\u8bfb\u3001\u7535\u5b50\u8868\u683c\u3001\u90ae\u4ef6\u3001\u804a\u5929\u3001\u65e5\u5386\u3001\u6587\u6863\u4e0e\u4ee3\u7801\u6267\u884c，\u6a2a\u8de8\u4e09\u4e2a\u4e13\u4e1a\u9886\u57df（Vidgen et al., 2026）\u3002\u4efb\u52a1\u7531\u4e00\u4e2a\u9010\u4efb\u52a1 rubric\u3001\u7ecf LLM \u88c1\u5224（Gemini-3.5-Flash）\u8bc4\u5224\u6253\u5206；\u53ea\u6709\u5355\u6b21\u91c7\u6837 rollout \u4e0a\u6ee1\u8db3 rubric，\u8be5\u4efb\u52a1\u5728 pass@1 \u4e0b\u624d\u7b97\u6210\u529f\u3002\u6211\u4eec\u8bc4\u4f30\u5168\u90e8 480 \u4e2a\u4efb\u52a1，\u5e76\u59cb\u7ec8\u5728\u8fd9\u4e2a\u5b8c\u6574\u5206\u6bcd\u4e0a\u62a5\u544a\u2014\u2014\u56e0\u57fa\u7840\u8bbe\u65bd\u6545\u969c\u800c\u7f3a\u5931 rollout \u7684\u4efb\u52a1\u8ba1\u4e3a\u5931\u8d25\u800c\u975e\u88ab\u6392\u9664，\u4ee5\u514d\u4e00\u4e2a\u5728\u96be\u4e16\u754c\u4e0a\u5d29\u6e83\u7684 harness \u770b\u8d77\u6765\u6bd4\u4e00\u4e2a\u6562\u4e8e\u5c1d\u8bd5\u7684 harness \u66f4\u597d\u3002<\/p>\n
A.7\u3000EngDesign\u3000\u6211\u4eec\u4f7f\u7528 EngDesign \u7684\u514d\u8bb8\u53ef\u5b50\u96c6（Guo et al., 2025），\u9009\u53d6\u5176\u4e2d 61 \u4e2a\u65e0\u9700\u4e13\u6709\u6a21\u62df\u5668\u5373\u53ef\u8fd0\u884c\u7684\u4efb\u52a1\u3002\u6bcf\u4e2a\u4efb\u52a1\u9648\u8ff0\u4e00\u4e2a\u8bbe\u8ba1\u76ee\u6807\u53ca\u8bbe\u8ba1\u5fc5\u987b\u6ee1\u8db3\u7684\u7269\u7406\u7ea6\u675f，\u7531\u5176\u81ea\u5e26\u7684\u51bb\u7ed3\u4eff\u771f\u5668\/\u6d4b\u8bd5\u53f0（\u800c\u975e\u88c1\u5224\u6a21\u578b）\u6253\u5206，\u56e0\u6b64\u6253\u5206\u662f\u786e\u5b9a\u6027\u7684，\u6211\u4eec\u6d4b\u5f97\u7684\u6bcf\u4e00\u5206\u65b9\u5dee\u90fd\u6765\u81ea\u7b56\u7565\u3002\u8fdb\u5316\u5728\u5168\u90e8 61 \u4e2a\u4efb\u52a1\u4e0a\u8fd0\u884c，\u4e0d\u505a\u5206\u5e03\u5185\u7559\u51fa\u5212\u5206\u2014\u2014\u5957\u4ef6\u592a\u5c0f，\u62bd\u4e0d\u51fa\u4efb\u52a1\u6765\u505a\u7559\u51fa\u3002<\/p>\n
A.8\u3000Frontier-Eng\u3000Frontier-Eng（Chi et al., 2026）\u4ece 26 \u4e2a\u9886\u57df\u6536\u96c6\u771f\u5b9e\u4e16\u754c\u5de5\u7a0b\u4f18\u5316\u95ee\u9898\u3002\u6bcf\u4e2a\u4efb\u52a1\u8981\u6c42\u667a\u80fd\u4f53\u4ea7\u51fa\u4e00\u4e2a\u8bbe\u8ba1\u6216\u7a0b\u5e8f，\u7531\u4e00\u4e2a\u51bb\u7ed3\u7684\u3001\u4efb\u52a1\u7279\u5b9a\u7684\u4eff\u771f\u5668\/\u8bc4\u4f30\u5668\u5728\u8fde\u7eed\u76ee\u6807\u4e0a\u6253\u5206\u2014\u2014\u4e0e EngDesign \u4e00\u6837\u4e0d\u6d89\u53ca\u88c1\u5224\u6a21\u578b\u3001\u6253\u5206\u786e\u5b9a\u6027\u3002\u7531\u4e8e\u5404\u4efb\u52a1\u76ee\u6807\u4e0d\u53ef\u516c\u5ea6，benchmark \u62a5\u544a Medal Score（\u5956\u724c\u5206）：\u5bf9\u6bcf\u4e2a\u4efb\u52a1，\u51bb\u7ed3 v1 \u5feb\u7167\u4e2d\u4e09\u4e2a\u6700\u4f73\u53ef\u884c\u7ed3\u679c\u5206\u522b\u4f5c\u4e3a\u91d1\u3001\u94f6\u3001\u94dc\u9608\u503c；\u63d0\u4ea4\u8fbe\u5230\u4efb\u4e00\u9608\u503c\u5206\u522b\u5f97 1\u30010.67\u30010.33；\u5206\u6570\u4e3a v1 \u96c6 47 \u4e2a\u4efb\u52a1\u4e0a\u7684\u5e73\u5747\u5f97\u5206，\u4ee5\u767e\u5206\u6bd4\u62a5\u544a\u3002\u6211\u4eec\u4ec5\u628a Frontier-Eng \u5f53\u4f5c\u5206\u5e03\u5916\u6d4b\u8bd5\u9762\u3002\u5176\u4e2d EngDesign \u9886\u57df\u590d\u7528\u4e86\u6211\u4eec evolve \u96c6\u7684\u4efb\u52a1，\u4e88\u4ee5\u6392\u9664；\u8bc4\u4f30\u73af\u5883\u65e0\u6cd5\u5728\u6211\u4eec\u6c99\u76d2\u4e2d\u642d\u5efa\u7684\u4efb\u52a1，\u4e24\u81c2\u90fd\u4e0d\u5f97\u5206\u3002\u6700\u7ec8 47 \u4e2a\u4efb\u52a1\u4e2d 38 \u4e2a\u8d21\u732e\u5f97\u5206，\u4e24\u81c2\u5728\u5b8c\u5168\u76f8\u540c\u7684\u4efb\u52a1\u96c6\u5408\u4e0a\u6253\u5206\u3002<\/p>\n
B. \u57fa\u7ebf\u65b9\u6cd5<\/h3>\n
\u7b80\u8981\u603b\u7ed3\u5b9e\u9a8c\u6240\u7528\u56db\u79cd harness \u8fdb\u5316\u57fa\u7ebf\u3002<\/p>\n
- \n
- Meta-Harness（Lee et al., 2026b）：\u628a harness \u5de5\u7a0b\u8868\u8ff0\u4e3a\u5bf9\u53ef\u6267\u884c harness \u4ee3\u7801\u7684\u5916\u5faa\u73af\u4f18\u5316\u95ee\u9898\u3002\u5176\u667a\u80fd\u4f53\u5f0f proposer \u53ef\u8bbf\u95ee\u5148\u524d\u5019\u9009\u7684\u6e90\u7801\u3001\u8bc4\u4f30\u5206\u6570\u4e0e\u6267\u884c\u8f68\u8ff9，\u5229\u7528\u8fd9\u4e9b\u79ef\u7d2f\u7ecf\u9a8c\u63d0\u51fa\u6539\u8fdb harness\u3002<\/li>\n
- Agentic Harness Engineering (AHE)（Lin et al., 2026a）：\u4e3a\u7f16\u7801\u667a\u80fd\u4f53 harness \u4f7f\u7528\u4e00\u4e2a\u53ef\u89c2\u6d4b\u6027\u9a71\u52a8\u7684\u8fdb\u5316\u5faa\u73af\u3002\u5b83\u628a harness \u7ec4\u4ef6\u3001\u6267\u884c\u7ecf\u9a8c\u4e0e\u7f16\u8f91\u7ed3\u679c\u7ec4\u7ec7\u6210\u663e\u5f0f\u8868\u793a，\u4f7f\u4e00\u4e2a\u4e0d\u65ad\u8fdb\u5316\u7684\u667a\u80fd\u4f53\u80fd\u591f\u8bca\u65ad\u5931\u8d25\u3001\u63d0\u51fa\u6539\u52a8\u3001\u5e76\u8bc4\u4f30\u5148\u524d\u7f16\u8f91\u7684\u6548\u679c\u3002<\/li>\n
