{"id":120419,"date":"2026-10-31T17:19:49","date_gmt":"2026-10-31T09:19:49","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/120419.html"},"modified":"2026-10-31T17:19:49","modified_gmt":"2026-10-31T09:19:49","slug":"%e6%b7%b1%e5%85%a5%e6%b5%85%e5%87%bagroundingpi%e6%8a%80%e6%9c%af%e8%a7%a3%e6%9e%90%ef%bc%9a%e7%b2%be%e7%a1%ae%e7%a9%ba%e9%97%b4%e6%84%9f%e7%9f%a5%e5%a6%82%e4%bd%95%e8%bf%81%e7%a7%bb%e5%88%b0%e6%9c%ba","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/120419.html","title":{"rendered":"\u6df1\u5165\u6d45\u51faGroundingPI\u6280\u672f\u89e3\u6790\uff1a\u7cbe\u786e\u7a7a\u95f4\u611f\u77e5\u5982\u4f55\u8fc1\u79fb\u5230\u673a\u5668\u4eba\u52a8\u4f5c"},"content":{"rendered":"

\u5199\u5728\u524d\u9762<\/h2>\n

\"\u56fe\u7247\"<\/p>\n

\u6b22\u8fce\u5927\u5bb6\u5173\u6ce8Rocky\u7684\u77e5\u4e4e：Rocky Ding \u300a\u4e09\u5e74\u9762\u8bd5\u4e94\u5e74\u6a21\u62df\u300bAIGC\/LLM\/AI Agent\u7b97\u6cd5\u5de5\u7a0b\u5e08\/\u5f00\u53d1\u5de5\u7a0b\u5e08\u6c42\u804c\u9762\u8bd5\u79d8\u7c4d\u72ec\u5bb6\u8d44\u6e90：\u3010\u4e09\u5e74\u9762\u8bd5\u4e94\u5e74\u6a21\u62df\u3011WeThinkIn\/AIGC-Interview-Book，\u6b22\u8fce\u5927\u5bb6Star～<\/p>\n

Rocky\u6700\u65b0\u64b0\u5199\u768410\u4e07\u5b57AI Agent（AI\u667a\u80fd\u4f53）\u6df1\u5165\u6d45\u51fa\u5168\u7ef4\u5ea6\u89e3\u6790\u6587\u7ae0： \u6df1\u5165\u6d45\u51fa\u5b8c\u6574\u89e3\u6790AI Agent（AI\u667a\u80fd\u4f53）\u7684\u6838\u5fc3\u57fa\u7840\u77e5\u8bc6<\/p>\n

AIGC\/LLM\/AI Agent\u7b97\u6cd5\u5c97\/\u5f00\u53d1\u5c97\u6c42\u804c\u9762\u8bd5\u5185\u63a8\u5b66\u4e60\u793e\u7fa4（\u6db5\u76d6AIGC\u3001LLM\u5927\u6a21\u578b\u3001AI Agent\u3001\u4f20\u7edf\u6df1\u5ea6\u5b66\u4e60\u3001\u81ea\u52a8\u9a7e\u9a76\u3001\u673a\u5668\u5b66\u4e60\u3001\u8ba1\u7b97\u673a\u89c6\u89c9\u3001\u81ea\u7136\u8bed\u8a00\u5904\u7406\u3001\u5f3a\u5316\u5b66\u4e60\u3001\u5927\u6570\u636e\u6316\u6398\u3001\u5177\u8eab\u667a\u80fd\u3001\u5143\u5b87\u5b99\u3001AGI\u7b49AI\u884c\u4e1a\u6700\u65b0\u9762\u8bd5\u5e72\u8d27\u7ecf\u9a8c\u4e0e\u6838\u5fc3\u77e5\u8bc6）\u6b22\u8fce\u5927\u5bb6\u52a0\u5165：https:\/\/t.zsxq.com\/33pJ0<\/p>\n

\n

\u5927\u5bb6\u597d，\u6211\u662fRocky\u3002<\/p>\n

1. \u673a\u5668\u4eba\u9700\u8981\u7684\u89c6\u89c9\u80fd\u529b，\u53ef\u80fd\u85cf\u5728\u6700\u4e0d\u8d77\u773c\u7684\u7ec6\u8282\u91cc<\/h3>\n

\u6574\u7406\u684c\u9762\u7684\u673a\u5668\u4eba，\u5fc5\u987b\u533a\u5206\u76f8\u90bb\u7684\u4e24\u4e2a\u676f\u5b50，\u627e\u51fa\u76ee\u6807\u676f\u628a，\u5e76\u5224\u65ad\u54ea\u4e2a\u5c40\u90e8\u4f4d\u7f6e\u9002\u5408\u4ea4\u4e92\u3002\u6a21\u578b\u5373\u4f7f\u80fd\u6d41\u7545\u89e3\u91ca\u201c\u676f\u5b50\u53ef\u4ee5\u88c5\u6c34\u201d，\u4e5f\u4e0d\u80fd\u7531\u6b64\u4fdd\u8bc1\u5b83\u80fd\u628a\u624b\u4f38\u5230\u6b63\u786e\u4f4d\u7f6e\u3002\u611f\u77e5\u548c\u52a8\u4f5c\u4e4b\u95f4，\u5b58\u5728\u4e00\u4e2a\u975e\u5e38\u5177\u4f53\u7684\u63a5\u53e3\u95ee\u9898：\u8bed\u8a00\u76ee\u6807\u80fd\u5426\u7a33\u5b9a\u5bf9\u5e94\u5230\u56fe\u50cf\u4e2d\u7684\u5bf9\u8c61\u3001\u4f4d\u7f6e\u548c\u5c40\u90e8\u7ed3\u6784\u3002<\/p>\n

GroundingPI \u628a\u8fd9\u4e2a\u95ee\u9898\u4f5c\u4e3a\u9884\u8bad\u7ec3\u7684\u4e2d\u5fc3\u3002\u5b83\u7528\u70b9\u548c\u6846\u8fd9\u4e9b\u89c6\u89c9\u57fa\u5143\u7edf\u4e00\u76ee\u6807\u68c0\u6d4b\u3001\u6307\u4ee3\u3001OCR\u3001\u754c\u9762\u5b9a\u4f4d\u4e0e\u89c6\u89c9\u63d0\u793a，\u518d\u8003\u5bdf\u8bad\u7ec3\u51fa\u7684\u611f\u77e5\u8868\u793a\u80fd\u5426\u5e2e\u52a9\u673a\u5668\u4eba\u548c\u81ea\u52a8\u9a7e\u9a76\u7684\u52a8\u4f5c\u5b66\u4e60\u3002<\/p>\n

