\nProgram Synthesis<\/p>\n<\/li>\n<\/ul>\n
\u7ed3\u679c：<\/p>\n
Math：
\nGTPO > GRPO \u2248 +3.0%<\/p>\n
Commonsense：
\n\u2248 +3.9%<\/p>\n
Program Synthesis：
\n\u540c\u6837\u53d6\u5f97\u63d0\u5347<\/p>\n
\u540c\u65f6\u8bba\u6587\u5f3a\u8c03：<\/p>\n
GTPO\u6ca1\u6709\u5f15\u5165\u660e\u663e\u989d\u5916\u8ba1\u7b97\u5f00\u9500\u3002<\/p>\n
\n3.7 \u4e00\u53e5\u8bdd\u7406\u89e3<\/h2>\n
GTPO\u89e3\u51b3\u7684\u662f\u201c\u591a\u8f6eRL\u77e5\u9053\u6700\u7ec8\u7ed3\u679c，\u4f46\u4e0d\u77e5\u9053\u6bcf\u4e00\u6b65\u5230\u5e95\u597d\u4e0d\u597d\u201d\u7684\u95ee\u9898\u3002<\/p>\n
\n4. Consistently Simulating Human Personas：\u5982\u4f55\u6784\u5efa\u771f\u6b63\u7a33\u5b9a\u7684\u201c\u4eba\u7c7b\u6a21\u62df\u5668\u201d？<\/h2>\n4.1 \u8bba\u6587\u4fe1\u606f<\/h3>\n
Consistently Simulating Human Personas with Multi-Turn Reinforcement Learning<\/p>\n
arXiv: 2511.00222<\/p>\n
\u8be5\u5de5\u4f5c\u53d1\u8868\u4e8e NeurIPS 2025\u3002<\/p>\n
ArXiv \u8bba\u6587\u4e3b\u9875<\/p>\n
\n4.2 \u4e3a\u4ec0\u4e48\u9700\u8981Human Simulator？<\/h2>\n
\u5982\u679c\u6211\u4eec\u60f3\u8bad\u7ec3\u4e00\u4e2a\u591a\u8f6e\u5bf9\u8bddAgent：<\/p>\n
Agent
\n \u2195
\nUser<\/p>\n
\u771f\u6b63\u7684\u7528\u6237\u4e0d\u53ef\u80fd\u6c38\u8fdc\u5728\u7ebf\u3002<\/p>\n
\u56e0\u6b64\u5f88\u591a\u7814\u7a76\u4f1a：<\/p>\n
LLM
\n\u2193
\n\u6a21\u62dfUser<\/p>\n
\u4e8e\u662f：<\/p>\n
Agent
\n \u2195
\nLLM User Simulator<\/p>\n
\u95ee\u9898\u6765\u4e86：<\/p>\n
LLM\u867d\u7136\u8bed\u8a00\u80fd\u529b\u5f88\u5f3a，\u4f46\u5b83\u771f\u7684\u80fd\u4e00\u76f4\u201c\u6f14\u597d\u4e00\u4e2a\u4eba\u201d\u5417？<\/p>\n
\u7b54\u6848\u5f80\u5f80\u662f\u5426\u5b9a\u7684\u3002<\/p>\n
\n4.3 Persona Drift<\/h2>\n
\u4f8b\u5982\u7ed9\u6a21\u578b：<\/p>\n
\u4f60\u662f\u4e00\u540d\u5bf9\u6570\u5b66\u6ca1\u6709\u5174\u8da3\u7684\u9ad8\u4e2d\u751f\u3002
\n\u4f60\u5bb3\u6015\u8003\u8bd5\u3002
\n\u4f60\u66f4\u559c\u6b22\u4f53\u80b2\u3002<\/p>\n
\u5f00\u59cb：<\/p>\n
User：
\n\u6211\u771f\u7684\u5f88\u8ba8\u538c\u6570\u5b66\u3002<\/p>\n
10\u8f6e\u4ee5\u540e：<\/p>\n
User：
\n\u6211\u6700\u559c\u6b22\u6570\u5b66\u4e86！<\/p>\n
\u8fd9\u5c31\u662f：<\/p>\n
Persona Drift<\/p>\n
\u66f4\u4e25\u91cd\u7684\u662f：<\/p>\n
Turn 2：
\n\u6211\u4ece\u6765\u6ca1\u53bb\u8fc7\u7f8e\u56fd\u3002<\/p>\n
Turn 15：
\n\u6211\u53bb\u5e74\u5728\u7ebd\u7ea6\u65c5\u884c\u3002<\/p>\n
\u4ea7\u751f：<\/p>\n
Self-Contradiction<\/p>\n
\n4.4 \u8bba\u6587\u63d0\u51fa\u4e09\u4e2aConsistency\u6307\u6807<\/h2>\n
\u8fd9\u662f\u8fd9\u7bc7\u8bba\u6587\u6700\u503c\u5f97\u501f\u9274\u7684\u5730\u65b9\u3002<\/p>\n
\n\u6307\u6807\u4e00：Prompt-to-Line Consistency<\/h3>\n
\u68c0\u67e5：<\/p>\n
\u5f53\u524d\u56de\u7b54\u662f\u5426\u7b26\u5408\u6700\u521d\u7684 Persona？<\/p>\n
\u5f62\u5f0f\u4e0a：<\/p>\n
Persona
\n \u2193
\nCurrent Response
\n \u2193
\nConsistency<\/p>\n
\u4f8b\u5982：<\/p>\n
Persona：
\n\u8ba8\u538c\u6570\u5b66<\/p>\n
Response：
\n\u6211\u5f88\u559c\u6b22\u6570\u5b66\u3002<\/p>\n
\u2192 Inconsistent<\/p>\n
\n\u6307\u6807\u4e8c：Line-to-Line Consistency<\/h3>\n
\u68c0\u67e5：<\/p>\n
\u5f53\u524d\u56de\u7b54\u548c\u8fc7\u53bb\u8bf4\u8fc7\u7684\u8bdd\u662f\u5426\u77db\u76fe？<\/p>\n
\u4f8b\u5982：<\/p>\n
Turn 3：
\n\u6211\u6ca1\u6709\u5ba0\u7269\u3002<\/p>\n
Turn 15：
\n\u6211\u7684\u72d7\u4eca\u5929\u751f\u75c5\u4e86\u3002<\/p>\n
\u2192 Contradiction<\/p>\n
\u5b83\u5173\u6ce8\u7684\u662f：<\/p>\n
Past Utterance
\n \u2195
\nCurrent Utterance<\/p>\n
\n\u6307\u6807\u4e09：Q&A Consistency<\/h3>\n
\u8fd9\u662f\u975e\u5e38\u6709\u610f\u601d\u7684\u4e00\u79cd\u65b9\u6cd5\u3002<\/p>\n
\u7cfb\u7edf\u6839\u636e Persona \u751f\u6210\u4e00\u4e9b\u95ee\u9898：<\/p>\n
\u4f60\u6700\u559c\u6b22\u4ec0\u4e48？
\n\u4f60\u662f\u5426\u559c\u6b22\u6570\u5b66？
\n\u4f60\u4f4f\u5728\u54ea\u91cc？<\/p>\n
\u7136\u540e\u4ece\u5bf9\u8bdd\u4e2d\u63a8\u65ad\u6a21\u578b\u7684\u7b54\u6848，\u518d\u4e0e Persona \u4e2d\u7684\u6b63\u786e\u7b54\u6848\u6bd4\u8f83\u3002<\/p>\n
\u4e5f\u5c31\u662f：<\/p>\n
Persona
\n \u2193
\nReference Answer<\/p>\n
Dialogue
\n \u2193
\nInferred Answer<\/p>\n
Reference
\n vs
\nInferred<\/p>\n
\u8fd9\u6837\u53ef\u4ee5\u5224\u65ad\u6a21\u578b\u957f\u671f\u7ef4\u6301\u7684：<\/p>\n
Belief
\nIdentity
\nPreference<\/p>\n
\u662f\u5426\u7a33\u5b9a\u3002<\/p>\n
\u76f8\u5173\u65b9\u6cd5\u7ec6\u8282\u548c\u4e09\u4e2a\u6307\u6807\u7684\u5b9a\u4e49\u4e5f\u5728\u8bba\u6587\u9644\u5f55\u4e2d\u8fdb\u884c\u4e86\u8bf4\u660e\u3002<\/p>\n
\n4.5 \u66f4\u5173\u952e\u7684\u4e00\u6b65：\u628aConsistency\u53d8\u6210Reward<\/h2>\n
\u8fd9\u7bc7\u8bba\u6587\u6ca1\u6709\u505c\u7559\u5728：<\/p>\n
Evaluation<\/p>\n
\u800c\u662f：<\/p>\n
Consistency Metric
\n \u2193
\nReward
\n \u2193
\nMulti-turn RL
\n \u2193