- Test-Time Harness Evolution (TTHE)（Nie et al., 2026）：\u5728\u6d4b\u8bd5\u671f\u81ea\u9002\u5e94\u671f\u95f4\u8fdb\u5316\u53ef\u6267\u884c harness，\u540c\u65f6\u56fa\u5b9a\u5e95\u5c42\u6a21\u578b\u6743\u91cd\u3002\u5b83\u7ef4\u62a4\u591a\u4e2a\u5019\u9009 harness，\u4ece\u6267\u884c\u8f68\u8ff9\u63d0\u51fa\u4fee\u6539，\u5e76\u7528\u4e00\u4e2a\u667a\u80fd\u4f53\u88c1\u5224\u9009\u51fa\u4e00\u4e2a harness \u6301\u4e45\u5316\u5230\u540e\u7eed\u8f93\u5165\u3002<\/li>\n
- HarnessX（Chen et al., 2026）：\u628a\u667a\u80fd\u4f53 harness \u8868\u793a\u4e3a\u8de8\u63d0\u793a\u8bcd\u3001\u5de5\u5177\u3001\u8bb0\u5fc6\u3001\u63a7\u5236\u6d41\u7b49\u7ec4\u4ef6\u7684\u3001\u6a21\u5757\u5316\u3001\u5e26\u7c7b\u578b\u7684\u539f\u8bed\u7ec4\u5408\u3002\u5176\u8f68\u8ff9\u9a71\u52a8\u7684\u81ea\u9002\u5e94\u673a\u5236\u7528\u6267\u884c\u53cd\u9988\u4fee\u6539\u5e76\u9009\u62e9 harness \u914d\u7f6e，\u4f7f\u8fd0\u884c\u65f6\u811a\u624b\u67b6\u968f\u65f6\u95f4\u6f14\u5316\u3002<\/li>\n<\/ul>\n
C. \u65b9\u6cd5\u7ec6\u8282<\/h3>\n
\u672c\u8282\u7ed9\u51fa\u7b2c 3 \u8282\u7565\u53bb\u7684\u8f6e\u6b21\u7ea7\u5f62\u5f0f\u5316\u4e0e\u5b9e\u73b0\u7ec6\u8282，\u5177\u4f53\u8bf4\u660e\u5b9e\u9a8c\u4e2d\u4f7f\u7528\u7684\u63d0\u51fa\u4fa7\u4e0e\u9009\u62e9\u4fa7\u6b63\u5219\u5316\u5668\u3002\u4e0e\u6b63\u6587\u4e00\u6837，𝐿\u2080\u3001Lasso\/𝐿\u2081\u3001Ridge\/𝐿\u2082 \u8fd9\u4e9b\u672f\u8bed\u4ec5\u7528\u4e8e\u6307\u793a\u5b83\u4eec\u5728\u590d\u6742\u5ea6\u63a7\u5236\u4e2d\u7684\u7c7b\u6bd4\u89d2\u8272；\u8be5\u8fc7\u7a0b\u5e76\u4e0d\u4f18\u5316\u5bf9\u5e94\u7684\u8303\u6570\u60e9\u7f5a\u76ee\u6807，\u5f02\u6784\u7684 harness \u7ec4\u4ef6\u4e5f\u4e0d\u88ab\u5f53\u4f5c\u67d0\u4e2a\u5171\u4eab\u8fde\u7eed\u53c2\u6570\u5411\u91cf\u7684\u5750\u6807\u3002<\/p>\n
C.1\u3000\u8f6e\u6b21\u7ea7\u5f62\u5f0f\u5316\u3000\u8bb0 \u03a9(𝐻) \u4e3a\u4ece 𝐻 \u7ecf\u4efb\u610f\u6e90\u7801\u7f16\u8f91\u53ef\u8fbe\u7684 harness \u96c6\u5408\u3002RRSI \u4fdd\u6301 \u03a9(𝐻) \u5f00\u653e，\u8f6c\u800c\u6b63\u5219\u5316\u7a7f\u8fc7\u8be5\u7a7a\u95f4\u7684\u8f6c\u79fb\u3002\u4e00\u8f6e\u7684\u5f62\u5f0f\u4e3a：<\/p>\n
𝐇\u209c \u223c P_reg(\u00b7 | 𝐻\u209c, F\u209c, L\u209c, 𝑏\u209c, E\u209c, B\u209c) \u2286 \u03a9(𝐻\u209c)，\u3000\u3000𝐻\u209c\u208a\u2081 = argmax_{𝐻\u2032 \u2208 𝐇\u209c \u2229 A\u209c} \u015c(𝐻\u2032)\u3000\u3000(8)<\/p>\n
\u82e5\u6ca1\u6709\u5019\u9009\u53ef\u63a5\u53d7，\u5219 𝐻\u209c\u208a\u2081 = 𝐻\u209c\u3002\u5176\u4e2d F\u209c \u4e3a\u672c\u8f6e\u53cd\u9988，L\u209c \u4e3a\u7f16\u8f91\u5386\u53f2，𝑏\u209c \u4e3a\u5f0f (4) \u7684\u9000\u706b\u7f16\u8f91\u9884\u7b97，E\u209c \u542b\u63a2\u7d22\u6307\u4ee4，B\u209c \u4e3a\u4ece\u8fd1\u671f\u5386\u53f2\u63a8\u65ad\u51fa\u7684\u7ed3\u6784\u5316\u526a\u679d\u76ee\u6807，A\u209c \u4e3a\u88ab\u5141\u8bb8\u66ff\u6362 incumbent \u7684\u5019\u9009\u96c6\u5408\u3002\u4e00\u6b21\u8fd0\u884c\u5bf9 𝑡 = 0,\u2026,𝑇\u22121 \u6267\u884c\u7b97\u6cd5 1 \u4e0e\u7b97\u6cd5 2，\u4ece 𝐻\u2080 \u51fa\u53d1，𝑆\u2605 = \u015c(𝐻\u2080)\u3002\u8fdb\u5316\u5f00\u59cb\u524d，\u53cd\u590d\u8bc4\u4f30\u672a\u6539\u52a8\u7684\u57fa\u5ea7 harness \u4ee5\u4f30\u8ba1\u7ecf\u9a8c\u566a\u58f0\u5bb9\u5fcd\u5ea6 𝛿\u3002<\/p>\n
\u7b97\u6cd5 1（RRSI，\u63d0\u51fa\u4fa7） \u8f93\u5165：𝐻\u209c\u3001\u5386\u53f2 L\u209c\u3001\u7b2c 𝑡 \u8f6e（\u5171 𝑇 \u8f6e）；𝑏_min\u3001𝑏_max\u3001\u505c\u6ede\u7a97\u53e3 𝑤\u3001\u566a\u58f0\u5e26 𝛿<\/p>\n
- F\u209c \u2190 Analyze(𝐻\u209c, D_evolve)<\/li>\n
- 𝑏\u209c \u2190 𝑏_min + (𝑏_max \u2212 𝑏_min)\u00b7\u00bd(1 + cos(𝜋𝑡\/𝑇))\u3000\u22b3 𝐿\u2080 \u5f0f\u7f16\u8f91\u57fa\u6570\u63a7\u5236<\/li>\n
- 𝜎\u209c \u2190 𝟙[ \u015c\u209c \u2212 \u015c\u209c\u208b𝑤 \u2264 𝛿 ]<\/li>\n
- T\u209c \u2190 { \u2113\u1d62 : (𝑡\u1d62, \u2113\u1d62, \u2026) \u2208 L\u209c }<\/li>\n
- U\u209c \u2190 K \\\\ T\u209c<\/li>\n
- E\u209c \u2190 (𝜎\u209c, U\u209c, 𝑚_draft)<\/li>\n
- B\u209c \u2190 { \u2113 \u2208 T\u209c : 𝑔\u209c(\u2113) \u2264 0 }\u3000\u22b3 Lasso\/𝐿\u2081 \u5f0f\u526a\u679d\u76ee\u6807<\/li>\n
- 𝐇\u209c \u223c P_reg(\u00b7 | 𝐻\u209c, F\u209c, L\u209c, 𝑏\u209c, E\u209c, B\u209c)<\/li>\n
- \u7528\u7ec4\u4ef6\u4e0e\u5047\u8bbe\u5143\u6570\u636e\u6807\u8bb0\u6bcf\u6761\u539f\u5b50\u7f16\u8f91<\/li>\n
- \u8fd4\u56de\u901a\u8fc7\u9884\u8bc4\u4f30\u7b5b\u67e5\u7684\u5019\u9009<\/li>\n
\u7b97\u6cd5 2（RRSI，\u9009\u62e9\u4fa7） \u8f93\u5165：\u7ecf\u7b5b\u67e5\u7684 𝐇\u209c\u3001(𝐻\u209c, \u015c\u209c, \u0108\u209c)\u3001𝑆\u2605\u3001𝛿\u3001𝑘；𝛽\u2080\u3001𝛽\u2081\u3001𝑤\u209b\u3001𝑤_c\u3001𝑤_n<\/p>\n
- A\u209c \u2190 \u2205<\/li>\n