Rocky\u8ba4\u4e3a，\u8fd9\u9879\u5de5\u4f5c\u7684\u4ef7\u503c，\u5728\u4e8e\u628a\u201c\u53ef\u8fc1\u79fb\u7684\u7cbe\u786e\u611f\u77e5\u201d\u5f53\u4f5c\u52a8\u4f5c\u6a21\u578b\u7684\u57fa\u7840\u8d44\u4ea7\u6765\u7814\u7a76\u3002 \u673a\u5668\u4eba\u6f14\u793a\u6570\u636e\u6602\u8d35，\u5982\u679c\u6bcf\u6b21\u8bad\u7ec3\u7b56\u7565\u90fd\u9700\u8981\u540c\u65f6\u5b66\u4f1a\u770b\u6e05\u5bf9\u8c61\u548c\u63a7\u5236\u5173\u8282，\u52a8\u4f5c\u6570\u636e\u5c31\u627f\u62c5\u4e86\u8fc7\u591a\u804c\u8d23\u3002\u4e00\u4e2a\u80fd\u63d0\u4f9b\u7a33\u5b9a\u7a7a\u95f4\u4fe1\u606f\u7684\u57fa\u7840\u6a21\u578b，\u6709\u673a\u4f1a\u8ba9\u540e\u7eed\u6f14\u793a\u66f4\u96c6\u4e2d\u5730\u6559\u4f1a\u7cfb\u7edf\u5982\u4f55\u6267\u884c\u3002<\/p>\n

\u4f46\u8fd9\u4e2a\u63a8\u7406\u6709\u8fb9\u754c：\u672c\u6587\u5206\u6790\u7684\u662f\u516c\u5f00\u62a5\u544a\u4e2d\u7684\u611f\u77e5\u4e0e\u8fc1\u79fb\u5b9e\u9a8c，\u6ca1\u6709\u72ec\u7acb\u8bad\u7ec3\u6216\u90e8\u7f72\u673a\u5668\u4eba\u3002\u8bba\u6587\u7ed3\u679c\u4e0d\u80fd\u76f4\u63a5\u7b49\u540c\u4e8e\u771f\u5b9e\u7269\u7406\u73af\u5883\u4e2d\u7684\u95ed\u73af\u53ef\u9760\u6027，\u611f\u77e5\u699c\u5355\u4e5f\u4e0d\u80fd\u66ff\u4ee3\u63a7\u5236\u7cfb\u7edf\u9a8c\u6536\u3002<\/p>\n

\"\u56fe1\uff1a\u7edf\u4e00\u7a7a\u95f4\u4efb\u52a1\u4e0e\u52a8\u4f5c\u8fc1\u79fb\u6982\u89c8\"<\/p>\n

\u56fe1｜\u7edf\u4e00\u7a7a\u95f4\u4efb\u52a1\u4e0e\u52a8\u4f5c\u8fc1\u79fb\u6982\u89c8\u3002\u6765\u81eaGroundingPI\u6280\u672f\u62a5\u544a\u3002<\/p>\n

2. \u7528\u4e00\u4e2a\u8bcd\u8868\u8868\u8fbe\u5bf9\u8c61\u3001\u4f4d\u7f6e\u4e0e\u7f3a\u5931<\/h3>\n

GroundingPI \u7531 MoonViT-V2 \u89c6\u89c9\u7f16\u7801\u5668\u3001\u53ef\u5b66\u4e60\u6295\u5f71\u5c42\u4e0e Qwen3-4B \u8bed\u8a00\u89e3\u7801\u5668\u7ec4\u6210\u3002\u6295\u5f71\u5668\u5408\u5e76\u76f8\u90bb <\/p>\n

2<\/p>\n

\u00d7<\/p>\n

2<\/p>\n

2\\\\times2<\/p>\n

<\/span><\/span>2<\/span><\/span>\u00d7<\/span><\/span><\/span><\/span>2<\/span><\/span><\/span><\/span><\/span> \u89c6\u89c9\u7279\u5f81，\u518d\u7528\u4e24\u5c42MLP\u6620\u5c04\u5230\u8bed\u8a00\u7a7a\u95f4\u3002\u7ed9\u5b9a\u56fe\u50cf <\/p>\n

I<\/p>\n

I<\/p>\n

<\/span><\/span>I<\/span><\/span><\/span><\/span><\/span> \u4e0e\u67e5\u8be2 <\/p>\n

P<\/p>\n

P<\/p>\n

<\/span><\/span>P<\/span><\/span><\/span><\/span><\/span>：<\/p>\n

<\/p>\n

V<\/p>\n

=<\/p>\n

C<\/p>\n

\u03d5<\/p>\n

(<\/p>\n

E<\/p>\n

\u03c8<\/p>\n

(<\/p>\n

I<\/p>\n

)<\/p>\n

)<\/p>\n

,<\/p>\n

p<\/p>\n

\u03b8<\/p>\n

(<\/p>\n

Y<\/p>\n

\u2223<\/p>\n

I<\/p>\n

,<\/p>\n

P<\/p>\n

)<\/p>\n

=<\/p>\n

\u220f<\/p>\n

i<\/p>\n

=<\/p>\n

1<\/p>\n

\u2223<\/p>\n

Y<\/p>\n

\u2223<\/p>\n

p<\/p>\n

\u03b8<\/p>\n

(<\/p>\n

y<\/p>\n

i<\/p>\n

\u2223<\/p>\n

V<\/p>\n

,<\/p>\n

P<\/p>\n

,<\/p>\n

y<\/p>\n

<<\/p>\n

i<\/p>\n

)<\/p>\n

.<\/p>\n

V=C_\\\\phi(E_\\\\psi(I)),\\\\qquad p_\\\\theta(Y\\\\mid I,P)=\\\\prod_{i=1}^{|Y|}p_\\\\theta(y_i\\\\mid V,P,y_{<i}). <\/p>\n

<\/span><\/span>V<\/span><\/span>=<\/span><\/span><\/span><\/span>C<\/span><\/span>\u03d5<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>E<\/span><\/span>\u03c8<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>I<\/span>))<\/span>,<\/span><\/span><\/span>p<\/span><\/span>\u03b8<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>Y<\/span><\/span>\u2223<\/span><\/span><\/span><\/span>I<\/span>,<\/span><\/span>P<\/span>)<\/span><\/span>=<\/span><\/span><\/span><\/span><\/span>i<\/span>=<\/span>1<\/span><\/span><\/span><\/span><\/span>\u220f<\/span><\/span><\/span><\/span>\u2223<\/span>Y<\/span>\u2223<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>p<\/span><\/span>\u03b8<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>y<\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u2223<\/span><\/span><\/span><\/span>V<\/span>,<\/span><\/span>P<\/span>,<\/span><\/span>y<\/span><\/span><<\/span>i<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>)<\/span>.<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