\nBetter User Simulator<\/p>\n
\u4e8e\u662f\u6574\u4e2a\u95ed\u73af：<\/p>\n
Persona
\n \u2193
\nLLM User Simulator
\n \u2193
\nMulti-turn Dialogue
\n \u2193
\nConsistency Evaluation
\n \u2193
\nReward
\n \u2193
\nRL
\n \u2193
\nMore Consistent Simulator<\/p>\n
\n4.6 \u5b9e\u9a8c\u4f7f\u7528\u4e86\u4e09\u7c7bUser<\/h2>\n
\u8bba\u6587\u8bad\u7ec3\u4e09\u4e2a\u7528\u6237\u89d2\u8272：<\/p>\n
Patient
\nStudent
\nSocial Chat Partner<\/p>\n
\u8fd9\u975e\u5e38\u91cd\u8981，\u56e0\u4e3a\u5982\u679c\u53ea\u9a8c\u8bc1\u4e00\u4e2aPersona，\u5f88\u96be\u8bc1\u660e\u65b9\u6cd5\u5177\u6709\u6cdb\u5316\u6027\u3002<\/p>\n
\n4.7 \u5b9e\u9a8c\u7ed3\u679c<\/h2>\n
\u8bba\u6587\u62a5\u544a：<\/p>\n
Multi-turn RL \u4f7f Persona inconsistency \u964d\u4f4e\u8d85\u8fc7 55%\u3002<\/p>\n
\u5e76\u4e14\u4e09\u4e2aConsistency\u6307\u6807\u90fd\u7ecf\u8fc7\u4e86\u4eba\u5de5\u6807\u6ce8\u9a8c\u8bc1\u3002<\/p>\n
\n4.8 \u548c\u666e\u901aPersona Prompt\u7684\u533a\u522b<\/h2>\n
\u4f20\u7edf\u65b9\u6cd5：<\/p>\n
System Prompt
\n+
\nPersona
\n\u2193
\nLLM<\/p>\n
\u8fd9\u76f8\u5f53\u4e8e：<\/p>\n
\u201c\u544a\u8bc9\u6a21\u578b\u4f60\u662f\u8c01\u3002\u201d<\/p>\n
\u672c\u6587：<\/p>\n
Persona
\n \u2193
\nDialogue
\n \u2193
\nConsistency Reward
\n \u2193
\nRL<\/p>\n
\u76f8\u5f53\u4e8e：<\/p>\n
\u201c\u901a\u8fc7\u8bad\u7ec3\u8ba9\u6a21\u578b\u771f\u6b63\u5b66\u4f1a\u6301\u7eed\u4fdd\u6301\u8fd9\u4e2a\u8eab\u4efd\u3002\u201d<\/p>\n
\u8fd9\u4e24\u8005\u662f\u5b8c\u5168\u4e0d\u540c\u7684\u3002<\/p>\n
\n5. EvolIF：\u4f20\u7edfMulti-Turn Benchmark\u4e3a\u4ec0\u4e48\u201c\u4e0d\u591f\u771f\u5b9e\u201d？<\/h2>\n5.1 \u8bba\u6587\u4fe1\u606f<\/h3>\n
One Battle After Another: Probing LLMs' Limits on Multi-Turn Instruction Following with a Benchmark Evolving Framework<\/p>\n
arXiv: 2511.03508<\/p>\n
\u8be5\u5de5\u4f5c\u540e\u6765\u53d1\u8868\u5728 ACL 2026 Long Papers\u3002<\/p>\n
ArXiv \u8bba\u6587\u4e3b\u9875<\/p>\n
\n5.2 \u5b83\u89e3\u51b3\u7684\u95ee\u9898<\/h2>\n
\u4f20\u7edf\u591a\u8f6eBenchmark\u901a\u5e38\u662f：<\/p>\n
Turn 1
\nTurn 2
\nTurn 3
\n…
\nTurn N<\/p>\n
\u95ee\u9898\u662f：<\/p>\n
N\u662f\u63d0\u524d\u89c4\u5b9a\u597d\u7684\u3002<\/p>\n
\u4f8b\u5982：<\/p>\n
\u6700\u591a10\u8f6e<\/p>\n
\u65e0\u8bba\u6a21\u578b\u8868\u73b0\u597d\u4e0d\u597d：<\/p>\n
10\u8f6e\u7ed3\u675f<\/p>\n
\u8fd9\u548c\u771f\u5b9e\u7528\u6237\u5e76\u4e0d\u4e00\u6837\u3002<\/p>\n
\u73b0\u5b9e\u60c5\u51b5\u662f：<\/p>\n
\u6a21\u578b\u8868\u73b0\u5f88\u597d
\n\u2192 \u7528\u6237\u7ee7\u7eed\u804a<\/p>\n
\u6a21\u578b\u5f00\u59cb\u72af\u9519
\n\u2192 \u7528\u6237\u7ea0\u6b63<\/p>\n
\u6a21\u578b\u7ee7\u7eed\u72af\u9519
\n\u2192 \u7528\u6237\u5931\u53bb\u8010\u5fc3
\n\u2192 \u79bb\u5f00<\/p>\n
\u6240\u4ee5：<\/p>\n
\u771f\u5b9e\u5bf9\u8bdd\u957f\u5ea6\u53d6\u51b3\u4e8e\u6a21\u578b\u8868\u73b0\u3002<\/p>\n
\n5.3 EvolIF\u7684\u6838\u5fc3\u601d\u60f3<\/h2>\n
EvolIF\u63d0\u51fa：<\/p>\n
\u8ba9Benchmark\u81ea\u5df1\u201c\u8fdb\u5316\u201d\u3002<\/p>\n
\u6838\u5fc3\u7ed3\u6784\u5305\u62ec：<\/p>\n
User Intent
\n \u2193
\nConstraint Tracking
\n \u2193
\nInstruction Tracking
\n \u2193
\nTopic Tracking
\n \u2193
\nQuery Synthesis Agent
\n \u2193
\nUser Query
\n \u2193
\nLLM Response
\n \u2193
\nState Update
\n \u2193
\nNext User Query<\/p>\n
\u8bba\u6587\u4f7f\u7528\u4e09\u5c42 tracking mechanism \u6765\u8ddf\u8e2a：<\/p>\n
Constraints
\nInstructions
\nTopics<\/p>\n
\u7136\u540e\u52a8\u6001\u751f\u6210\u4e0b\u4e00\u8f6e\u7528\u6237\u884c\u4e3a\u3002<\/p>\n
\n5.4 \u4e3a\u4ec0\u4e48\u53ebEvolving？<\/h2>\n
\u56e0\u4e3a：<\/p>\n
State_t
\n \u2193
\nLLM Response
\n \u2193
\nUser Reaction
\n \u2193
\nState_{t+1}
\n \u2193
\nNew Constraint
\n \u2193
\nNew Topic
\n \u2193
\nNew Instruction<\/p>\n
\u4e5f\u5c31\u662f\u8bf4：<\/p>\n
\u4e0b\u4e00\u8f6e\u4e0d\u662f\u63d0\u524d\u5199\u597d\u7684\u3002<\/p>\n
\u800c\u662f\u7531\u5f53\u524d\u4ea4\u4e92\u72b6\u6001\u51b3\u5b9a\u3002<\/p>\n
\n5.5 \u6700\u91cd\u8981\u7684\u521b\u65b0：User Patience<\/h2>\n
EvolIF\u501f\u9274 Flow Theory，\u5f15\u5165\u7528\u6237\u8010\u5fc3\u7684\u6982\u5ff5\u3002<\/p>\n
\u4f20\u7edfBenchmark：<\/p>\n
Fixed Turns<\/p>\n
EvolIF：<\/p>\n
Model performs well
\n \u2193
\nUser continues<\/p>\n
Model fails
\n \u2193
\nUser gives correction<\/p>\n
Repeated failure
\n \u2193
\nUser patience decreases<\/p>\n
Patience exhausted
\n \u2193
\nConversation ends<\/p>\n
\u56e0\u6b64：<\/p>\n
\u5bf9\u8bdd\u7ec8\u6b62\u6761\u4ef6\u4ece\u201c\u56fa\u5b9aTurn\u6570\u201d\u53d8\u6210\u201c\u7528\u6237\u662f\u5426\u8fd8\u613f\u610f\u7ee7\u7eed\u201d\u3002<\/p>\n
\u8fd9\u4e00\u6b65\u975e\u5e38\u63a5\u8fd1\u771f\u5b9eHuman-Agent Interaction\u3002<\/p>\n
\n5.6 \u5b9e\u9a8c\u9a8c\u8bc1\u8def\u7ebf<\/h2>\n