- \u5bf9 𝐇\u209c \u4e2d\u6bcf\u4e2a 𝐻\u2032 \u5e76\u884c：<\/li>\n
- \u015c\u2032, \u0108\u2032 \u2190 Evaluate(𝐻\u2032, D_evolve, 𝑘)<\/li>\n
- \u0394𝑆 \u2190 \u015c\u2032 \u2212 \u015c\u209c；\u0394𝐶 \u2190 (\u0108\u2032 \u2212 \u0108\u209c)\/\u0108\u209c<\/li>\n
- 𝜈 \u2190 𝜈\u209c(𝐻\u2032)\u3000\u22b3 \u65b0\u7684\u7ed3\u6784\u7ec4\u4ef6\u7c7b\u578b\u6570<\/li>\n
- \u82e5 \u0394𝑆 > 𝛿：<\/li>\n
- 𝑐 \u2190 [ \u0394𝐶 \u2264 𝛽\u2080 + 𝛽\u2081\u0394𝑆 ]\u3000\u22b3 \u589e\u76ca\u76f8\u5173\u6210\u672c\u89c4\u5219，\u5f0f (7)<\/li>\n
- \u5426\u5219：<\/li>\n
- 𝑐 \u2190 [ 𝑤\u209b\u0394𝑆 \u2212 𝑤_c\u0394𝐶 + 𝑤_n𝜈 > 0 ]\u3000\u22b3 \u5e26\u5185\u89c4\u5219，\u5f0f (17)<\/li>\n
- \u7ed3\u675f\u82e5<\/li>\n
- 𝑔 \u2190 DomainGuard(𝐻\u209c, 𝐻\u2032)<\/li>\n
- \u82e5 \u015c\u2032 \u2265 𝑆\u2605 \u2212 𝛿 \u4e14 𝑐 \u4e14 𝑔：<\/li>\n
- A\u209c \u2190 A\u209c \u222a {𝐻\u2032}<\/li>\n
- \u7ed3\u675f\u82e5<\/li>\n
- \u7ed3\u675f\u5faa\u73af<\/li>\n
- 𝐻\u209c\u208a\u2081 \u2190 A\u209c \u4e2d \u015c\u2032 \u6700\u5927\u8005；\u82e5 A\u209c = \u2205 \u5219\u53d6 𝐻\u209c<\/li>\n
- 𝑆\u2605 \u2190 max(𝑆\u2605, \u015c\u209c\u208a\u2081)<\/li>\n
- \u8bb0\u5f55\u6bcf\u6761\u6d4b\u5f97\u7f16\u8f91；\u5f53\u4e14\u4ec5\u5f53\u5176\u5019\u9009\u4e3a 𝐻\u209c\u208a\u2081 \u2260 𝐻\u209c \u65f6\u8bb0 𝑎 = 1<\/li>\n
- \u8fd4\u56de 𝐻\u209c\u208a\u2081<\/li>\n
C.2\u3000\u63d0\u51fa\u4fa7\u7c3f\u8bb0<\/p>\n
\u539f\u5b50\u7f16\u8f91\u8868\u793a\u3002 \u7b2c 𝑡 \u8f6e，proposer \u8d77\u8349\u4e00\u4e2a\u9488\u5bf9 𝐻\u209c \u7684\u539f\u5b50\u7f16\u8f91\u6c60 E\u209c，\u4e00\u4e2a\u5019\u9009\u5e94\u7528\u8be5\u6c60\u7684\u4e00\u4e2a\u5b50\u96c6\u3002\u628a\u8be5\u5b50\u96c6\u8bb0\u4e3a 𝑧\u209c \u2208 {0,1}^|E\u209c|，𝑧\u209c,\u2c7c = 1 \u8868\u793a\u7f16\u8f91 𝑗 \u88ab\u7eb3\u5165\u3002\u7f16\u8f91\u6c60\u6bcf\u8f6e\u4ece\u5f00\u653e\u7a7a\u95f4 \u03a9(𝐻\u209c) \u91cd\u65b0\u62bd\u53d6，\u6545 |E\u209c| \u4e0d\u5fc5\u8de8\u8f6e\u56fa\u5b9a\u3002\u5f0f (4) \u7684\u9000\u706b\u9884\u7b97\u65bd\u52a0：<\/p>\n
\u2016𝑧\u209c\u2016\u2080 \u2264 𝑏\u209c\u3000\u3000(9)<\/p>\n
\u56e0\u6b64 𝑏\u209c \u9650\u5236\u7684\u662f\u6346\u8fdb\u5355\u4e2a\u5019\u9009\u7684\u3001\u53ef\u72ec\u7acb\u5f52\u56e0\u7684\u7f16\u8f91\u6570，\u800c\u975e"\u6700\u7ec8\u53ef\u80fd\u88ab\u4fee\u6539\u7684\u7ec4\u4ef6\u96c6\u5408"\u3002\u8fd9\u662f\u6211\u4eec\u6700\u76f4\u63a5\u7684\u7ecf\u5178\u7c7b\u6bd4：\u5b83\u662f\u5bf9\u66f4\u65b0\u7684\u57fa\u6570\u7ea6\u675f，\u800c\u975e\u5bf9\u56fa\u5b9a\u6a21\u578b\u53c2\u6570\u5411\u91cf\u7684 𝐿\u2080 \u60e9\u7f5a\u3002<\/p>\n
\u7f16\u8f91\u5386\u53f2\u4e0e\u7ec4\u4ef6\u7ea7\u6458\u8981\u3002 \u88ab\u8bc4\u4f30\u5019\u9009\u4e2d\u7684\u6bcf\u6761\u539f\u5b50\u7f16\u8f91\u90fd\u6807\u8bb0\u4e00\u4e2a\u7ec4\u4ef6 \u2113\u3001\u4e00\u4e2a\u5047\u8bbe \u210e\u3001\u4ee5\u53ca\u5019\u9009\u6e90\u7801 diff 𝑑\u3002\u542b\u591a\u6761\u7f16\u8f91\u7684\u5019\u9009\u6309\u6bcf\u6761\u7f16\u8f91\u8d21\u732e\u4e00\u6761\u5386\u53f2\u8bb0\u5f55；\u8be5\u5019\u9009\u4e2d\u6240\u6709\u7f16\u8f91\u5171\u4eab\u540c\u4e00\u4e2a\u6d4b\u5f97\u7684 \u0394𝑆\u3001\u0394𝐶 \u4e0e\u8f6e\u6b21\u7ed3\u679c\u3002\u5ffd\u7565\u90a3\u4e9b\u5728\u83b7\u5f97\u6709\u6548\u6d4b\u91cf\u524d\u5c31\u5931\u8d25\u7684\u5019\u9009，\u7b2c 𝑡 \u8f6e\u4e4b\u524d\u7684\u5386\u53f2\u53ef\u5199\u4e3a：<\/p>\n
L\u209c = { (𝑡\u1d62, \u2113\u1d62, \u210e\u1d62, 𝑑\u1d62, \u0394𝑆\u1d62, \u0394𝐶\u1d62, 𝑎\u1d62) : 𝑖 \u2264 𝑛\u209c }，\u3000𝑎\u1d62 \u2208 {0,1}\u3000\u3000(10)<\/p>\n
\u5176\u4e2d 𝑎\u1d62 = 1 \u5f53\u4e14\u4ec5\u5f53\u643a\u5e26\u7f16\u8f91 𝑖 \u7684\u5019\u9009\u88ab\u9009\u4e3a\u8be5\u8f6e\u80dc\u8005\u3001\u4ece\u800c\u8fdb\u5165\u88ab\u63a5\u53d7\u7684\u8fdb\u5316\u8def\u5f84\u3002\u90a3\u4e9b\u53ef\u63a5\u53d7\u4f46\u8f93\u7ed9\u5206\u6570\u66f4\u9ad8\u7684\u53ef\u63a5\u53d7\u5019\u9009\u7684，𝑎\u1d62 = 0\u3002<\/p>\n
proposer \u4f7f\u7528\u7684\u4e24\u4e2a\u6458\u8981\u4e3a：<\/p>\n
T\u209c = { \u2113\u1d62 : 𝑖 \u2264 𝑛\u209c }，\u3000\u3000𝑔\u209c(\u2113) = max{ \u0394𝑆\u1d62 : \u2113\u1d62 = \u2113, 𝑡 \u2212 𝑡\u1d62 \u2264 𝑛_prune }，max \u2205 = \u2212\u221e\u3000\u3000(11)<\/p>\n
\u8fd9\u91cc T\u209c \u662f\u81f3\u5c11\u6709\u8fc7\u4e00\u6b21\u6d4b\u5f97\u7f16\u8f91\u7684\u7ec4\u4ef6\u96c6\u5408，𝑔\u209c(\u2113) \u662f\u526a\u679d\u7a97\u53e3\u5185\u4e0e\u7ec4\u4ef6 \u2113 \u5173\u8054\u7684\u6700\u4f73\u8fd1\u671f\u6d4b\u5f97\u589e\u76ca\u3002\u7531\u4e8e\u6346\u7ed1\u7f16\u8f91\u7ee7\u627f\u5019\u9009\u7ea7\u6d4b\u91cf，\u968f\u7740\u7f16\u8f91\u9884\u7b97\u9000\u706b\u8d8b\u8fd1\u4e8e\u4e00，\u8fd9\u4e9b\u8bc1\u636e\u4e5f\u53d8\u5f97\u66f4\u53ef\u5f52\u56e0\u3002<\/p>\n