\u8bed\u4e49\u7c7b\u522b\u3001\u7ed3\u6784\u6807\u8bb0\u548c1,000\u4e2a\u5750\u6807token\u5171\u4eab\u8f93\u51fa\u5934\u3002\u6846\u7528xyxy\u56db\u4e2a\u5750\u6807，\u70b9\u7528\u4e24\u4e2a\u5750\u6807；\u660e\u786e\u67e5\u8be2\u5374\u4e0d\u5b58\u5728\u7684\u7c7b\u522b\u8f93\u51faNone\u3002\u540c\u7c7b\u591a\u4e2a\u5b9e\u4f8b\u5171\u4eab\u6807\u7b7e\u5305\u88c5，\u51cf\u5c11\u91cd\u590d\u6587\u672c\u5f00\u9500\u3002<\/p>\n

\"\u56fe2\uff1a\u89c6\u89c9\u7f16\u7801\u4e0e\u539f\u5b50\u5750\u6807\u8bcd\u8868\"<\/p>\n

\u56fe2｜\u89c6\u89c9\u7f16\u7801\u4e0e\u539f\u5b50\u5750\u6807\u8bcd\u8868\u3002\u6765\u81eaGroundingPI\u6280\u672f\u62a5\u544a\u3002<\/p>\n

2.1 \u91cf\u5316\u5750\u6807\u5e26\u6765\u7684\u6536\u76ca\u4e0e\u8bef\u5dee<\/h4>\n

\u6bcf\u4e2a <0> \u5230 <999> \u90fd\u662f\u4e00\u4e2a\u539f\u5b50token，\u800c\u4e0d\u662f\u82e5\u5e72\u6570\u5b57\u5b57\u7b26\u3002\u5b83\u7f29\u77ed\u7a7a\u95f4\u5e8f\u5217，\u4e5f\u8ba9\u8bad\u7ec3\u76f4\u63a5\u76d1\u7763\u51e0\u4f55\u79bb\u6563\u503c\u3002\u539f\u5b50\u5750\u6807\u4ecd\u662f\u6709\u9650\u7cbe\u5ea6\u8868\u793a；\u6781\u5c0f\u76ee\u6807\u3001\u7f29\u653e\u53d6\u6574\u4e0e\u8fb9\u7f18\u4f4d\u7f6e\u90fd\u53ef\u80fd\u653e\u5927\u91cf\u5316\u8bef\u5dee\u3002<\/p>\n

\u8f93\u51fa\u534f\u8bae\u4e5f\u5c5e\u4e8e\u80fd\u529b\u7684\u4e00\u90e8\u5206\u3002\u591a\u6846\u6392\u5e8f\u3001\u7c7b\u522b\u5206\u9694\u7b26\u3001\u7ed3\u675f\u6807\u8bb0\u548c\u7f3a\u5931\u76ee\u6807\u7684\u8868\u793a\u5fc5\u987b\u4e00\u81f4；\u8c03\u7528\u65b9\u4e0d\u80fd\u53ea\u7528\u4e00\u4e2a\u5bbd\u677e\u6b63\u5219，\u6293\u5230\u56db\u4e2a\u6570\u5b57\u5c31\u9ed8\u8ba4\u8f93\u51fa\u6709\u6548\u3002<\/p>\n

\"\u88686\uff1a\u7a7a\u95f4\u534f\u8bae\u8bcd\u8868\"<\/p>\n

\u88686｜\u7a7a\u95f4\u534f\u8bae\u8bcd\u8868\u3002\u661f\u53f7\u3001\u7f3a\u5931\u9879\u4e0e\u534f\u8bae\u4f8b\u5916\u6cbf\u7528\u8bba\u6587，\u9700\u7ed3\u5408\u6b63\u6587\u89e3\u91ca\u3002<\/p>\n

\"\u88687\uff1a\u4efb\u52a1\u63d0\u793a\u4e0e\u51e0\u4f55\u8f93\u51fa\"<\/p>\n

\u88687｜\u4efb\u52a1\u63d0\u793a\u4e0e\u51e0\u4f55\u8f93\u51fa\u3002\u661f\u53f7\u3001\u7f3a\u5931\u9879\u4e0e\u534f\u8bae\u4f8b\u5916\u6cbf\u7528\u8bba\u6587，\u9700\u7ed3\u5408\u6b63\u6587\u89e3\u91ca\u3002<\/p>\n

\u5750\u6807\u534f\u8bae\u7ed9\u4e86\u6a21\u578b\u7edf\u4e00\u4efb\u52a1\u63a5\u53e3，\u4f46\u6ca1\u6709\u53d6\u6d88\u4efb\u52a1\u5dee\u5f02\u3002OCR\u6807\u7b7e\u662f\u6587\u5b57\u8f6c\u5199，\u68c0\u6d4b\u6807\u7b7e\u662f\u5bf9\u8c61\u7c7b\u522b，\u8f68\u8ff9\u70b9\u53c8\u9700\u8981\u4fdd\u6301\u65f6\u95f4\u6b21\u5e8f\u3002\u628a\u8fd9\u4e9b\u90fd\u5e8f\u5217\u5316，\u4e0d\u610f\u5473\u7740\u8bc4\u4ef7\u65f6\u53ef\u4ee5\u6df7\u7528\u5339\u914d\u65b9\u5f0f\u3002<\/p>\n

2.2 4B\u540d\u79f0\u4e0d\u7b49\u4e8e\u5b8c\u6574\u53c2\u6570\u8ba1\u6570<\/h4>\n

GroundingPI \u768436\u5c42\u8bed\u8a00\u4e3b\u5e72\u5747\u4e3a\u5168\u6ce8\u610f\u529b；\u52a8\u6001\u56fe\u50cf\u5904\u7406\u6700\u591a4096\u4e2apatch，\u5408\u5e76\u540e\u6700\u591a1024\u4e2a\u89c6\u89c9token\u3002\u4f4d\u7f6e\u5bb9\u91cf262,144\u4e0e\u8bad\u7ec3\u5b9e\u9645\u4f7f\u7528\u76848192\/32768\u5e8f\u5217\u957f\u5ea6\u4e0d\u540c\u3002<\/p>\n

\"\u88688\uff1a\u4e3b\u5e72\u7ed3\u6784\u4e0e\u89c6\u89c9\u9884\u7b97\"<\/p>\n

\u88688｜\u4e3b\u5e72\u7ed3\u6784\u4e0e\u89c6\u89c9\u9884\u7b97\u3002\u661f\u53f7\u3001\u7f3a\u5931\u9879\u4e0e\u534f\u8bae\u4f8b\u5916\u6cbf\u7528\u8bba\u6587，\u9700\u7ed3\u5408\u6b63\u6587\u89e3\u91ca\u3002<\/p>\n

\"\u88689\uff1a\u5b8c\u6574\u53c2\u6570\u8ba1\u6570\"<\/p>\n

\u88689｜\u5b8c\u6574\u53c2\u6570\u8ba1\u6570\u3002\u661f\u53f7\u3001\u7f3a\u5931\u9879\u4e0e\u534f\u8bae\u4f8b\u5916\u6cbf\u7528\u8bba\u6587，\u9700\u7ed3\u5408\u6b63\u6587\u89e3\u91ca\u3002<\/p>\n