EvolIF\u5305\u542b\u591a\u4e2a\u7ea6\u675f\u7c7b\u522b，\u5f53\u524d\u7248\u672c\u8986\u76d6 12\u4e2a constraint groups\u3002<\/p>\n
\u8bba\u6587\u6d4b\u8bd5：<\/p>\n
GPT-5
\nGemini-3-Pro
\nMiniMax-M2
\nKimi-K2
\nQwen3-235B
\nGrok-4-Fast
\nDeepSeek-V3.2
\nSeed-1.6
\nLlama-4-Maverick
\nMistral-Large-3<\/p>\n
\u7ed3\u679c\u4e2d：<\/p>\n
GPT-5
\nRobustness = 66.40%<\/p>\n
\u9ad8\u4e8e：<\/p>\n
Gemini-3.0-Pro = 60.81%<\/p>\n
\u5dee\u8ddd：<\/p>\n
5.59 percentage points<\/p>\n
\n5.7 EvolIF\u771f\u6b63\u91cd\u8981\u7684\u5730\u65b9<\/h2>\n
\u8fd9\u7bc7\u8bba\u6587\u7684\u4ef7\u503c\u5176\u5b9e\u4e0d\u53ea\u662f：<\/p>\n
\u201c\u63d0\u51fa\u4e00\u4e2aBenchmark\u3002\u201d<\/p>\n
\u800c\u662f\u63d0\u51fa\u4e00\u79cd\u65b0\u7684\u8bc4\u6d4b\u601d\u60f3：<\/p>\n
Traditional Evaluation<\/h4>\n
Input
\n \u2193
\nN Turns
\n \u2193
\nScore<\/p>\n
EvolIF<\/h4>\n
User State
\n \u2193
\nInteraction
\n \u2193
\nModel Response
\n \u2193
\nUser Reaction
\n \u2193
\nState Evolution
\n \u2193
\nNext Interaction
\n \u2193
\n…
\n \u2193
\nUser Patience Exhausted<\/p>\n
\u56e0\u6b64\u5b83\u8bc4\u4ef7\u7684\u4e0d\u662f\u5355\u7eaf：<\/p>\n
\u201c\u6a21\u578b\u7b2cN\u8f6e\u56de\u7b54\u5f97\u597d\u4e0d\u597d？\u201d<\/p>\n
\u800c\u662f：<\/p>\n
\u201c\u6a21\u578b\u80fd\u4e0d\u80fd\u8ba9\u7528\u6237\u613f\u610f\u7ee7\u7eed\u548c\u5b83\u4ea4\u4e92？\u201d<\/p>\n
\u8fd9\u662f\u591a\u8f6e\u5bf9\u8bdd\u8bc4\u4ef7\u975e\u5e38\u91cd\u8981\u7684\u8f6c\u53d8\u3002<\/p>\n
\n6. FunReason-MT：\u9ad8\u8d28\u91cf\u591a\u8f6eTool-use\u6570\u636e\u5230\u5e95\u600e\u4e48\u6784\u9020？<\/h2>\n6.1 \u8bba\u6587\u4fe1\u606f<\/h3>\n
FunReason-MT Technical Report: Advanced Data Synthesis Solution for Real-world Multi-Turn Tool-use<\/p>\n
arXiv: 2510.24645<\/p>\n
ArXiv \u8bba\u6587\u4e3b\u9875<\/p>\n
\n6.2 \u5b83\u89e3\u51b3\u4ec0\u4e48\u95ee\u9898？<\/h2>\n
\u73b0\u5728\u5f88\u591aAgent\u8bad\u7ec3\u7684\u95ee\u9898\u4e0d\u662f：<\/p>\n
\u201c\u6ca1\u6709\u6a21\u578b\u3002\u201d<\/p>\n
\u800c\u662f：<\/p>\n
\u6ca1\u6709\u8db3\u591f\u597d\u7684\u591a\u8f6e Tool-use \u6570\u636e\u3002<\/p>\n
\u4f20\u7edf\u6570\u636e\u751f\u6210：<\/p>\n
Random Environment Sampling<\/p>\n
\u6216\u8005：<\/p>\n
Multi-Agent Role Playing<\/p>\n
\u5bb9\u6613\u751f\u6210：<\/p>\n
\u7b80\u5355Tool Call
\n\u7b80\u5355Task
\n\u4f4e\u903b\u8f91\u4f9d\u8d56<\/p>\n
\u4f46\u771f\u5b9e\u4efb\u52a1\u53ef\u80fd\u662f：<\/p>\n
User Goal
\n \u2193
\nSearch
\n \u2193
\nTool Result
\n \u2193
\nReason
\n \u2193
\nCall Tool B
\n \u2193
\nTool Result
\n \u2193
\n\u4fee\u6539\u7b56\u7565
\n \u2193
\nCall Tool C
\n \u2193
\nVerification
\n \u2193
\nFinal Answer<\/p>\n
\u4e5f\u5c31\u662f\u8bf4：<\/p>\n
\u771f\u6b63\u56f0\u96be\u7684\u662fMulti-turn Logical Dependency\u3002<\/p>\n
\n6.3 FunReason-MT\u7684\u4e09\u9636\u6bb5<\/h2>\n
\u8bba\u6587\u63d0\u51fa\u4e09\u4e2a\u6838\u5fc3\u7ec4\u4ef6\u3002<\/p>\n
\n\u7b2c\u4e00\u9636\u6bb5：Environment-API Graph Interaction<\/h3>\n
\u9996\u5148\u4e0d\u518d\u968f\u673a\u751f\u6210 Tool Call\u3002<\/p>\n
\u800c\u662f\u5efa\u7acb：<\/p>\n
Environment
\n \u2193
\nAvailable APIs
\n \u2193
\nAPI Graph<\/p>\n
\u4f8b\u5982：<\/p>\n
Search Flight
\n \u2193
\nGet Flight Detail
\n \u2193
\nBook Flight
\n \u2193
\nPayment<\/p>\n
\u8fd9\u6837\u53ef\u4ee5\u4ea7\u751f：<\/p>\n
\u6709\u771f\u5b9e\u4f9d\u8d56\u5173\u7cfb\u7684Tool trajectory\u3002<\/p>\n
\n6.4 \u7b2c\u4e8c\u9636\u6bb5：Advanced Tool-Query Synthesis<\/h2>\n
\u666e\u901a\u65b9\u6cd5\u53ef\u80fd：<\/p>\n
\u968f\u673a\u751f\u6210User Query<\/p>\n
\u7ed3\u679c：<\/p>\n
Query
\n\u2193
\n\u4e00\u4e2aTool
\n\u2193
\n\u7ed3\u675f<\/p>\n
FunReason-MT\u5e0c\u671b\u6784\u9020：<\/p>\n
\u590d\u6742User Goal
\n \u2193
\n\u591a\u4e2a\u5de5\u5177
\n \u2193
\n\u591a\u4e2a\u4f9d\u8d56
\n \u2193
\n\u591a\u8f6e\u4ea4\u4e92<\/p>\n
\u56e0\u6b64\u5b83\u4e13\u95e8\u8bbe\u8ba1 Tool-Query Synthesis\u3002<\/p>\n
\n6.5 \u7b2c\u4e09\u9636\u6bb5：Guided Iterative Chain<\/h2>\n
\u6700\u7ec8\u8fd8\u9700\u8981\u8ba9\u6a21\u578b\u4ea7\u751f：<\/p>\n
Reasoning
\n+
\nTool Call
\n+
\nTool Result
\n+
\nNext Reasoning<\/p>\n
\u56e0\u6b64\u91c7\u7528 Guided Iterative Chain：<\/p>\n
Goal
\n \u2193
\nReason
\n \u2193
\nTool Call
\n \u2193
\nObservation
\n \u2193
\nReason
\n \u2193
\nTool Call
\n \u2193
\nObservation
\n \u2193
\nFinal<\/p>\n
\u6700\u7ec8\u5f62\u6210\u9ad8\u8d28\u91cf Multi-turn Agent Trajectory\u3002<\/p>\n
\n6.6 \u6570\u636e\u751f\u6210\u6574\u4f53Pipeline<\/h2>\n
\u53ef\u4ee5\u603b\u7ed3\u6210：<\/p>\n
Real-world Environment
\n \u2502
\n \u2193
\nEnvironment-API Graph
\n \u2502
\n \u2193
\nTargeted Tool Trajectories
\n \u2502
\n \u2193
\nAdvanced Tool-Query Synthesis
\n \u2502