\u7ed3\u6784\u5316\u63a2\u7d22\u72b6\u6001\u3002 \u8bb0 K \u4e3a\u53ef\u7f16\u8f91\u7ec4\u4ef6\u8bcd\u8868\u3002\u5b9e\u73b0\u4e2d：<\/p>\n
K = { prompt, control_flow, config, output_plumbing, context_mgmt, client_tool, skill, memory, subagent }\u3000\u3000(12)<\/p>\n
\u63a2\u7d22\u6307\u4ee4\u4e3a：<\/p>\n
E\u209c = (𝜎\u209c, U\u209c, 𝑚_draft)，\u3000𝜎\u209c = 𝟙[ \u015c\u209c \u2212 \u015c\u209c\u208b𝑤 \u2264 𝛿 ]，\u3000U\u209c = K \\\\ T\u209c\u3000\u3000(13)<\/p>\n
\u5176\u4e2d 𝜎\u209c \u8868\u793a\u8fc7\u53bb 𝑤 \u8f6e\u7684\u8fdb\u5c55\u672a\u8d85\u51fa\u7ecf\u9a8c\u566a\u58f0\u5bb9\u5fcd\u5ea6，U\u209c \u542b\u5c1a\u672a\u88ab\u4efb\u4f55\u6d4b\u5f97\u7f16\u8f91\u52a8\u7528\u8fc7\u7684\u7ec4\u4ef6，𝑚_draft \u5728\u641c\u7d22\u505c\u6ede\u65f6\u4e3a\u63a2\u7d22\u6027\u7f16\u8f91\u9884\u7559\u5019\u9009\u540d\u989d\u3002<\/p>\n
\u7ed3\u6784\u5316\u526a\u679d\u3002 \u526a\u679d\u76ee\u6807\u96c6\u4e3a：<\/p>\n
B\u209c = { \u2113 \u2208 T\u209c : 𝑔\u209c(\u2113) \u2264 0 }\u3000\u3000(14)<\/p>\n
\u4e8e\u662f，\u4e00\u4e2a\u7ec4\u4ef6\u82e5\u5df2\u88ab\u52a8\u7528\u3001\u4f46\u5728\u8fd1\u671f\u526a\u679d\u7a97\u53e3\u5185\u672a\u4ea7\u751f\u4e25\u683c\u4e3a\u6b63\u7684\u6d4b\u5f97\u589e\u76ca，\u5c31\u88ab\u6807\u8bb0\u4e3a\u65e0\u4ea7\u51fa\u3002proposer \u6536\u5230 B\u209c \u4ee5\u53ca\u4e0e\u8fd9\u4e9b\u7ec4\u4ef6\u5173\u8054\u7684\u4efb\u4f55\u5148\u524d\u88ab\u63a5\u53d7\u7684\u7f16\u8f91，\u5e76\u88ab\u6307\u793a\u5728\u540e\u7eed\u63d0\u6848\u4e2d\u79fb\u9664\u65e0\u4ea7\u51fa\u7684\u673a\u68b0\u7ed3\u6784\u3002\u8fd9\u5728\u89d2\u8272\u4e0a\u7c7b\u6bd4 Lasso\/𝐿\u2081 \u5f0f\u7a00\u758f\u5316，\u56e0\u4e3a\u8be5\u673a\u5236\u901a\u8fc7\u4ece\u4fdd\u7559\u7684 harness \u4e2d\u5220\u9664\u79bb\u6563\u7ed3\u6784\u6765\u8d77\u4f5c\u7528；\u5b83\u4e0d\u662f\u4e00\u4e2a 𝐿\u2081 \u60e9\u7f5a\u7684\u8fde\u7eed\u4f18\u5316\u95ee\u9898\u3002<\/p>\n
C.3\u3000\u9009\u62e9\u4fa7\u7c3f\u8bb0<\/p>\n
\u566a\u58f0\u8c03\u6574\u540e\u7684\u5e95\u7ebf\u3002 \u6cc4\u6f0f critic \u5728\u5b8c\u6574\u8bc4\u4f30\u4e4b\u524d\u65bd\u52a0\u3002\u5bf9\u6bcf\u4e2a\u8fdb\u5165\u9009\u62e9\u7684\u5019\u9009，\u7b2c\u4e00\u4e2a\u975e\u8865\u507f\u6027\u6027\u80fd\u8981\u6c42\u662f\u5f0f (5) \u7684\u7a33\u5b9a\u6027\u5e95\u7ebf：\u015c(𝐻\u2032) \u2265 𝑆\u2605 \u2212 𝛿\u3002\u5b83\u5141\u8bb8\u5728\u7ecf\u9a8c\u6821\u51c6\u5bb9\u5fcd\u5ea6\u5185\u7684\u6ce2\u52a8，\u540c\u65f6\u9632\u6b62\u641c\u7d22\u7d2f\u79ef\u4e00\u8fde\u4e32\u5c0f\u9000\u6b65\u3002<\/p>\n
\u5e26\u5185\u89c4\u5219\u6240\u7528\u7684\u65b0\u9896\u5ea6\u3002 \u5851\u5f62\u89c4\u5219\u53ea\u5bf9\u7ed3\u6784\u7ec4\u4ef6\u4f7f\u7528\u65b0\u9896\u5ea6\u3002\u4ee4 K_str = { client_tool, skill, memory, subagent } (15)，\u8bb0 𝑁\u209c(\u2113) \u4e3a\u7b2c 𝑡 \u8f6e\u4e4b\u524d\u3001\u6807\u8bb0\u4e3a\u7ec4\u4ef6 \u2113 \u7684\u5df2\u63a5\u53d7\u7f16\u8f91\u8bb0\u5f55\u6570\u3002\u82e5 comp(𝐻\u2032) \u4e3a\u5019\u9009 𝐻\u2032 \u89e6\u53ca\u7684\u7ec4\u4ef6\u7c7b\u578b\u96c6\u5408，\u5b9e\u73b0\u8ba1\u7b97：<\/p>\n
𝜈\u209c(𝐻\u2032) = \u03a3_{\u2113 \u2208 K_str} 𝟙[ \u2113 \u2208 comp(𝐻\u2032) \u2227 𝑁\u209c(\u2113) = 0 ]\u3000\u3000(16)<\/p>\n
\u56e0\u6b64 𝜈\u209c(𝐻\u2032) \u7edf\u8ba1\u8be5\u5019\u9009\u89e6\u53ca\u7684\u3001\u6b64\u524d\u4ece\u672a\u51fa\u73b0\u5728\u83b7\u80dc\u7f16\u8f91\u4e2d\u7684\u4e0d\u540c\u7ed3\u6784\u7ec4\u4ef6\u7c7b\u578b\u6570\u3002\u63d0\u793a\u8bcd\u3001\u63a7\u5236\u6d41\u3001\u914d\u7f6e\u3001\u8f93\u51fa\u7ba1\u7ebf\u4e0e\u4e0a\u4e0b\u6587\u7ba1\u7406\u7c7b\u7f16\u8f91\u4e0d\u83b7\u5f97\u8fd9\u4e00\u65b0\u9896\u5ea6\u5956\u52b1\u3002<\/p>\n
\u589e\u76ca\u8d85\u8fc7\u566a\u58f0\u5bb9\u5fcd\u5ea6\u65f6\u7684\u63a5\u53d7\u3002 \u5bf9 \u0394𝑆 > 𝛿，\u9009\u62e9\u5668\u4f7f\u7528\u5f0f (7) \u7684\u589e\u76ca\u76f8\u5173\u6210\u672c\u6761\u4ef6：\u0394𝐶 \u2264 𝛽\u2080 + 𝛽\u2081\u0394𝑆\u3002\u8be5\u89c4\u5219\u53ea\u5728\u4f34\u968f\u66f4\u5927\u6d4b\u5f97\u6539\u8fdb\u65f6\u624d\u5141\u8bb8\u66f4\u591a\u63a8\u7406\u6210\u672c\u3002\u8fd9\u6b63\u662f\u6b63\u6587\u5f15\u51fa Ridge\/𝐿\u2082 \u5f0f\u7c7b\u6bd4\u7684\u90e8\u5206：\u5b83\u6291\u5236\u805a\u5408\u8d44\u6e90\u5360\u7528\u7684\u65e0\u7ea6\u675f\u589e\u957f，\u800c\u4e0d\u8981\u6c42\u5254\u9664\u5355\u4e2a\u7ec4\u4ef6\u3002\u7c7b\u6bd4\u662f\u529f\u80fd\u6027\u7684\u800c\u975e\u6570\u5b66\u7684：\u8be5\u89c4\u5219\u4e0d\u662f\u4e00\u4e2a\u5e73\u65b9\u8303\u6570\u60e9\u7f5a\u3002<\/p>\n