\u7b97\u4e0a\u89c6\u89c9\u7f16\u7801\u5668\u3001\u6295\u5f71\u5668\u4ee5\u53ca\u4e0d\u5171\u4eab\u6743\u91cd\u7684\u8f93\u5165\u5d4c\u5165\u548c\u8f93\u51fa\u5934，\u5b8c\u6574\u6a21\u578b\u7ea64.844B\u53c2\u6570\u3002\u201c4B\u201d\u6307\u8bed\u8a00\u4e3b\u5e72\u5bb6\u65cf，\u4e0d\u80fd\u76f4\u63a5\u62ff\u5b83\u63a8\u7b97\u5b8c\u6574\u6a21\u578b\u7684\u663e\u5b58\u548c\u541e\u5410\u3002<\/p>\n

\u5bf9\u63a8\u7406\u6210\u672c，\u5173\u952e\u8fd8\u6709\u56fe\u50cftoken\u6570\u4e0e\u76ee\u6807\u6570\u3002\u5bc6\u96c6\u573a\u666f\u53ef\u80fd\u8981\u6c42\u8f93\u51fa\u4e0a\u767e\u4e2a\u6846，\u9010token\u751f\u6210\u4f1a\u7d2f\u79ef\u5ef6\u8fdf\u3002\u7a7a\u95f4\u8868\u793a\u66f4\u7d27\u51d1\u662f\u4e00\u79cd\u4f18\u5316，\u4f46\u662f\u5426\u6ee1\u8db3\u63a7\u5236\u5468\u671f，\u8981\u770b\u6574\u6761\u89c2\u6d4b\u5230\u52a8\u4f5c\u94fe\u7684\u6d4b\u91cf\u3002<\/p>\n

3. \u6570\u636e\u5f15\u64ce\u7684\u96be\u70b9：\u786e\u8ba4\u201c\u5168\u90e8\u201d\u4e0e\u786e\u8ba4\u201c\u6ca1\u6709\u201d<\/h3>\n

\u591a\u6559\u5e08\u6807\u6ce8\u4e0d\u662f\u7b80\u5355\u6295\u7968\u3002GroundingPI \u6309\u5b57\u6bb5\u878d\u5408\u5b9a\u4f4d\u3001\u5206\u5272\u3001\u6587\u5b57\u8bc6\u522b\u3001\u754c\u9762\u548c\u6587\u6863\u6807\u6ce8；\u8bed\u4e49\u4e0e\u51e0\u4f55\u5206\u522b\u9a8c\u8bc1\u3002\u4e00\u4e2a\u5bf9\u8c61\u7684\u7c7b\u522b\u786e\u8ba4\u4e86\u3001\u6846\u8fd8\u4e0d\u786e\u5b9a，\u53ea\u80fd\u4fdd\u7559\u5df2\u786e\u8ba4\u5b57\u6bb5，\u4e0d\u80fd\u628a\u6574\u4e2a\u8bb0\u5f55\u90fd\u89c6\u4e3a\u771f\u503c\u3002<\/p>\n

\"\u56fe3\uff1a\u591a\u6559\u5e08\u5b57\u6bb5\u6821\u9a8c\u4e0e\u6570\u636e\u8fed\u4ee3\"<\/p>\n

\u56fe3｜\u591a\u6559\u5e08\u5b57\u6bb5\u6821\u9a8c\u4e0e\u6570\u636e\u8fed\u4ee3\u3002\u6765\u81eaGroundingPI\u6280\u672f\u62a5\u544a\u3002<\/p>\n

\u4efb\u52a1\u5bf9\u8bc1\u636e\u7684\u8981\u6c42\u4e5f\u4e0d\u540c\u3002\u4e00\u4e2a\u5df2\u786e\u8ba4\u7684\u676f\u5b50\u53ef\u4ee5\u652f\u6491\u5c40\u90e8\u6307\u4ee3\u8bad\u7ec3，\u4f46\u201c\u627e\u51fa\u6240\u6709\u676f\u5b50\u201d\u8fd8\u9700\u8981\u5b8c\u6574\u8986\u76d6；None\u8981\u6c42\u786e\u8ba4\u5bf9\u8c61\u4e0d\u5b58\u5728，\u7a7a\u9884\u6d4b\u672c\u8eab\u4e0d\u80fd\u8bc1\u660e\u4e0d\u5b58\u5728\u3002\u7269\u4f53\u2014\u90e8\u4ef6\u548c\u5355\u8bcd\u2014\u6587\u672c\u884c\u5173\u7cfb\u4e5f\u5e94\u4fdd\u7559，\u4e0d\u80fd\u88ab\u53bb\u91cd\u7b97\u6cd5\u8bef\u5408\u5e76\u3002<\/p>\n

\u5f53\u7ec6\u8282\u4e0d\u6e05\u65f6，\u5f15\u5165\u5c40\u90e8\u89c2\u5bdf，\u5e76\u5c06\u88c1\u526a\u7ed3\u679c\u6620\u5c04\u56de\u539f\u56fe\u5750\u6807\u3002\u5982\u679c\u76d1\u7763\u4f9d\u8d56\u539f\u59cb\u8f93\u5165\u6839\u672c\u4e0d\u53ef\u89c1\u7684\u7ec6\u8282，\u5e94\u628a\u6240\u9700\u5c40\u90e8\u89c6\u56fe\u4e00\u8d77\u4fdd\u7559，\u907f\u514d\u5f3a\u8feb\u6a21\u578b\u9884\u6d4b\u8f93\u5165\u4e2d\u4e0d\u5b58\u5728\u7684\u4fe1\u606f\u3002\u6807\u7b7e\u4fee\u8ba2\u8fd8\u8981\u4f20\u9012\u7ed9\u4f9d\u8d56\u5b83\u7684\u4e0b\u6e38\u4efb\u52a1\u3002<\/p>\n

\u8fd9\u7ed9\u6570\u636e\u56e2\u961f\u7684\u542f\u53d1\u5f88\u5b9e\u9645：\u6807\u6ce8\u8d28\u91cf\u4e0d\u4ec5\u662f\u5355\u4e2a\u6846\u753b\u5f97\u51c6，\u66f4\u662f\u4efb\u52a1\u4f9d\u8d56\u662f\u5426\u6ee1\u8db3\u3002\u591a\u4e2a\u6559\u5e08\u540c\u610f\u4e00\u4e2a\u7b54\u6848，\u4ecd\u53ef\u80fd\u5171\u4eab\u504f\u5dee；\u8d28\u91cf\u4f53\u7cfb\u9700\u8981\u8bb0\u5f55\u672a\u89e3\u51b3\u72b6\u6001，\u800c\u4e0d\u662f\u5c06\u4e0d\u786e\u5b9a\u6027\u5f3a\u884c\u53d8\u6210\u6807\u7b7e\u3002<\/p>\n