\n \u2193
\nHard Multi-turn Query
\n \u2502
\n \u2193
\nGuided Iterative Chain
\n \u2502
\n \u2193
\nReasoning + Tool Calls
\n \u2502
\n \u2193
\nQuality Filtering
\n \u2502
\n \u2193
\nTraining Dataset
\n \u2502
\n \u2193
\nRL \/ SFT
\n \u2502
\n \u2193
\nAgent Model<\/p>\n
\n6.7 \u5b9e\u9a8c\u9a8c\u8bc1<\/h2>\n
\u8bba\u6587\u4f7f\u7528：<\/p>\n
Berkeley Function-Calling Leaderboard（BFCL）<\/p>\n
\u8fdb\u884c\u9a8c\u8bc1\u3002<\/p>\n
\u4e00\u4e2a\u975e\u5e38\u6709\u4ee3\u8868\u6027\u7684\u7ed3\u679c：<\/p>\n
Qwen3-4B
\nBFCLv3 Multi-turn = 15.75<\/p>\n
\u52a0\u5165 FunReason-MT \u8bad\u7ec3\u540e：<\/p>\n
Qwen3-4B + FunReason-MT
\n\u2248 57.75<\/p>\n
\u5728\u4f5c\u8005\u5f53\u524d\u516c\u5f00\u6a21\u578b\u7ed3\u679c\u4e2d，4B\u6a21\u578b\u7684\u591a\u8f6e\u6027\u80fd\u5df2\u7ecf\u8d85\u8fc7\u90e8\u5206\u66f4\u5927\u6a21\u578b；\u540c\u65f6\u5728 BFCLv4 OOD Agentic Evaluation \u4e0a\u4e5f\u8868\u73b0\u7a81\u51fa\u3002<\/p>\n
\n6.8 \u4e3a\u4ec0\u4e48\u8fd9\u7bc7\u8bba\u6587\u503c\u5f97\u5173\u6ce8？<\/h2>\n
\u56e0\u4e3a\u5b83\u8bf4\u660e：<\/p>\n
\u591a\u8f6eAgent\u80fd\u529b\u4e0d\u4ec5\u53d6\u51b3\u4e8e\u6a21\u578bArchitecture，\u8fd8\u9ad8\u5ea6\u53d6\u51b3\u4e8e\u8bad\u7ec3\u6570\u636e\u4e2d\u662f\u5426\u771f\u6b63\u5b58\u5728\u201c\u591a\u8f6e\u903b\u8f91\u201d\u3002<\/p>\n
\u5982\u679c\u8bad\u7ec3\u6570\u636e\u53ea\u662f：<\/p>\n
User
\n \u2193
\nTool
\n \u2193
\nAnswer<\/p>\n
\u6a21\u578b\u5f88\u96be\u5b66\u4f1a：<\/p>\n
Goal
\n \u2193
\nPlan
\n \u2193
\nTool
\n \u2193
\nObservation
\n \u2193
\nRe-plan
\n \u2193
\nTool
\n \u2193
\nVerification<\/p>\n
\u800c\u8fd9\u6070\u6070\u662f Agent \u4e0e\u666e\u901a Chatbot \u7684\u6838\u5fc3\u533a\u522b\u3002<\/p>\n
\n7. \u516d\u7bc7\u8bba\u6587\u653e\u5728\u4e00\u8d77，\u5230\u5e95\u6709\u4ec0\u4e48\u5173\u7cfb？<\/h2>\n
\u73b0\u5728\u628a\u516d\u7bc7\u8bba\u6587\u653e\u5728\u540c\u4e00\u5f20\u56fe\u91cc：<\/p>\n
Multi-Turn Agent
\n \u2502
\n \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510
\n \u2502 \u2502 \u2502
\n \u2193 \u2193 \u2193
\n Data Memory RL
\n \u2502 \u2502 \u2502
\n \u2193 \u2193 \u2193
\n FunReason-MT Rhea GTPO
\n \u2502 \u2502 \u2502
\n \u2502 \u2193 \u2193
\n \u2502 Context Turn-level
\n \u2502 Memory Reward
\n \u2502 \u2502 \u2502
\n \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510 \u2502 \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518
\n \u2193 \u2193 \u2193
\n H-EPM
\n \u2502
\n \u2193
\n Experience Reuse
\n \u2502
\n \u2193
\n Long-horizon Agent
\n \u2502
\n \u251c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510
\n \u2193 \u2193
\n Human Simulator Evaluation
\n \u2502 \u2502
\n \u2193 \u2193
\n Persona RL EvolIF
\n \u2502 \u2502
\n \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518
\n \u2193
\n Multi-turn System<\/p>\n
\n8. \u516d\u7bc7\u8bba\u6587\u6700\u6838\u5fc3\u7684\u5dee\u5f02<\/h2>\n
\n\u8bba\u6587\u4e3b\u8981\u89e3\u51b3\u95ee\u9898\u6838\u5fc3\u5bf9\u8c61\u65b9\u6cd5\u5173\u952e\u8bcd\u9a8c\u8bc1\u91cd\u70b9<\/tr>\n
\n\n| H-EPM<\/td>\n | \u5386\u53f2\u7ecf\u9a8c\u65e0\u6cd5\u590d\u7528<\/td>\n | Agent Memory<\/td>\n | Episodic + Procedural<\/td>\n | Tool-use \/ OOD<\/td>\n<\/tr>\n |
\n| Rhea<\/td>\n | Context\u9010\u6e10\u8870\u51cf<\/td>\n | Conversation Memory<\/td>\n | Instruction + Episodic<\/td>\n | Long Dialogue<\/td>\n<\/tr>\n |
\n| GTPO<\/td>\n | RL Reward\u592a\u7c97<\/td>\n | RL Training<\/td>\n | Turn-level Reward<\/td>\n | Multi-turn TIR<\/td>\n<\/tr>\n |
\n| Persona RL<\/td>\n | \u4eba\u7c7b\u6a21\u62df\u5668Persona\u6f02\u79fb<\/td>\n | User Simulator<\/td>\n | Consistency Reward<\/td>\n | Persona Consistency<\/td>\n<\/tr>\n |
\n| EvolIF<\/td>\n | Benchmark\u4e0d\u771f\u5b9e<\/td>\n | Evaluation<\/td>\n | Evolving User<\/td>\n | Long-horizon IF<\/td>\n<\/tr>\n |
\n| FunReason-MT<\/td>\n | \u7f3a\u4e4f\u9ad8\u8d28\u91cf\u6570\u636e<\/td>\n | Training Data<\/td>\n | API Graph + Synthesis<\/td>\n | BFCL<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
\n9. \u4e00\u4e2a\u975e\u5e38\u91cd\u8981\u7684\u7814\u7a76\u8d8b\u52bf：\u4eceContext\u8d70\u5411State<\/h2>\n\u5982\u679c\u628a\u8fd9\u4e9b\u8bba\u6587\u653e\u5728\u65f6\u95f4\u7ebf\u4e0a，\u4f1a\u53d1\u73b0\u4e00\u4e2a\u975e\u5e38\u660e\u663e\u7684\u53d8\u5316\u3002<\/p>\n \u65e9\u671f：<\/p>\n Conversation History<\/p>\n \u7814\u7a76\u91cd\u70b9：<\/p>\n \u5982\u4f55\u8ba9\u6a21\u578b\u770b\u66f4\u591a\u5386\u53f2？<\/p>\n \u968f\u540e：<\/p>\n Memory<\/p>\n \u7814\u7a76\u91cd\u70b9：<\/p>\n \u54ea\u4e9b\u5386\u53f2\u5e94\u8be5\u4fdd\u5b58？<\/p>\n \u8fdb\u4e00\u6b65：<\/p>\n Structured Memory<\/p>\n \u7814\u7a76\u91cd\u70b9：<\/p>\n \u4e0d\u540c\u5386\u53f2\u4fe1\u606f\u5e94\u8be5\u627f\u62c5\u4ec0\u4e48\u529f\u80fd？<\/p>\n \u518d\u8fdb\u4e00\u6b65：<\/p>\n State Evolution<\/p>\n \u7814\u7a76\u91cd\u70b9：<\/p>\n \u5f53\u524d\u7528\u6237\u72b6\u6001\u3001\u4efb\u52a1\u72b6\u6001\u3001\u73af\u5883\u72b6\u6001\u53d1\u751f\u4e86\u4ec0\u4e48\u53d8\u5316？<\/p>\n \u6700\u7ec8：<\/p>\n State