\u589e\u76ca\u672a\u8d85\u8fc7\u566a\u58f0\u5bb9\u5fcd\u5ea6\u65f6\u7684\u63a5\u53d7\u3002 \u5bf9\u901a\u8fc7\u7a33\u5b9a\u6027\u5e95\u7ebf\u3001\u4f46\u6d4b\u5f97\u589e\u76ca\u672a\u8d85\u51fa\u7ecf\u9a8c\u5bb9\u5fcd\u5ea6（\u0394𝑆 \u2264 𝛿）\u7684\u5019\u9009，\u5b9e\u73b0\u4e0d\u4f7f\u7528\u5f0f (7)，\u800c\u65bd\u52a0\u5851\u5f62\u53ef\u63a5\u53d7\u6761\u4ef6：<\/p>\n
𝑤\u209b\u0394𝑆 \u2212 𝑤_c\u0394𝐶 + 𝑤_n𝜈\u209c(𝐻\u2032) > 0\u3000\u3000(17)<\/p>\n
\u5176\u4e2d 𝑤\u209b\u3001𝑤_c\u3001𝑤_n \u2265 0 \u5206\u522b\u63a7\u5236\u6d4b\u5f97\u5206\u6570\u53d8\u5316\u3001\u76f8\u5bf9\u63a8\u7406\u6210\u672c\u53d8\u5316\u3001\u4ee5\u53ca\u6b64\u524d\u672a\u7528\u8fc7\u7684\u7ed3\u6784\u7ec4\u4ef6\u7c7b\u578b\u7684\u8d21\u732e\u3002\u8be5\u5206\u652f\u7684\u76ee\u7684\u662f\u907f\u514d\u628a\u4e00\u4e2a\u5c0f\u7684\u5206\u6570\u6ce2\u52a8\u672c\u8eab\u5f53\u4f5c\u5145\u5206\u8bc1\u636e\u3002\u5728\u8fd9\u4e00\u533a\u57df\u5185，\u964d\u4f4e\u6210\u672c\u901a\u8fc7 \u2212𝑤_c\u0394𝐶 \u6b63\u5411\u8d21\u732e；\u5c1d\u8bd5\u4e00\u4e2a\u4ece\u672a\u8fdb\u5165\u8fc7\u88ab\u63a5\u53d7\u8fdb\u5316\u8def\u5f84\u7684\u7ed3\u6784\u673a\u5236\u901a\u8fc7 𝑤_n𝜈\u209c(𝐻\u2032) \u8d21\u732e\u3002\u4f9d\u8fdb\u5316\u5b9e\u4f8b\u4e0d\u540c，\u5e26\u5185\u5206\u6570\u53d8\u5316\u4e5f\u53ef\u80fd\u901a\u8fc7 𝑤\u209b\u0394𝑆 \u8d21\u732e\u3002<\/p>\n
\u7f16\u7801\u5b9e\u4f8b\u8bbe 𝑤\u209b = 0\u3002\u56e0\u6b64\u4e00\u4e2a\u4ecd\u843d\u5728 𝛿 \u5185\u7684\u5206\u6570\u63d0\u5347，\u672c\u8eab\u4e0d\u80fd\u8ba9\u4e00\u4e2a\u7f16\u7801\u5019\u9009\u53d8\u5f97\u53ef\u63a5\u53d7；\u5019\u9009\u5fc5\u987b\u8f6c\u800c\u4ece\u66f4\u4f4e\u6210\u672c\u548c\/\u6216\u7ed3\u6784\u65b0\u9896\u5ea6\u83b7\u5f97\u8db3\u591f\u4fe1\u7528\u3002\u667a\u80fd\u4f53\u5de5\u4f5c\u533a\u4e0e\u5de5\u7a0b\u8bbe\u8ba1\u5b9e\u4f8b\u4f7f\u7528\u6b63\u7684 𝑤\u209b\u3002\u4e09\u4e2a\u6743\u91cd\u5bf9\u4e00\u4e2a\u8fdb\u5316\u5b9e\u4f8b\u56fa\u5b9a，\u5217\u4e8e\u8868 5\u3002\u5f0f (17) \u662f\u4e0d\u786e\u5b9a\u6027\u533a\u57df\u5185\u90e8\u4e00\u4e2a\u5b9e\u73b0\u5c42\u9762\u7684\u51b3\u80dc\/\u53ef\u63a5\u53d7\u89c4\u5219，\u672c\u8eab\u4e0d\u7b49\u540c\u4e8e\u67d0\u4e2a 𝐿\u209a \u60e9\u7f5a\u3002<\/p>\n
\u9886\u57df\u7279\u5b9a\u7684\u975e\u8865\u507f\u6027\u5b88\u536b\u3002 \u5728\u7a33\u5b9a\u6027\u4e0e\u6210\u672c\u68c0\u67e5\u4e4b\u540e，\u5b9e\u73b0\u53ef\u65bd\u52a0\u4e00\u4e2a\u9886\u57df\u7279\u5b9a\u5b88\u536b 𝑔(𝐻\u209c, 𝐻\u2032) \u2208 {0,1}\u3002\u7f16\u7801\u4e0e\u667a\u80fd\u4f53\u5de5\u4f5c\u533a\u5b9e\u4f8b\u4e0d\u4f7f\u7528\u989d\u5916\u5b88\u536b，\u6545 𝑔 = 1\u3002\u5de5\u7a0b\u8bbe\u8ba1\u5b9e\u4f8b\u989d\u5916\u62d2\u7edd\u8fd9\u6837\u7684\u5019\u9009：\u5176\u6709\u6548\u8f93\u51fa\u7387\u76f8\u5bf9 incumbent \u4e0b\u964d\u8d85\u8fc7 0.03，\u6216\u65e0\u63d0\u4ea4\u7387\u4e0a\u5347\u8d85\u8fc7 0.02\u3002\u8fd9\u4e9b\u5b88\u536b\u9632\u6b62"\u4e3b\u901a\u8fc7\u7387\u76ee\u6807\u4e0a\u7684\u4e00\u70b9\u63d0\u5347"\u53bb\u8865\u507f"\u57fa\u672c\u6267\u884c\u6709\u6548\u6027\u4e0a\u7684\u5927\u5e45\u9000\u5316"\u3002<\/p>\n
\u672b\u8f6e\u9009\u62e9\u3002 \u4e00\u4e2a\u5019\u9009\u53ea\u6709\u540c\u65f6\u6ee1\u8db3\u566a\u58f0\u8c03\u6574\u540e\u7684\u5e95\u7ebf\u3001\u590d\u6742\u5ea6\u611f\u77e5\u89c4\u5219\u7684\u76f8\u5e94\u5206\u652f\u3001\u4ee5\u53ca\u6240\u6709\u6fc0\u6d3b\u7684\u9886\u57df\u5b88\u536b，\u624d\u7b97\u53ef\u63a5\u53d7\u3002\u5728\u53ef\u63a5\u53d7\u5019\u9009\u4e2d，\u9009\u62e9\u5668\u9009\u6d4b\u5f97\u5206\u6570\u6700\u9ad8\u8005；\u82e5\u65e0\u53ef\u63a5\u53d7\u5019\u9009，\u5219\u4fdd\u7559 incumbent\u3002\u968f\u540e\u66f4\u65b0\u8fd0\u884c\u6700\u4f73\u5206\u6570 𝑆\u2605 \u2190 max(𝑆\u2605, \u015c(𝐻\u209c\u208a\u2081))\u3002<\/p>\n
D. \u5b9e\u9a8c<\/h3>\n
D.1\u3000\u8d85\u53c2\u6570\u8bbe\u7f6e\u3000RRSI \u5f15\u5165\u5c11\u91cf\u8d85\u53c2，\u63a7\u5236\u66f4\u65b0\u7a00\u758f\u6027\u3001\u63a2\u7d22\u3001\u526a\u679d\u4ee5\u53ca\u6210\u672c\u2014\u6027\u80fd\u6743\u8861\u3002\u6211\u4eec\u4ec5\u7528\u8fdb\u5316\u73af\u5883\u4e0e\u5de5\u7a0b\u8003\u8651\u6765\u9009\u8fd9\u4e9b\u53c2\u6570；\u7559\u51fa\u4e0e OOD benchmark \u4e0d\u7528\u4e8e\u8c03\u53c2\u3002\u566a\u58f0\u5bb9\u5fcd\u5ea6 𝛿 \u7531\u53cd\u590d\u8bc4\u4f30\u672a\u6539\u52a8\u57fa\u5ea7 harness \u6821\u51c6\u3002\u7f16\u8f91\u9884\u7b97\u53c2\u6570 (𝑏_min, 𝑏_max) \u51b3\u5b9a\u5355\u4e2a\u5019\u9009\u80fd\u6346\u7ed1\u591a\u5c11\u72ec\u7acb\u6539\u52a8；𝑤 \u4e0e 𝑚_draft \u63a7\u5236\u641c\u7d22\u4f55\u65f6\u3001\u591a\u5f3a\u5730\u63a2\u7d22\u4f7f\u7528\u4e0d\u8db3\u7684\u7ec4\u4ef6\u3002\u526a\u679d\u7a97\u53e3 𝑛_prune \u51b3\u5b9a\u4e00\u4e2a\u7ec4\u4ef6\u8981\u79ef\u7d2f\u591a\u5c11\u8fd1\u671f\u8bc1\u636e\u624d\u88ab\u89c6\u4e3a\u65e0\u4ea7\u51fa\u3002\u6700\u540e，(𝛽\u2080, 