4. \u4ece\u8bed\u8a00\u76d1\u7763\u8d70\u5411\u7a7a\u95f4\u7ed3\u679c\u4f18\u5316<\/h3>\n

\u8bad\u7ec3\u603b\u4f53\u5305\u542b\u57fa\u7840VLM\u9884\u8bad\u7ec3\u3001\u5750\u6807\u5bf9\u9f50SFT\u4e0e\u5f3a\u5316\u540e\u8bad\u7ec3\u3002\u57fa\u7840VLM\u5148\u51bb\u7ed3\u4e24\u4e2a\u4e3b\u5e72，\u53ea\u8bad\u7ec3\u6295\u5f71\u5668；\u968f\u540e\u8fdb\u884c\u5168\u53c2\u6570\u591a\u6a21\u6001\u9884\u8bad\u7ec3\u548c\u4e00\u822c\u56fe\u50cf\/\u89c6\u9891\u7406\u89e3\u3002\u5750\u6807\u5bf9\u9f50\u9636\u6bb5\u518d\u5b66\u4e60\u7a7a\u95f4\u4efb\u52a1\u53ca\u8f93\u51fa\u534f\u8bae\u3002<\/p>\n

\"\u886810\uff1a\u57fa\u7840VLM\u4e0eSFT\u8bad\u7ec3\"<\/p>\n

\u886810｜\u57fa\u7840VLM\u4e0eSFT\u8bad\u7ec3\u3002\u661f\u53f7\u3001\u7f3a\u5931\u9879\u4e0e\u534f\u8bae\u4f8b\u5916\u6cbf\u7528\u8bba\u6587，\u9700\u7ed3\u5408\u6b63\u6587\u89e3\u91ca\u3002<\/p>\n

\u76d1\u7763\u9636\u6bb5\u4ecd\u7136\u662f\u81ea\u56de\u5f52\u4ea4\u53c9\u71b5，\u5750\u6807token\u4e0e\u666e\u901a\u54cd\u5e94token\u4f7f\u7528\u76f8\u540c\u7684\u9884\u6d4b\u6846\u67b6，\u6ca1\u6709\u989d\u5916\u628a\u6bcf\u4e2atoken\u76f4\u63a5\u53d8\u6210\u8fde\u7eedIoU\u56de\u5f52\u3002\u5b83\u5bb9\u6613\u5b9e\u73b0\u7edf\u4e00\u5e8f\u5217\u8bad\u7ec3，\u4f46token\u4f3c\u7136\u4e0e\u6700\u7ec8\u96c6\u5408\u8d28\u91cf\u5e76\u4e0d\u5b8c\u5168\u4e00\u81f4：\u91cd\u590d\u4e00\u4e2a\u6846\u53ef\u80fd\u53ea\u9020\u6210\u5c11\u91cftoken\u5dee\u5f02，\u5374\u7834\u574f\u5b9e\u4f8b\u5b8c\u6574\u6027\u3002<\/p>\n

\"\u56fe4\uff1aSFT\u548cGRPO\u7a7a\u95f4\u540e\u8bad\u7ec3\"<\/p>\n

\u56fe4｜SFT\u548cGRPO\u7a7a\u95f4\u540e\u8bad\u7ec3\u3002\u6765\u81eaGroundingPI\u6280\u672f\u62a5\u544a\u3002<\/p>\n

4.1 GRPO\u4f18\u5316\u7684\u662f\u53ef\u68c0\u67e5\u7684\u7ed3\u679c<\/h4>\n

\u6bcf\u4e2a\u56fe\u50cf\u2014\u67e5\u8be2\u91c7\u68378\u6761\u54cd\u5e94\u3002grounding\u6709\u4e24\u79cd\u5956\u52b1：\u96c6\u5408\u8986\u76d6 <\/p>\n

R<\/p>\n

s<\/p>\n

e<\/p>\n

t<\/p>\n

R_{set}<\/p>\n

<\/span><\/span>R<\/span><\/span>se<\/span>t<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span> \u4e0e\u4e25\u683c\u8f93\u51fa\u8d28\u91cf <\/p>\n

R<\/p>\n

s<\/p>\n

t<\/p>\n

r<\/p>\n

i<\/p>\n

c<\/p>\n

t<\/p>\n

R_{strict}<\/p>\n

<\/span><\/span>R<\/span><\/span>s<\/span>t<\/span>r<\/span>i<\/span>c<\/span>t<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u3002\u5148\u5206\u522b\u5728\u540c\u4e00\u63d0\u793a\u7684\u54cd\u5e94\u7ec4\u5185\u6807\u51c6\u5316，\u518d\u63090.7\u548c0.3\u7ec4\u5408；OCR\u5219\u4f7f\u7528\u4e00\u4e2a\u7efc\u5408\u5956\u52b1\u3002\u7ec4\u5408\u4f18\u52bf\u8fd8\u5728\u6574\u6279\u54cd\u5e94\u4e0a\u6807\u51c6\u5316\u3002<\/p>\n

<\/p>\n

A<\/p>\n

~<\/p>\n

i<\/p>\n

=<\/p>\n

0.7<\/p>\n

Z<\/p>\n

(<\/p>\n

R<\/p>\n

s<\/p>\n

e<\/p>\n

t<\/p>\n

,<\/p>\n

i<\/p>\n

)<\/p>\n

+<\/p>\n

0.3<\/p>\n

Z<\/p>\n

(<\/p>\n

R<\/p>\n

s<\/p>\n

t<\/p>\n

r<\/p>\n

i<\/p>\n

c<\/p>\n

t<\/p>\n

,<\/p>\n

i<\/p>\n

)<\/p>\n

,<\/p>\n

A<\/p>\n

i<\/p>\n

=<\/p>\n

A<\/p>\n

~<\/p>\n

i<\/p>\n

\u2212<\/p>\n

\u03bc<\/p>\n

A<\/p>\n

~<\/p>\n

s<\/p>\n

A<\/p>\n

~<\/p>\n

+<\/p>\n

10<\/p>\n

\u2212<\/p>\n

8<\/p>\n

.<\/p>\n

\\\\widetilde A_i=0.7Z(R_{set,i})+0.3Z(R_{strict,i}),\\\\qquad A_i=\\\\frac{\\\\widetilde A_i-\\\\mu_{\\\\widetilde A}}{s_{\\\\widetilde A}+10^{-8}}. <\/p>\n

<\/span><\/span><\/span>A<\/span><\/span><\/span><\/p>\n

<\/span><\/span><\/span><\/span><\/span><\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>0.7<\/span>Z<\/span>(<\/span>R<\/span><\/span>se<\/span>t<\/span>,<\/span>i<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>)<\/span><\/span>+<\/span><\/span><\/span><\/span>0.3<\/span>Z<\/span>(<\/span>R<\/span><\/span>s<\/span>t<\/span>r<\/span>i<\/span>c<\/span>t<\/span>,<\/span>i<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>)<\/span>,<\/span><\/span><\/span>A<\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span><\/span><\/span>s<\/span><\/span><\/span>A<\/span><\/span><\/span><\/p>\n