\n \u2193
\nAction
\n \u2193
\nEnvironment
\n \u2193
\nObservation
\n \u2193
\nState Update
\n \u2193
\nAction
\n…<\/p>\n \u8fd9\u5b9e\u9645\u4e0a\u5df2\u7ecf\u8d8a\u6765\u8d8a\u63a5\u8fd1：<\/p>\n Reinforcement Learning \/ Agentic Interaction<\/p>\n
\n10. \u5bf9\u201c\u591a\u8f6e\u5bf9\u8bdd\u6a21\u578b\u201d\u7814\u7a76\u6700\u91cd\u8981\u7684\u542f\u53d1<\/h2>\n\u5982\u679c\u4f60\u7684\u7814\u7a76\u76ee\u6807\u662f：<\/p>\n \u4efb\u52a1\u5bfc\u5411\u7684\u591a\u8f6e\u5bf9\u8bdd\u6a21\u578b \/ \u591a\u8f6e\u4ea4\u4e92Agent<\/p>\n \u90a3\u4e48\u8fd9 6 \u7bc7\u8bba\u6587\u53ef\u4ee5\u63d0\u4f9b\u4e00\u4e2a\u975e\u5e38\u5b8c\u6574\u7684\u7814\u7a76\u6846\u67b6\u3002<\/p>\n
\n10.1 \u7b2c\u4e00\u5c42：User Simulator<\/h3>\n\u9996\u5148\u9700\u8981\u89e3\u51b3：<\/p>\n \u8c01\u548cAgent\u5bf9\u8bdd？<\/p>\n \u53ef\u4ee5\u53c2\u8003：<\/p>\n Persona RL<\/p>\n \u6784\u5efa：<\/p>\n User Persona
\n+
\nUser Goal
\n+
\nUser State
\n+
\nDialogue History<\/p>\n \u5e76\u4e14\u901a\u8fc7：<\/p>\n Prompt-to-Line
\nLine-to-Line
\nQ&A Consistency<\/p>\n \u4fdd\u8bc1\u6a21\u62df\u7528\u6237\u4e0d\u4f1a\u968f\u5bf9\u8bdd\u53d1\u751f\u4e25\u91cd\u6f02\u79fb\u3002<\/p>\n
\n10.2 \u7b2c\u4e8c\u5c42：Conversation State<\/h2>\n\u7136\u540e\u89e3\u51b3：<\/p>\n Agent\u5230\u5e95\u5e94\u8be5\u8bb0\u4f4f\u4ec0\u4e48？<\/p>\n \u53ef\u4ee5\u53c2\u8003：<\/p>\n Rhea<\/p>\n \u628a\u5386\u53f2\u62c6\u6210：<\/p>\n Global Instruction
\nTask State
\nUser Preference
\nRecent Episode<\/p>\n \u800c\u4e0d\u662f\u7b80\u5355：<\/p>\n \u5168\u90e8Conversation History<\/p>\n
\n10.3 \u7b2c\u4e09\u5c42：Experience Memory<\/h2>\n\u8fdb\u4e00\u6b65\u89e3\u51b3：<\/p>\n \u4ee5\u524d\u505a\u8fc7\u7c7b\u4f3c\u4efb\u52a1\u600e\u4e48\u529e？<\/p>\n \u53ef\u4ee5\u53c2\u8003：<\/p>\n H-EPM<\/p>\n \u6784\u5efa：<\/p>\n Episodic Memory
\n+
\nProcedural Memory<\/p>\n \u8ba9\u6a21\u578b\u5b66\u4e60：<\/p>\n \u8fc7\u53bb\u505a\u8fc7\u4ec0\u4e48
\n+
\n\u8fc7\u53bb\u4e3a\u4ec0\u4e48\u8fd9\u4e48\u505a
\n+
\n\u54ea\u4e9b\u884c\u4e3a\u53ef\u4ee5\u6cdb\u5316<\/p>\n
\n10.4 \u7b2c\u56db\u5c42：Policy Learning<\/h2>\n\u5982\u679c\u8fdb\u4e00\u6b65\u4f7f\u7528 RL：<\/p>\n \u4e00\u4e2a\u5b8c\u6574Trajectory\u6700\u7ec8\u6210\u529f，\u4f46\u5230\u5e95\u54ea\u4e00\u6b65\u8d21\u732e\u6700\u5927？<\/p>\n \u53c2\u8003：<\/p>\n GTPO<\/p>\n \u628a：<\/p>\n Trajectory Reward<\/p>\n \u8fdb\u4e00\u6b65\u53d8\u6210：<\/p>\n Turn-level Reward<\/p>\n \u8fd9\u6837\u5c31\u53ef\u4ee5\u7814\u7a76：<\/p>\n Turn 1
\nTurn 2
\nTurn 3
\n…<\/p>\n \u6bcf\u4e00\u6b65\u7684\u7b56\u7565\u5b66\u4e60\u3002<\/p>\n
\n10.5 \u7b2c\u4e94\u5c42：Training Data<\/h2>\n\u7136\u540e\u9700\u8981\u5927\u91cf\u8bad\u7ec3\u6570\u636e\u3002<\/p>\n \u53c2\u8003：<\/p>\n FunReason-MT<\/p>\n \u6784\u9020：<\/p>\n Goal
\n \u2193
\nUser Query
\n \u2193
\nAgent Reasoning
\n \u2193
\nTool
\n \u2193
\nObservation
\n \u2193
\nNext User Query
\n \u2193
\nAgent Reasoning
\n \u2193
\nTool
\n \u2193
\nFinal<\/p>\n \u771f\u6b63\u5f62\u6210：<\/p>\n \u591a\u8f6e\u903b\u8f91\u4f9d\u8d56\u3002<\/p>\n
\n10.6 \u7b2c\u516d\u5c42：Evaluation<\/h2>\n\u6700\u540e\u4e0d\u80fd\u53ea\u770b：<\/p>\n Final Answer Accuracy<\/p>\n \u800c\u5e94\u8be5\u50cf EvolIF \u4e00\u6837\u8003\u8651：<\/p>\n Task Completion
\n+
\nInstruction Following
\n+
\nContext Consistency
\n+
\nRecovery
\n+
\nUser Satisfaction
\n+
\nInteraction Length
\n+
\nRobustness<\/p>\n \u4e5f\u5c31\u662f：<\/p>\n \u4ece\u201c\u8bc4\u4ef7\u7b54\u6848\u201d\u8f6c\u5411\u201c\u8bc4\u4ef7\u6574\u4e2a\u4ea4\u4e92\u8fc7\u7a0b\u201d\u3002<\/p>\n
\n11. \u5982\u679c\u628a\u8fd96\u7bc7\u8bba\u6587\u8fdb\u4e00\u6b65\u62bd\u8c61，\u53ef\u4ee5\u5f97\u5230\u4e00\u4e2a\u7814\u7a76\u95ed\u73af<\/h2>\n\u6700\u7ec8\u53ef\u4ee5\u5f97\u5230\u8fd9\u6837\u4e00\u4e2a\u5b8c\u6574\u67b6\u6784：<\/p>\n \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510
\n \u2502 User Simulator \u2502
\n \u2502 Persona + Goal \u2502
\n \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518
\n \u2502
\n \u2193
\n Multi-turn Dialogue
\n \u2502
\n \u2193
\n \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510
\n \u2502 State Manager \u2502
\n \u2502 Instruction Memory \u2502
\n \u2502 Episodic Memory \u2502
\n \u2502 Task State \u2502
\n \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518
\n \u2502
\n \u2193
\n Agent Reasoning
\n \u2502
\n \u2193