𝛽\u2081) \u7f16\u7801\u6d4b\u5f97\u589e\u76ca\u4e0e\u989d\u5916\u63a8\u7406\u6210\u672c\u4e4b\u95f4\u5141\u8bb8\u7684\u6743\u8861\u3002\u8868 5 \u5217\u51fa\u5404\u5b9e\u4f8b\u6240\u7528\u503c\u3002\u5206\u6570 \u015c \u4e3a [0,1] \u4e0a\u7684\u6bd4\u4f8b，\u0394𝐶 \u4e3a\u6bcf\u6b21\u8bd5\u9a8c\u7b56\u7565 token \u7684\u76f8\u5bf9\u53d8\u5316，\u6545 𝛿 \u4e0e 𝛽\u2081 \u4ee5\u8fd9\u4e9b\u5355\u4f4d\u8868\u793a：\u7f16\u7801\u5b9e\u4f8b\u4e0a 𝛿 \u5bf9\u5e94 89\u00d7𝑘=178 \u6b21\u8bd5\u9a8c\u4e2d\u7684 3 \u6b21\u901a\u8fc7；\u667a\u80fd\u4f53\u5de5\u4f5c\u533a\u5b9e\u4f8b\u4e0a\u5bf9\u5e94\u7ea6 14100 \u6761\u6807\u51c6\u5224\u5b9a\u4e2d\u7684 60 \u6761；\u5de5\u7a0b\u8bbe\u8ba1\u5b9e\u4f8b\u4e0a\u5bf9\u5e94 61\u00d7𝑘=244 \u6b21\u8bd5\u9a8c\u4e2d\u7684 5 \u6b21\u901a\u8fc7\u3002\u540c\u7406，𝛽\u2081 \u5bf9\u5e94：\u7f16\u7801\u4e2d\u6bcf\u591a\u4e00\u6b21\u901a\u8fc7\u5141\u8bb8 25% token；\u667a\u80fd\u4f53\u5de5\u4f5c\u533a\u4e2d\u6bcf\u591a 100 \u6761\u6807\u51c6；\u5de5\u7a0b\u8bbe\u8ba1\u4e2d\u6bcf\u591a\u4e00\u6b21\u901a\u8fc7\u5141\u8bb8 10%\u3002<\/p>\n
\u8868 5｜RRSI \u5728\u5404\u8fdb\u5316\u8bbe\u7f6e\u4e2d\u4f7f\u7528\u7684\u8d85\u53c2\u6570\u3002\u6240\u6709\u9009\u62e9\u5747\u5728\u4e0d\u53c2\u8003\u7559\u51fa\u6216 OOD benchmark \u7684\u60c5\u51b5\u4e0b\u56fa\u5b9a\u3002<\/p>\n
\n
\u8d85\u53c2\u4f5c\u7528\u7f16\u7801\u667a\u80fd\u4f53\u5de5\u4f5c\u533a\u5de5\u7a0b\u8bbe\u8ba1<\/tr>\n \n\n 𝑇<\/td>\n \u8fdb\u5316\u8f6e\u6b21<\/td>\n 20<\/td>\n 20<\/td>\n 40<\/td>\n<\/tr>\n \n 𝑘<\/td>\n \u6bcf\u6b21\u8bc4\u4f30\u6bcf\u4efb\u52a1\u8bd5\u9a8c\u6570<\/td>\n 2<\/td>\n 2<\/td>\n 4<\/td>\n<\/tr>\n \n 𝛿<\/td>\n \u7ecf\u9a8c\u566a\u58f0\u5bb9\u5fcd\u5ea6<\/td>\n 0.017<\/td>\n 0.004<\/td>\n 0.020<\/td>\n<\/tr>\n \n 𝑏_min<\/td>\n \u672b\u8f6e\u7f16\u8f91\u9884\u7b97<\/td>\n 1<\/td>\n 1<\/td>\n 1<\/td>\n<\/tr>\n \n 𝑏_max<\/td>\n \u521d\u59cb\u7f16\u8f91\u9884\u7b97<\/td>\n 4<\/td>\n 3<\/td>\n 4<\/td>\n<\/tr>\n \n 𝑤<\/td>\n \u505c\u6ede\u68c0\u6d4b\u7a97\u53e3<\/td>\n 3<\/td>\n 3<\/td>\n 3<\/td>\n<\/tr>\n \n 𝑚_draft<\/td>\n \u9884\u7559\u63a2\u7d22\u6027\u63d0\u6848\u6570<\/td>\n 1<\/td>\n 1<\/td>\n 1<\/td>\n<\/tr>\n \n 𝑛_prune<\/td>\n \u526a\u679d\u7a97\u53e3<\/td>\n 4<\/td>\n 4<\/td>\n 5<\/td>\n<\/tr>\n \n 𝛽\u2080<\/td>\n \u57fa\u7840\u6210\u672c\u8bb8\u53ef<\/td>\n 0.10<\/td>\n 0.10<\/td>\n 0.15<\/td>\n<\/tr>\n \n 𝛽\u2081<\/td>\n \u589e\u76ca\u76f8\u5173\u6210\u672c\u8bb8\u53ef<\/td>\n 44.5<\/td>\n 35.4<\/td>\n 24.4<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n E. \u5b9a\u6027\u6848\u4f8b\u7814\u7a76<\/h3>\n
\u4e3a\u8865\u5145\u805a\u5408\u7ed3\u679c，\u6211\u4eec\u68c0\u89c6 RRSI \u8fdb\u5316\u8fc7\u7a0b\u4e2d\u505a\u51fa\u7684\u4ee3\u8868\u6027\u51b3\u7b56\u3002\u8868 6 \u603b\u7ed3\u4e86\u53d1\u5e03\u8f68\u8ff9\u4e2d\u7684\u82e5\u5e72\u4f8b\u5b50\u3002\u5b8c\u6574\u7684\u9010\u8f6e\u8bb0\u5f55\u2014\u2014\u5305\u62ec\u63d0\u6848\u3001critic \u5224\u5b9a\u3001\u63a5\u53d7\u51b3\u7b56\u4e0e\u7cbe\u786e harness diff\u2014\u2014\u89c1\u9879\u76ee\u7f51\u7ad9\u3002<\/p>\n
\u8fd9\u4e9b\u4f8b\u5b50\u8ba9\u6211\u4eec\u66f4\u5177\u4f53\u5730\u770b\u5230\u6b63\u5219\u5316\u884c\u4e3a\u3002\u7279\u522b\u662f\u7f16\u7801\u7b2c 0 \u8f6e\u7684\u4e24\u4e2a\u5019\u9009\u8868\u9762\u4e0a\u76f8\u4f3c，\u4f46\u53ea\u6709\u90a3\u4e2a\u6d4b\u5f97\u6539\u8fdb\u8db3\u591f\u5927\u7684\u5019\u9009\u901a\u8fc7\u4e86\u6210\u672c\u611f\u77e5\u9009\u62e9\u89c4\u5219\u3002\u53cd\u4e4b，\u7b2c 8 \u8f6e\u7684\u5019\u9009\u964d\u4f4e\u4e86\u63a8\u7406\u6210\u672c，\u5374\u4ecd\u88ab\u62d2\u7edd，\u56e0\u4e3a\u5176\u6027\u80fd\u843d\u5728\u53ef\u63a5\u53d7\u5e95\u7ebf\u4e4b\u4e0b\u3002\u5de5\u7a0b\u4f8b\u5b50\u5219\u5c55\u793a\u4e86\u4e92\u8865\u60c5\u5f62：\u4e00\u4e2a\u5c0f\u800c\u53ef\u590d\u7528\u7684\u63a7\u5236\u6d41\u4fee\u6b63\u88ab\u4fdd\u7559\u4e0b\u6765，\u8d44\u6e90\u51e0\u4e4e\u6ca1\u6709\u589e\u957f\u3002\u8fd9\u4e9b\u8f68\u8ff9\u5171\u540c\u8868\u660e：RRSI \u5e76\u4e0d\u7b80\u5355\u5730\u7d2f\u79ef\u4e00\u5207\u80fd\u63d0\u9ad8 evolve \u96c6\u5206\u6570\u7684\u7f16\u8f91，\u800c\u662f\u9009\u62e9\u6027\u5730\u4fdd\u7559\u90a3\u4e9b"\u6d4b\u5f97\u6536\u76ca\u76f8\u5bf9\u5176\u590d\u6742\u5ea6\u8db3\u591f\u7a33\u5065"\u7684\u6539\u52a8\u3002<\/p>\n
\u8868 6｜RRSI \u4e2d\u6709\u4ee3\u8868\u6027\u7684 harness \u8fdb\u5316\u51b3\u7b56\u3002\u8fd9\u4e9b\u4f8b\u5b50\u8868\u660e\u8fdb\u5316\u5e76\u975e\u53ea\u7531\u5206\u6570\u9a71\u52a8：\u5019\u9009\u7684\u7279\u5f02\u6027\u3001\u8bc4\u4f30\u7a33\u5b9a\u6027\u4e0e\u63a8\u7406\u6210\u672c\u5171\u540c\u51b3\u5b9a\u4e00\u5904\u6539\u52a8\u662f\u5426\u88ab\u4fdd\u7559\u3002<\/p>\n