<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>+<\/span><\/span>1<\/span>0<\/span><\/span>\u2212<\/span>8<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>A<\/span><\/span><\/span><\/p>\n

<\/span><\/span><\/span><\/span><\/span><\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u2212<\/span><\/span>\u03bc<\/span><\/span><\/span>A<\/span><\/span><\/span><\/p>\n

<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>.<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

\u5176\u4e2d\u540c\u7ec4\u5e38\u6570\u5956\u52b1\u5728\u6807\u51c6\u5316\u540e\u8d21\u732e\u96f6\u3002\u6bcf\u6761\u54cd\u5e94\u7684\u4f18\u52bf\u8d4b\u7ed9\u5176\u6240\u6709\u6709\u6548\u8f93\u51fatoken，\u91c7\u7528\u6309\u54cd\u5e94\u957f\u5ea6\u5f52\u4e00\u7684\u88c1\u526a\u76ee\u6807：<\/p>\n

\u8bba\u6587\u5f0f（1）：<\/p>\n

<\/p>\n

J<\/p>\n

(<\/p>\n

\u03b8<\/p>\n

)<\/p>\n

=<\/p>\n

E<\/p>\n

B<\/p>\n

\u2009\u2063<\/p>\n

[<\/p>\n

1<\/p>\n

N<\/p>\n

\u2211<\/p>\n

i<\/p>\n

=<\/p>\n

1<\/p>\n

N<\/p>\n

1<\/p>\n

\u2223<\/p>\n

Y<\/p>\n

i<\/p>\n

\u2223<\/p>\n

\u2211<\/p>\n

t<\/p>\n

=<\/p>\n

1<\/p>\n

\u2223<\/p>\n

Y<\/p>\n

i<\/p>\n

\u2223<\/p>\n

{<\/p>\n

min<\/p>\n

\u2061
\n \u2009\u2063<\/p>\n

(<\/p>\n

r<\/p>\n

i<\/p>\n

,<\/p>\n

t<\/p>\n

A<\/p>\n

i<\/p>\n

,<\/p>\n

clip<\/p>\n

\u2061<\/p>\n

(<\/p>\n

r<\/p>\n

i<\/p>\n

,<\/p>\n

t<\/p>\n

,<\/p>\n

1<\/p>\n

\u2212<\/p>\n

\u03f5<\/p>\n

,<\/p>\n

1<\/p>\n

+<\/p>\n

\u03f5<\/p>\n

)<\/p>\n

A<\/p>\n

i<\/p>\n

)<\/p>\n

\u2212<\/p>\n

\u03b2<\/p>\n

d<\/p>\n

i<\/p>\n

,<\/p>\n

t<\/p>\n

}<\/p>\n

]<\/p>\n

.<\/p>\n

\\\\mathcal J(\\\\theta)=\\\\mathbb E_{\\\\mathcal B}\\\\!\\\\left[\\\\frac{1}{N}\\\\sum_{i=1}^{N}\\\\frac{1}{|Y_i|}\\\\sum_{t=1}^{|Y_i|}\\\\left\\\\{\\\\min\\\\!\\\\left(r_{i,t}A_i,\\\\operatorname{clip}(r_{i,t},1-\\\\epsilon,1+\\\\epsilon)A_i\\\\right)-\\\\beta d_{i,t}\\\\right\\\\}\\\\right]. <\/p>\n

<\/span><\/span>J<\/span>(<\/span>\u03b8<\/span>)<\/span><\/span>=<\/span><\/span><\/span><\/span>E<\/span><\/span>B<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

<\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>N<\/span><\/span><\/span><\/span><\/span><\/span><\/span>1<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>i<\/span>=<\/span>1<\/span><\/span><\/span><\/span><\/span>\u2211<\/span><\/span><\/span><\/span>N<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u2223<\/span>Y<\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u2223<\/span><\/span><\/span><\/span><\/span><\/span><\/span>1<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>t<\/span>=<\/span>1<\/span><\/span><\/span><\/span><\/span>\u2211<\/span><\/span><\/span><\/span>\u2223<\/span>Y<\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u2223<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>{<\/span>min<\/span><\/span><\/span>(<\/span>r<\/span><\/span>i<\/span>,<\/span>t<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>A<\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>,<\/span><\/span>clip<\/span><\/span>(<\/span>r<\/span><\/span>i<\/span>,<\/span>t<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>,<\/span><\/span>1<\/span><\/span>\u2212<\/span><\/span>\u03f5<\/span>,<\/span><\/span>1<\/span><\/span>+<\/span><\/span>\u03f5<\/span>)<\/span>A<\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>)<\/span><\/span><\/span>\u2212<\/span><\/span>\u03b2<\/span>d<\/span><\/span>i<\/span>,<\/span>t<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>}<\/span><\/span><\/span><\/p>\n

<\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>.<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

<\/p>\n

r<\/p>\n

i<\/p>\n

,<\/p>\n

t<\/p>\n

r_{i,t}<\/p>\n

<\/span><\/span>r<\/span><\/span>i<\/span>,<\/span>t<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span> \u4e3a\u5f53\u524d\u7b56\u7565\u4e0e\u65e7\u7b56\u7565\u7684token\u6982\u7387\u6bd4，\u88c1\u526a\u53c2\u6570 <\/p>\n

\u03f5<\/p>\n

=<\/p>\n

0.2<\/p>\n

\\\\epsilon=0.2<\/p>\n

<\/span><\/span>\u03f5<\/span><\/span>=<\/span><\/span><\/span><\/span>0.2<\/span><\/span><\/span><\/span><\/span>，\u53c2\u8003\u6b63\u5219\u6743\u91cd <\/p>\n

\u03b2<\/p>\n

=<\/p>\n

0.02<\/p>\n

\\\\beta=0.02<\/p>\n

<\/span><\/span>\u03b2<\/span><\/span>=<\/span><\/span><\/span><\/span>0.02<\/span><\/span><\/span><\/span><\/span>\u3002\u53c2\u8003\u6a21\u578b\u51bb\u7ed3\u4e3aSFT checkpoint；RL\u51bb\u7ed3\u89c6\u89c9\u7f16\u7801\u5668\u4e0e\u6295\u5f71\u5668，\u4ec5\u66f4\u65b0\u8bed\u8a00\u4fa7\u53c2\u6570\u3002<\/p>\n