\n Tool \/ Environment
\n \u2502
\n \u2193
\n Observation
\n \u2502
\n \u2193
\n State Evolution
\n \u2502
\n \u2193
\n Next User Turn
\n \u2502
\n \u2193
\n Reward
\n \u2502
\n \u2193
\n Turn-level RL
\n \u2502
\n \u2193
\n Better Policy
\n \u2502
\n \u2193
\n Experience Memory
\n \u2502
\n \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2192 \u4e0b\u4e00\u6b21\u4efb\u52a1<\/p>\n \u7136\u540e\u5916\u90e8\u518d\u5957\u4e00\u4e2a：<\/p>\n \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510
\n \u2502 EvolIF-style \u2502
\n \u2502 Multi-turn Eval \u2502
\n \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518
\n \u2193
\n Long-horizon Interaction Score<\/p>\n \u8fd9\u5df2\u7ecf\u975e\u5e38\u63a5\u8fd1\u4e00\u4e2a\u5b8c\u6574\u7684：<\/p>\n Self-evolving Multi-turn Agent<\/p>\n
\n12. \u6700\u503c\u5f97\u5173\u6ce8\u7684\u4e09\u4e2a\u7814\u7a76\u7a7a\u767d<\/h2>\n\u9605\u8bfb\u8fd9\u516d\u7bc7\u8bba\u6587\u4e4b\u540e，\u6211\u8ba4\u4e3a\u8fd8\u6709\u4e09\u4e2a\u660e\u663e\u7684\u7814\u7a76\u7a7a\u95f4\u3002<\/p>\n 12.1 Memory\u548cUser State\u8fd8\u6ca1\u6709\u771f\u6b63\u7edf\u4e00<\/h3>\n\u73b0\u5728：<\/p>\n Rhea
\n\u2192 Conversation Memory<\/p>\n H-EPM
\n\u2192 Tool Experience Memory<\/p>\n Persona RL
\n\u2192 Persona Consistency<\/p>\n \u4f46\u771f\u6b63\u7684\u4eba\u7c7b\u7528\u6237\u72b6\u6001\u5e94\u8be5\u662f：<\/p>\n User Persona
\n+
\nUser Goal
\n+
\nUser Preference
\n+
\nEmotion
\n+
\nKnowledge
\n+
\nDialogue History
\n+
\nTask Progress<\/p>\n \u5982\u4f55\u628a\u8fd9\u4e9b\u7edf\u4e00\u6210：<\/p>\n Structured User State<\/p>\n \u4ecd\u7136\u503c\u5f97\u7814\u7a76\u3002<\/p>\n
\n12.2 \u201c\u7528\u6237\u6a21\u62df\u5668\u201d\u4e0e\u201c\u8bc4\u4f30\u5668\u201d\u53ef\u4ee5\u8fdb\u4e00\u6b65\u7edf\u4e00<\/h2>\n\u73b0\u5728：<\/p>\n Persona Simulator<\/p>\n \u8d1f\u8d23\u751f\u6210User\u3002<\/p>\n \u800c：<\/p>\n EvolIF<\/p>\n \u8d1f\u8d23\u8bc4\u4f30Interaction\u3002<\/p>\n \u4e00\u4e2a\u66f4\u8fdb\u4e00\u6b65\u7684\u65b9\u5411\u662f：<\/p>\n User Simulator
\n \u2193
\nInteraction
\n \u2193
\nUser Satisfaction
\n \u2193
\nState Update
\n \u2193
\nNext Turn<\/p>\n \u8ba9 User Simulator \u672c\u8eab\u6210\u4e3a：<\/p>\n Dynamic Evaluator<\/p>\n \u8fd9\u6837\u5c31\u4e0d\u9700\u8981\u4eba\u4e3a\u89c4\u5b9a\u56fa\u5b9a10\u8f6e\u300120\u8f6e\u3002<\/p>\n
\n12.3 \u591a\u8f6e\u5bf9\u8bdd\u7684Reward\u4ecd\u7136\u975e\u5e38\u503c\u5f97\u7814\u7a76<\/h2>\nGTPO\u89e3\u51b3\u4e86：<\/p>\n Trajectory Reward
\n\u2193
\nTurn Reward<\/p>\n \u4f46\u4e00\u4e2a\u771f\u6b63\u590d\u6742\u7684\u591a\u8f6e\u4efb\u52a1\u53ef\u80fd\u8fd8\u9700\u8981：<\/p>\n Turn Reward
\n+
\nTask Progress
\n+
\nUser Satisfaction
\n+
\nInstruction Following
\n+
\nPersona Consistency
\n+
\nRecovery
\n+
\nTool Efficiency<\/p>\n \u56e0\u6b64\u672a\u6765\u5f88\u53ef\u80fd\u4ece：<\/p>\n Turn-level Reward<\/p>\n \u7ee7\u7eed\u53d1\u5c55\u5230：<\/p>\n Multi-objective Interaction Reward<\/p>\n \u4f8b\u5982：<\/p>\n [ R_t = \\\\lambda_1 R_{\\\\text{task}} + \\\\lambda_2 R_{\\\\text{user}} + \\\\lambda_3 R_{\\\\text{instruction}} + \\\\lambda_4 R_{\\\\text{consistency}} + \\\\lambda_5 R_{\\\\text{efficiency}} ]<\/p>\n \u6700\u7ec8\u8bad\u7ec3\u7684\u5c31\u4e0d\u518d\u662f\u7b80\u5355：<\/p>\n \u201c\u56de\u7b54\u6b63\u786e\u7684\u6a21\u578b\u201d\u3002<\/p>\n \u800c\u662f：<\/p>\n \u80fd\u591f\u6301\u7eed\u5b8c\u6210\u4efb\u52a1\u3001\u7406\u89e3\u7528\u6237\u72b6\u6001\u3001\u4fdd\u6301\u4e00\u81f4\u6027\u3001\u53ca\u65f6\u6062\u590d\u9519\u8bef，\u5e76\u4e14\u8ba9\u7528\u6237\u613f\u610f\u7ee7\u7eed\u4ea4\u4e92\u7684\u6a21\u578b\u3002<\/p>\n
\n13. \u6700\u7ec8\u603b\u7ed3：\u8fd96\u7bc7\u8bba\u6587\u5206\u522b\u5728\u8865\u54ea\u4e00\u5757？<\/h2>\n\u53ef\u4ee5\u7528\u4e00\u53e5\u8bdd\u6982\u62ec：<\/p>\n FunReason-MT<\/h4>\n\u89e3\u51b3\u201c\u6ca1\u6709\u597d\u7684\u591a\u8f6e\u8bad\u7ec3\u6570\u636e\u201d\u3002<\/p>\n Persona RL<\/h4>\n\u89e3\u51b3\u201cUser Simulator\u4e0d\u80fd\u4e00\u76f4\u4fdd\u6301\u81ea\u5df1\u662f\u8c01\u201d\u3002<\/p>\n Rhea<\/h4>\n\u89e3\u51b3\u201cConversation\u8d8a\u957f，\u91cd\u8981\u4fe1\u606f\u8d8a\u5bb9\u6613\u88ab\u6df9\u6ca1\u201d\u3002<\/p>\n H-EPM<\/h4>\n\u89e3\u51b3\u201c\u8fc7\u53bb\u6210\u529f\u7ecf\u9a8c\u65e0\u6cd5\u6709\u6548\u8fc1\u79fb\u5230\u65b0\u4efb\u52a1\u201d\u3002<\/p>\n GTPO<\/h4>\n\u89e3\u51b3\u201c\u591a\u8f6eRL\u53ea\u77e5\u9053\u6700\u7ec8\u6210\u529f，\u5374\u4e0d\u77e5\u9053\u54ea\u4e00\u6b65\u505a\u5f97\u597d\u201d\u3002<\/p>\n EvolIF<\/h4>\n\u89e3\u51b3\u201c\u4f20\u7edfBenchmark\u65e0\u6cd5\u771f\u6b63\u6a21\u62df\u957f\u671f\u4eba\u673a\u4ea4\u4e92\u201d\u3002<\/p>\n \u56e0\u6b64，\u8fd9\u516d\u7bc7\u8bba\u6587\u5171\u540c\u63ed\u793a\u4e86\u4e00\u4e2a\u975e\u5e38\u91cd\u8981\u7684\u8d8b\u52bf：<\/p>\n \u672a\u6765\u7684\u591a\u8f6e\u5bf9\u8bdd\u7814\u7a76，\u6838\u5fc3\u7ade\u4e89\u529b\u53ef\u80fd\u4e0d\u518d\u662f\u5355\u7eaf\u7684Context Length，\u800c\u662f\u6a21\u578b\u80fd\u5426\u5f62\u6210\u201c\u72b6\u6001\u2014\u8bb0\u5fc6\u2014\u884c\u52a8\u2014\u53cd\u9988\u2014\u7ecf\u9a8c\u201d\u7684\u95ed\u73af\u3002<\/p>\n \u4ece：<\/p>\n Context \u2192 Response<\/p>\n \u9010\u6e10\u6f14\u5316\u4e3a：<\/p>\n State