\n
\u9886\u57df\/\u8f6e\u6b21Harness \u6539\u52a8\u7ed3\u679c\u8bf4\u660e\u4e86\u4ec0\u4e48<\/tr>\n \n\n \u7f16\u7801，R0-A<\/td>\n \u589e\u52a0\u4e00\u4e2a\u6709\u754c\u7684"\u5b8c\u6210\u524d\u9a8c\u8bc1\u5ba1\u8ba1"，\u5e76\u5bf9\u957f\u65f6\u8fd0\u884c\u4efb\u52a1\u63d0\u4f9b\u975e\u963b\u585e\u8f6e\u8be2\u6307\u5f15\u3002<\/td>\n \u63a5\u53d7：evolve \u96c6 +3.93 \u5206\u3002<\/td>\n \u5f53\u589e\u76ca\u8d85\u8fc7\u566a\u58f0\u9608\u503c\u65f6，\u4e00\u4e2a\u53ef\u590d\u7528\u7684\u884c\u4e3a\u673a\u5236\u53ef\u4ee5\u652f\u6491\u4e00\u6b21\u76f8\u5bf9\u5bbd\u6cdb\u7684\u65e9\u671f\u66f4\u65b0\u3002<\/td>\n<\/tr>\n \n \u7f16\u7801，R0-B<\/td>\n \u589e\u52a0\u7c7b\u4f3c\u7684\u9a8c\u8bc1\u63d0\u9192\u4e0e\u957f\u65f6\u8fd0\u884c\u6307\u5f15，\u4f46\u6d4b\u5f97\u589e\u76ca\u66f4\u5c0f\u3001\u989d\u5916\u63a8\u7406\u6210\u672c\u66f4\u9ad8\u3002<\/td>\n \u88ab\u6210\u672c\u89c4\u5219\u62d2\u7edd：+1.69 \u5206，\u6210\u672c +26.1%\u3002<\/td>\n \u5f53\u8868\u9762\u6539\u8fdb\u843d\u5728\u566a\u58f0\u5e26\u5185\u3001\u4e14\u9700\u8981\u5927\u91cf\u989d\u5916\u8ba1\u7b97\u65f6，\u4e0d\u4f1a\u88ab\u81ea\u52a8\u4fdd\u7559\u3002<\/td>\n<\/tr>\n \n \u7f16\u7801，R8-B<\/td>\n \u628a\u539f\u59cb\u4efb\u52a1\u6307\u4ee4\u9489\u6b7b\u5728"\u5b8c\u6210\u95e8"\u4e2d，\u4f7f\u7b56\u7565\u5728\u63d0\u4ea4\u524d\u91cd\u65b0\u6838\u5bf9\u5b57\u9762\u89c4\u683c\u3002<\/td>\n \u88ab\u5e95\u7ebf\u62d2\u7edd：\u22122.81 \u5206，\u5c3d\u7ba1\u6210\u672c \u221213.6%\u3002<\/td>\n \u5bf9\u4e00\u4e2a\u6027\u80fd\u843d\u5728\u566a\u58f0\u8c03\u6574\u540e\u63a5\u53d7\u5e95\u7ebf\u4e4b\u4e0b\u7684\u5019\u9009，\u5355\u9760\u66f4\u4f4e\u6210\u672c\u65e0\u6cd5\u8865\u507f\u3002<\/td>\n<\/tr>\n \n \u5de5\u7a0b\u8bbe\u8ba1，R2<\/td>\n \u4e3a\u53cd\u590d\u51fa\u73b0\u7684 \u201cworkdir must be an existing directory\u201d \u5de5\u5177\u4f7f\u7528\u9519\u8bef\u589e\u52a0\u4e00\u4e2a\u6709\u754c\u6062\u590d\u63d0\u793a\u3002<\/td>\n \u63a5\u53d7：122\/244 \u2192 128\/244 \u901a\u8fc7，token +1.6%\u3002<\/td>\n \u641c\u7d22\u53ef\u4ee5\u4fdd\u7559\u90a3\u4e9b\u5c0f\u800c\u4e0e\u4efb\u52a1\u65e0\u5173\u3001\u5374\u4ee5\u6781\u5c11\u989d\u5916\u590d\u6742\u5ea6\u63d0\u5347\u53ef\u9760\u6027\u7684\u63a7\u5236\u6d41\u4fee\u6b63\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
\n（\u5168\u6587\u5b8c\u3002\u539f\u6587\u5171 24 \u9875，\u542b\u56fe 1\u20134 \u4e0e\u8868 1\u20136）<\/p>\n","protected":false},"excerpt":{"rendered":"
RRSI \u7684\u6838\u5fc3\u8d21\u732e\u4e0d\u5728\u4e8e\\”\u8ba9 harness \u6f14\u5316\u5f97\u66f4\u5f3a\\”，\u800c\u5728\u4e8e\u628a harness \u6f14\u5316\u4ece\\”\u5728\u6f14\u5316\u96c6\u4e0a\u5237\u5206\\”\u6539\u6210\\”\u5728\u5206\u5e03\u5916\u4ecd\u6210\u7acb\u7684\u673a\u5236\u641c\u7d22\\”\u3002\u6700\u786c\u7684\u8bc1\u636e\u6765\u81ea agentic workspace：OOD \u5747\u5206\u7531 H0 \u7684 39.7 \u63d0\u5347\u81f3 43.6，\u800c\u5404\u57fa\u7ebf\u5747\u503c\u7ea6 39.4；\u8868 2 \u540c\u65f6\u663e\u793a，\u53bb\u6389\u4efb\u4e00\u4fa7\u6b63\u5219\u90fd\u4f1a\u62ac\u9ad8\u6f14\u5316\u96c6\u5206\u6570\u5374\u538b\u4f4e OOD \u8868\u73b0\u3002\u4ee3\u4ef7\u4e5f\u6e05\u695a：\u6700\u7ec8 harness \u4e3a 2.42<\/p>\n","protected":false},"author":2,"featured_media":111121,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[10848,11999,11653,74,66,50],"topic":[],"class_list":["post-111126","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-harness","tag-rsi","tag-11653","tag-agent","tag-ai","tag-50"],"yoast_head":"\n
\u3010\u6df1\u5ea6\u89e3\u8bfb\u8c37\u6b4c\u6700\u65b0\u8bba\u6587\u3011RRSI \u667a\u80fd\u4f53 Harness \u7684\u6b63\u5219\u5316\u9012\u5f52\u81ea\u6211\u6539\u8fdb - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/111126.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u3010\u6df1\u5ea6\u89e3\u8bfb\u8c37\u6b4c\u6700\u65b0\u8bba\u6587\u3011RRSI \u667a\u80fd\u4f53 Harness \u7684\u6b63\u5219\u5316\u9012\u5f52\u81ea\u6211\u6539\u8fdb - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"RRSI \u7684\u6838\u5fc3\u8d21\u732e\u4e0d\u5728\u4e8e"\u8ba9 harness \u6f14\u5316\u5f97\u66f4\u5f3a"，\u800c\u5728\u4e8e\u628a harness \u6f14\u5316\u4ece"\u5728\u6f14\u5316\u96c6\u4e0a\u5237\u5206"\u6539\u6210"\u5728\u5206\u5e03\u5916\u4ecd\u6210\u7acb\u7684\u673a\u5236\u641c\u7d22"\u3002\u6700\u786c\u7684\u8bc1\u636e\u6765\u81ea agentic workspace：OOD \u5747\u5206\u7531 H0 \u7684 39.7 \u63d0\u5347\u81f3 43.6，\u800c\u5404\u57fa\u7ebf\u5747\u503c\u7ea6 39.4；\u8868 2 \u540c\u65f6\u663e\u793a，\u53bb\u6389\u4efb\u4e00\u4fa7\u6b63\u5219\u90fd\u4f1a\u62ac\u9ad8\u6f14\u5316\u96c6\u5206\u6570\u5374\u538b\u4f4e OOD \u8868\u73b0\u3002\u4ee3\u4ef7\u4e5f\u6e05\u695a：\u6700\u7ec8 harness \u4e3a 2.42\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/111126.