\u6b63\u5219\u9879\u4f7f\u7528 <\/p>\n

d<\/p>\n

=<\/p>\n

\u03c1<\/p>\n

\u2212<\/p>\n

log<\/p>\n

\u2061<\/p>\n

\u03c1<\/p>\n

\u2212<\/p>\n

1<\/p>\n

d=\\\\rho-\\\\log\\\\rho-1<\/p>\n

<\/span><\/span>d<\/span><\/span>=<\/span><\/span><\/span><\/span>\u03c1<\/span><\/span>\u2212<\/span><\/span><\/span><\/span>log<\/span><\/span><\/span>\u03c1<\/span><\/span>\u2212<\/span><\/span><\/span><\/span>1<\/span><\/span><\/span><\/span><\/span>，\u5176\u4e2d <\/p>\n

\u03c1<\/p>\n

=<\/p>\n

\u03c0<\/p>\n

r<\/p>\n

e<\/p>\n

f<\/p>\n

\/<\/p>\n

\u03c0<\/p>\n

\u03b8<\/p>\n

\\\\rho=\\\\pi_{ref}\/\\\\pi_\\\\theta<\/p>\n

<\/span><\/span>\u03c1<\/span><\/span>=<\/span><\/span><\/span><\/span>\u03c0<\/span><\/span>r<\/span>e<\/span>f<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\/<\/span>\u03c0<\/span><\/span>\u03b8<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u3002\u5b83\u5728\u5f53\u524d\u7b56\u7565\u91c7\u6837\u65f6\u7684\u671f\u671b\u5bf9\u5e94KL；\u5b9e\u9645\u7528\u65e7\u7b56\u7565rollout\u8bc4\u4f30\u65f6，\u5e94\u7406\u89e3\u4e3a\u91c7\u6837\u6b63\u5219，\u4e0d\u80fd\u76f4\u63a5\u79f0\u4e3a\u5f53\u524d\u7b56\u7565KL\u7684\u65e0\u504f\u4f30\u8ba1\u3002<\/p>\n

\"\u886811\uff1a\u5f3a\u5316\u540e\u8bad\u7ec3\u8bbe\u7f6e\"<\/p>\n

\u886811｜\u5f3a\u5316\u540e\u8bad\u7ec3\u8bbe\u7f6e\u3002\u661f\u53f7\u3001\u7f3a\u5931\u9879\u4e0e\u534f\u8bae\u4f8b\u5916\u6cbf\u7528\u8bba\u6587，\u9700\u7ed3\u5408\u6b63\u6587\u89e3\u91ca\u3002<\/p>\n

4.2 \u8986\u76d6\u5956\u52b1\u4e0d\u662f\u6807\u51c6\u68c0\u6d4bF1<\/h4>\n

<\/p>\n

R<\/p>\n

s<\/p>\n

e<\/p>\n

t<\/p>\n

R_{set}<\/p>\n

<\/span><\/span>R<\/span><\/span>se<\/span>t<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span> \u4e3a\u6bcf\u4e2a\u771f\u503c\u6846\u72ec\u7acb\u9009\u6700\u9ad8IoU\u9884\u6d4b，\u518d\u6838\u9a8c\u7c7b\u522b；\u540c\u4e00\u9884\u6d4b\u53ef\u4ee5\u88ab\u591a\u4e2a\u771f\u503c\u590d\u7528\u3002\u4ee4\u5339\u914d\u5206\u6570\u548c\u4e3a <\/p>\n

S<\/p>\n

S<\/p>\n

<\/span><\/span>S<\/span><\/span><\/span><\/span><\/span>，\u771f\u503c\u6570 <\/p>\n

n<\/p>\n

n<\/p>\n

<\/span><\/span>n<\/span><\/span><\/span><\/span><\/span>，\u9884\u6d4b\u6570 <\/p>\n

m<\/p>\n

m<\/p>\n

<\/span><\/span>m<\/span><\/span><\/span><\/span><\/span>，\u8ba1\u7b97 <\/p>\n

R<\/p>\n

s<\/p>\n

=<\/p>\n

S<\/p>\n

\/<\/p>\n

n<\/p>\n

R_s=S\/n<\/p>\n

<\/span><\/span>R<\/span><\/span>s<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>S<\/span>\/<\/span>n<\/span><\/span><\/span><\/span><\/span>\u3001<\/p>\n

P<\/p>\n

s<\/p>\n

=<\/p>\n

S<\/p>\n

\/<\/p>\n

m<\/p>\n

P_s=S\/m<\/p>\n

<\/span><\/span>P<\/span><\/span>s<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>S<\/span>\/<\/span>m<\/span><\/span><\/span><\/span><\/span>，\u518d\u53d6\u7c7b\u4f3cF1\u7684\u8c03\u548c\u5e73\u5747\u3002<\/p>\n

\u7531\u4e8e\u5339\u914d\u5141\u8bb8\u590d\u7528，\u8fd9\u4e2a\u8986\u76d6\u4ee3\u7406\u5956\u52b1\u53ef\u80fd\u8d85\u8fc71\u3002\u56e0\u6b64\u5b83\u4e0d\u80fd\u76f4\u63a5\u5f53\u6210\u4e00\u5bf9\u4e00\u5339\u914d\u4e0b\u7684\u6807\u51c6F1，\u4e5f\u9700\u8981\u4e25\u683c\u5956\u52b1\u8865\u5145\u91cd\u590d\u6846\u3001\u8d85\u5927\u6846\u3001\u6570\u91cf\u3001\u683c\u5f0f\u548c\u6392\u5e8f\u7ea6\u675f\u3002<\/p>\n

\"\u886812\uff1a\u4e25\u683cgrounding\u5956\u52b1\"<\/p>\n

\u886812｜\u4e25\u683cgrounding\u5956\u52b1\u3002\u661f\u53f7\u3001\u7f3a\u5931\u9879\u4e0e\u534f\u8bae\u4f8b\u5916\u6cbf\u7528\u8bba\u6587，\u9700\u7ed3\u5408\u6b63\u6587\u89e3\u91ca\u3002<\/p>\n

\u4e25\u683c\u5956\u52b1\u628a0.50\/0.75\/0.95\u4e09\u4e2aIoU\u9608\u503c\u7684F1\u3001\u5339\u914dIoU\u7b49\u5206\u9879\u52a0\u6743，\u5e76\u88c1\u526a\u52300\u81f31\u3002\u5148\u5206\u522b\u6807\u51c6\u5316\u518d\u7ec4\u5408，\u610f\u5473\u77400.7\/0.3\u662f\u6807\u51c6\u5316\u540e\u7684\u76f8\u5bf9\u6743\u91cd，\u4e0d\u80fd\u6309\u539f\u59cb\u91cf\u7eb2\u76f4\u89c9\u89e3\u91ca\u3002<\/p>\n

4.3 OCR\u540c\u65f6\u5bf9\u9f50\u6587\u5b57\u4e0e\u51e0\u4f55<\/h4>\n

OCR\u91c7\u7528Hungarian\u4e00\u5bf9\u4e00\u5339\u914d，\u8054\u5408\u8003\u8651\u6587\u5b57\u7f16\u8f91\u76f8\u4f3c\u5ea6\u4e0e\u6846IoU，\u517c\u987e\u4e25\u683c\u6587\u672c\u5339\u914d\u548c\u90e8\u5206\u5339\u914d\u53cd\u9988\u3002\u5176\u4e00\u822c\u5f62\u5f0f\u4e3a <\/p>\n