\n \u2193
\nMemory
\n \u2193
\nReasoning
\n \u2193
\nAction
\n \u2193
\nEnvironment
\n \u2193
\nObservation
\n \u2193
\nUser Feedback
\n \u2193
\nReward
\n \u2193
\nLearning
\n \u2193
\nUpdated State<\/p>\n \u8fd9\u4e5f\u662f\u4ece Chatbot \u2192 Agent \u2192 Continual Interactive Agent \u7684\u5173\u952e\u4e00\u6b65\u3002<\/p>\n
\n14. \u516d\u7bc7\u8bba\u6587\u94fe\u63a5\u6c47\u603b<\/h3>\n\nH-EPM \u2014 Experience-Evolving Multi-Turn Tool-Use Agent with Hybrid Episodic-Procedural Memory ArXiv 2512.07287<\/p>\n<\/li>\n \nRhea \u2014 Role-aware Heuristic Episodic Attention for Conversational LLMs ArXiv 2512.06869<\/p>\n<\/li>\n \nGTPO \u2014 Empowering Multi-Turn Tool-Integrated Agentic Reasoning with Group Turn Policy Optimization ArXiv 2511.14846<\/p>\n<\/li>\n \nOne Battle After Another \/ EvolIF ArXiv 2511.03508<\/p>\n<\/li>\n \nConsistently Simulating Human Personas with Multi-Turn Reinforcement Learning ArXiv 2511.00222<\/p>\n<\/li>\n \nFunReason-MT \u2014 Advanced Data Synthesis Solution for Real-world Multi-Turn Tool-use ArXiv 2510.24645<\/p>\n<\/li>\n
\n15. \u4e00\u5f20\u56fe\u8bb0\u4f4f\u6574\u4e2a\u7814\u7a76\u65b9\u5411<\/h3>\n \u591a\u8f6e\u5bf9\u8bdd \/ Agent
\n \u2502
\n \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510
\n \u2193 \u2193 \u2193
\n \u6570\u636e\u6784\u9020 Memory RL
\n \u2502 \u2502 \u2502
\n FunReason-MT Rhea \/ H-EPM GTPO
\n \u2502 \u2502 \u2502
\n \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518
\n \u2193
\n Long-horizon Agent
\n \u2502
\n \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2534\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510
\n \u2193 \u2193
\n User Simulator Evaluation
\n \u2502 \u2502
\n Persona RL EvolIF
\n \u2502 \u2502
\n \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518
\n \u2193
\n Self-Evolving Agent
\n \u2193
\n State \u2192 Memory \u2192 Action \u2192 Feedback
\n \u2191 \u2193
\n \u2514\u2500\u2500\u2500\u2500\u2500\u2500 Learning \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518<\/p>\n \u5982\u679c\u4ece\u79d1\u7814\u9009\u9898\u89d2\u5ea6\u770b，\u6700\u503c\u5f97\u7ee7\u7eed\u6df1\u5165\u7684\u5e76\u4e0d\u662f\u5355\u72ec\u590d\u73b0\u5176\u4e2d\u67d0\u4e00\u7bc7，\u800c\u662f\u628a\u8fd9\u516d\u7bc7\u5de5\u4f5c\u8fde\u63a5\u8d77\u6765：\u52a8\u6001User Simulator + Structured User State + Episodic\/Procedural Memory + Turn-level Reward + Evolving Evaluation\u3002\u8fd9\u6761\u8def\u7ebf\u5df2\u7ecf\u975e\u5e38\u63a5\u8fd1\u201c\u9762\u5411\u957f\u671f\u4ea4\u4e92\u7684\u591a\u8f6e\u5bf9\u8bddAgent\u201d\u8fd9\u4e00\u5b8c\u6574\u7814\u7a76\u6846\u67b6\u3002<\/p>\n","protected":false},"excerpt":{"rendered":" \u8bba\u6587\u65b9\u5411：Multi-Turn Dialogue \/ Agent \/ Memory \/ RL \/ Human Simulator \/ Evaluation \u9605\u8bfb\u8bba\u6587：6\u7bc7 \u6838\u5fc3\u5206\u6790\u6846\u67b6：\u89e3\u51b3\u7684\u95ee\u9898 \u2192 \u89e3\u51b3\u65b9\u6cd5 \u2192 \u4e0e\u5176\u5b83\u65b9\u6cd5\u7684\u533a\u522b \u2192 \u5b9e\u9a8c\u9a8c\u8bc1\u8def\u7ebf 0. \u4e3a\u4ec0\u4e48\u8981\u628a\u8fd9 6 \u7bc7\u8bba\u6587\u653e\u5728\u4e00\u8d77\u770b？
\n\u8fc7\u53bb\u7684\u5927\u8bed\u8a00\u6a21\u578b\u591a\u8f6e\u5bf9\u8bdd\u7814\u7a76，\u4e00\u4e2a\u975e\u5e38\u5178\u578b\u7684\u601d\u8def\u662f：
\nUser\u2193
\nConversation Hi<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[50,86,427,3546],"topic":[],"class_list":["post-95993","post","type-post","status-publish","format-standard","hentry","category-server","tag-50","tag-86","tag-427","tag-3546"],"yoast_head":"\n \u3010\u591a\u8f6e\u5bf9\u8bdd\u8bba\u6587\u5bfc\u8bfb\uff08\u516b\uff09\u3011\u591a\u8f6e\u5bf9\u8bdd\u6b63\u5728\u4ece\u201c\u4e0a\u4e0b\u6587\u8bb0\u5fc6\u201d\u8d70\u5411\u201c\u6301\u7eed\u4ea4\u4e92\u667a\u80fd\u201d - \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\/95993.