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-09-30T02:57:21+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260930025715-6abc7a8b2cb0f.png\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"28 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/111126.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/111126.html\",\"name\":\"\u3010\u6df1\u5ea6\u89e3\u8bfb\u8c37\u6b4c\u6700\u65b0\u8bba\u6587\u3011RRSI \u667a\u80fd\u4f53 Harness \u7684\u6b63\u5219\u5316\u9012\u5f52\u81ea\u6211\u6539\u8fdb - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-09-30T02:57:21+00:00\",\"dateModified\":\"2026-09-30T02:57:21+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/111126.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/111126.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/111126.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u3010\u6df1\u5ea6\u89e3\u8bfb\u8c37\u6b4c\u6700\u65b0\u8bba\u6587\u3011RRSI \u667a\u80fd\u4f53 Harness \u7684\u6b63\u5219\u5316\u9012\u5f52\u81ea\u6211\u6539\u8fdb\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u3010\u6df1\u5ea6\u89e3\u8bfb\u8c37\u6b4c\u6700\u65b0\u8bba\u6587\u3011RRSI \u667a\u80fd\u4f53 Harness \u7684\u6b63\u5219\u5316\u9012\u5f52\u81ea\u6211\u6539\u8fdb - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/111126.html","og_locale":"zh_CN","og_type":"article","og_title":"\u3010\u6df1\u5ea6\u89e3\u8bfb\u8c37\u6b4c\u6700\u65b0\u8bba\u6587\u3011RRSI \u667a\u80fd\u4f53 Harness \u7684\u6b63\u5219\u5316\u9012\u5f52\u81ea\u6211\u6539\u8fdb - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"RRSI \u7684\u6838\u5fc3\u8d21\u732e\u4e0d\u5728\u4e8e\"\u8ba9 harness \u6f14\u5316\u5f97\u66f4\u5f3a\"，\u800c\u5728\u4e8e\u628a harness \u6f14\u5316\u4ece\"\u5728\u6f14\u5316\u96c6\u4e0a\u5237\u5206\"\u6539\u6210\"\u5728\u5206\u5e03\u5916\u4ecd\u6210\u7acb\u7684\u673a\u5236\u641c\u7d22\"\u3002\u6700\u786c\u7684\u8bc1\u636e\u6765\u81ea agentic workspace：OOD \u5747\u5206\u7531 H0 \u7684 39.7 \u63d0\u5347\u81f3 43.6，\u800c\u5404\u57fa\u7ebf\u5747\u503c\u7ea6 39.4；\u8868 2 \u540c\u65f6\u663e\u793a，\u53bb\u6389\u4efb\u4e00\u4fa7\u6b63\u5219\u90fd\u4f1a\u62ac\u9ad8\u6f14\u5316\u96c6\u5206\u6570\u5374\u538b\u4f4e OOD \u8868\u73b0\u3002\u4ee3\u4ef7\u4e5f\u6e05\u695a：\u6700\u7ec8 harness \u4e3a 2.42","og_url":"https:\/\/www.wsisp.com\/helps\/111126.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-09-30T02:57:21+00:00","og_image":[{"url":"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260930025715-6abc7a8b2cb0f.png"}],"author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"28 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/111126.html","url":"https:\/\/www.wsisp.com\/helps\/111126.html","name":"\u3010\u6df1\u5ea6\u89e3\u8bfb\u8c37\u6b4c\u6700\u65b0\u8bba\u6587\u3011RRSI \u667a\u80fd\u4f53 Harness \u7684\u6b63\u5219\u5316\u9012\u5f52\u81ea\u6211\u6539\u8fdb - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-09-30T02:57:21+00:00","dateModified":"2026-09-30T02:57:21+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/111126.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/111126.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/111126.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"\u3010\u6df1\u5ea6\u89e3\u8bfb\u8c37\u6b4c\u6700\u65b0\u8bba\u6587\u3011RRSI \u667a\u80fd\u4f53 Harness \u7684\u6b63\u5219\u5316\u9012\u5f52\u81ea\u6211\u6539\u8fdb"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/111126","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=111126"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/111126\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media\/111121"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=111126"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=111126"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=111126"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=111126"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}
- \u63d0\u51fa\u4fa7\u6b63\u5219\u5316（Proposal-side）\n
- \n
- \n
- \n
- \n
- \n
- DeepThink v1.3.0 \u2014 \u4e91\u539f\u751f\u4e0e\u534f\u4f5c<\/li>\n<\/ul>\n
- DeepThink v1.4.0 \u2014 \u5206\u5e03\u5f0f\u80fd\u529b\u843d\u5730\u4e0e\u4f01\u4e1a\u534f\u4f5c\u5de5\u4f5c\u8d1f\u8f7d<\/li>\n<\/ul>\n
- DeepThink v1.5.0 \u2014 Agent \u7fa4\u7ec4\u534f\u4f5c（Swarm）\u6267\u884c\u94fe\u8def\u8d2f\u901a<\/li>\n<\/ul>\n