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u3010\u591a\u8f6e\u5bf9\u8bdd\u8bba\u6587\u5bfc\u8bfb\uff08\u516b\uff09\u3011\u591a\u8f6e\u5bf9\u8bdd\u6b63\u5728\u4ece\u201c\u4e0a\u4e0b\u6587\u8bb0\u5fc6\u201d\u8d70\u5411\u201c\u6301\u7eed\u4ea4\u4e92\u667a\u80fd\u201d - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u8bba\u6587\u65b9\u5411：Multi-Turn Dialogue \/ Agent \/ Memory \/ RL \/ Human Simulator \/ Evaluation \u9605\u8bfb\u8bba\u6587：6\u7bc7 \u6838\u5fc3\u5206\u6790\u6846\u67b6：\u89e3\u51b3\u7684\u95ee\u9898 \u2192 \u89e3\u51b3\u65b9\u6cd5 \u2192 \u4e0e\u5176\u5b83\u65b9\u6cd5\u7684\u533a\u522b \u2192 \u5b9e\u9a8c\u9a8c\u8bc1\u8def\u7ebf 0. \u4e3a\u4ec0\u4e48\u8981\u628a\u8fd9 6 \u7bc7\u8bba\u6587\u653e\u5728\u4e00\u8d77\u770b？ \u8fc7\u53bb\u7684\u5927\u8bed\u8a00\u6a21\u578b\u591a\u8f6e\u5bf9\u8bdd\u7814\u7a76，\u4e00\u4e2a\u975e\u5e38\u5178\u578b\u7684\u601d\u8def\u662f： User\u2193 Conversation Hi\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/95993.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-08-27T01:00:28+00:00\" \/>\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=\"11 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/95993.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/95993.html\",\"name\":\"\u3010\u591a\u8f6e\u5bf9\u8bdd\u8bba\u6587\u5bfc\u8bfb\uff08\u516b\uff09\u3011\u591a\u8f6e\u5bf9\u8bdd\u6b63\u5728\u4ece\u201c\u4e0a\u4e0b\u6587\u8bb0\u5fc6\u201d\u8d70\u5411\u201c\u6301\u7eed\u4ea4\u4e92\u667a\u80fd\u201d - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-08-27T01:00:28+00:00\",\"dateModified\":\"2026-08-27T01:00:28+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/95993.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/95993.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/95993.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u3010\u591a\u8f6e\u5bf9\u8bdd\u8bba\u6587\u5bfc\u8bfb\uff08\u516b\uff09\u3011\u591a\u8f6e\u5bf9\u8bdd\u6b63\u5728\u4ece\u201c\u4e0a\u4e0b\u6587\u8bb0\u5fc6\u201d\u8d70\u5411\u201c\u6301\u7eed\u4ea4\u4e92\u667a\u80fd\u201d\"}]},{\"@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\u591a\u8f6e\u5bf9\u8bdd\u8bba\u6587\u5bfc\u8bfb\uff08\u516b\uff09\u3011\u591a\u8f6e\u5bf9\u8bdd\u6b63\u5728\u4ece\u201c\u4e0a\u4e0b\u6587\u8bb0\u5fc6\u201d\u8d70\u5411\u201c\u6301\u7eed\u4ea4\u4e92\u667a\u80fd\u201d - \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\/95993.html","og_locale":"zh_CN","og_type":"article","og_title":"\u3010\u591a\u8f6e\u5bf9\u8bdd\u8bba\u6587\u5bfc\u8bfb\uff08\u516b\uff09\u3011\u591a\u8f6e\u5bf9\u8bdd\u6b63\u5728\u4ece\u201c\u4e0a\u4e0b\u6587\u8bb0\u5fc6\u201d\u8d70\u5411\u201c\u6301\u7eed\u4ea4\u4e92\u667a\u80fd\u201d - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u8bba\u6587\u65b9\u5411：Multi-Turn Dialogue \/ Agent \/ Memory \/ RL \/ Human Simulator \/ Evaluation \u9605\u8bfb\u8bba\u6587：6\u7bc7 \u6838\u5fc3\u5206\u6790\u6846\u67b6：\u89e3\u51b3\u7684\u95ee\u9898 \u2192 \u89e3\u51b3\u65b9\u6cd5 \u2192 \u4e0e\u5176\u5b83\u65b9\u6cd5\u7684\u533a\u522b \u2192 \u5b9e\u9a8c\u9a8c\u8bc1\u8def\u7ebf 0. \u4e3a\u4ec0\u4e48\u8981\u628a\u8fd9 6 \u7bc7\u8bba\u6587\u653e\u5728\u4e00\u8d77\u770b？ \u8fc7\u53bb\u7684\u5927\u8bed\u8a00\u6a21\u578b\u591a\u8f6e\u5bf9\u8bdd\u7814\u7a76，\u4e00\u4e2a\u975e\u5e38\u5178\u578b\u7684\u601d\u8def\u662f： User\u2193 Conversation Hi","og_url":"https:\/\/www.wsisp.com\/helps\/95993.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-08-27T01:00:28+00:00","author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"11 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/95993.html","url":"https:\/\/www.wsisp.com\/helps\/95993.html","name":"\u3010\u591a\u8f6e\u5bf9\u8bdd\u8bba\u6587\u5bfc\u8bfb\uff08\u516b\uff09\u3011\u591a\u8f6e\u5bf9\u8bdd\u6b63\u5728\u4ece\u201c\u4e0a\u4e0b\u6587\u8bb0\u5fc6\u201d\u8d70\u5411\u201c\u6301\u7eed\u4ea4\u4e92\u667a\u80fd\u201d - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-08-27T01:00:28+00:00","dateModified":"2026-08-27T01:00:28+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/95993.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/95993.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/95993.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"\u3010\u591a\u8f6e\u5bf9\u8bdd\u8bba\u6587\u5bfc\u8bfb\uff08\u516b\uff09\u3011\u591a\u8f6e\u5bf9\u8bdd\u6b63\u5728\u4ece\u201c\u4e0a\u4e0b\u6587\u8bb0\u5fc6\u201d\u8d70\u5411\u201c\u6301\u7eed\u4ea4\u4e92\u667a\u80fd\u201d"}]},{"@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\/95993","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=95993"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/95993\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=95993"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=95993"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=95993"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=95993"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}} |