F<\/p>\n

(<\/p>\n

A<\/p>\n

)<\/p>\n

=<\/p>\n

2<\/p>\n

M<\/p>\n

(<\/p>\n

A<\/p>\n

)<\/p>\n

\/<\/p>\n

(<\/p>\n

n<\/p>\n

+<\/p>\n

m<\/p>\n

)<\/p>\n

F(A)=2M(A)\/(n+m)<\/p>\n

<\/span><\/span>F<\/span>(<\/span>A<\/span>)<\/span><\/span>=<\/span><\/span><\/span><\/span>2<\/span>M<\/span>(<\/span>A<\/span>)<\/span>\/<\/span>(<\/span>n<\/span><\/span>+<\/span><\/span><\/span><\/span>m<\/span>)<\/span><\/span><\/span><\/span><\/span>，\u5176\u4e2d <\/p>\n

M<\/p>\n

(<\/p>\n

A<\/p>\n

)<\/p>\n

M(A)<\/p>\n

<\/span><\/span>M<\/span>(<\/span>A<\/span>)<\/span><\/span><\/span><\/span><\/span> \u662f\u6700\u4f18\u5206\u914d\u7684\u4eb2\u548c\u5ea6\u603b\u548c\u3002<\/p>\n

\u786c\u5339\u914d\u8981\u6c42\u89c4\u8303\u5316\u5b57\u7b26\u4e32\u76f8\u540c\u4e14IoU\u8fc7\u9608\u503c；\u8f6f\u4eb2\u548c\u5ea6\u4f7f\u7528 <\/p>\n

I<\/p>\n

o<\/p>\n

U<\/p>\n

\\\\sqrt{\\\\mathrm{IoU}}<\/p>\n

<\/span><\/span><\/span>IoU<\/span><\/span><\/span><\/span><\/span><\/p>\n

<\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span> \u4e58\u7f16\u8f91\u76f8\u4f3c\u5ea6，\u5e76\u5206\u522b\u8bbe\u51e0\u4f55\u4e0e\u6587\u672c\u9608\u503c\u3002Unicode\u89c4\u8303\u5316\u3001\u5927\u5c0f\u5199\u6298\u53e0\u548c\u6807\u70b9\u5904\u7406\u5c5e\u4e8e\u5956\u52b1\u5b9a\u4e49，\u4e0d\u80fd\u5728\u590d\u73b0\u65f6\u968f\u610f\u66f4\u6362\u3002<\/p>\n

\"\u886813\uff1aOCR\u7c92\u5ea6\u805a\u5408\"<\/p>\n

\u886813｜OCR\u7c92\u5ea6\u805a\u5408\u3002\u661f\u53f7\u3001\u7f3a\u5931\u9879\u4e0e\u534f\u8bae\u4f8b\u5916\u6cbf\u7528\u8bba\u6587，\u9700\u7ed3\u5408\u6b63\u6587\u89e3\u91ca\u3002<\/p>\n

\u6587\u672c\u53ef\u80fd\u5b58\u5728\u8bcd\u7ea7\u548c\u884c\u7ea7\u4e24\u79cd\u5408\u7406\u7c92\u5ea6\u3002\u62a5\u544a\u5728\u5b8c\u6574\u89c6\u56fe\u4e4b\u95f4\u805a\u5408，\u5e76\u5bf9\u51b2\u7a81\u6807\u6ce8\u964d\u6743，\u800c\u975e\u628a\u4e92\u4e0d\u517c\u5bb9\u7684\u96f6\u6563\u5339\u914d\u62fc\u6210\u4e00\u4e2a\u9ad8\u5206\u3002\u6700\u540e\u8fd8\u7528\u683c\u5f0f\u95e8\u63a7\u53ca\u91cd\u590d\u3001\u8fc7\u91cf\u3001\u65e0\u6548\u6846\u60e9\u7f5a\u3002\u8fd9\u4f53\u73b0\u4e00\u4e2a\u73b0\u5b9e\u95ee\u9898：\u5956\u52b1\u9700\u8981\u9f13\u52b1\u53ef\u4ea4\u4ed8\u7ed3\u6784，\u800c\u4e0d\u53ea\u662f\u5c40\u90e8\u8bc6\u522b\u6b63\u786e\u3002<\/p>\n

\u8bad\u7ec3\u5956\u52b1\u4e0e\u6d4b\u8bd5\u6307\u6807\u662f\u4e24\u5957\u5b9a\u4e49\u3002\u524d\u8005\u670d\u52a1\u4f18\u5316，\u540e\u8005\u670d\u52a1\u6bd4\u8f83；\u5982\u679c\u53ea\u5c55\u793a\u5956\u52b1\u4e0a\u5347，\u4e0d\u80fd\u8bc1\u660e\u72ec\u7acb\u57fa\u51c6\u80fd\u529b\u63d0\u5347\u3002<\/p>\n

5. \u611f\u77e5\u7ed3\u679c\u8981\u62c6\u5f00\u770b，\u5e73\u5747\u5206\u4f1a\u63a9\u76d6\u80fd\u529b\u5dee\u5f02<\/h3>\n

\u62a5\u544a\u572834\u4e2a\u57fa\u51c6\u300111\u7c7b\u80fd\u529b\u4e0a\u6bd4\u8f8344\u4e2a\u57fa\u7ebf，GroundingPI\u7efc\u5408\u5747\u520673.68，GPT-6 Astra\u4e3a71.54\u3002Astra\u5728\u8fd9\u9879\u5b9e\u9a8c\u4e2d\u662f\u5f3a\u611f\u77e5\u53c2\u7167，\u5e76\u975e\u88ab\u4f5c\u4e3a\u673a\u5668\u4eba\u52a8\u4f5c\u4e3b\u5e72\u505a\u4e86\u76f8\u540c\u9002\u914d\u3002<\/p>\n

\"\u56fe5\uff1a\u5404\u7c7b\u611f\u77e5\u4efb\u52a1\u6bd4\u8f83\"<\/p>\n

\u56fe5｜\u5404\u7c7b\u611f\u77e5\u4efb\u52a1\u6bd4\u8f83\u3002\u6765\u81eaGroundingPI\u6280\u672f\u62a5\u544a\u3002<\/p>\n

\u8fd9\u4e5f\u4e0d\u662f\u6240\u6709\u4efb\u52a1\u5168\u9762\u9886\u5148\u3002Dense200\u6846\u5b9a\u4f4dF1mIoU\u4e3a74.53，\u5bf9\u7167Astra65.04；\u5728GUI\u4e0e\u90e8\u5206\u7a7a\u95f4\u4efb\u52a1\u4e0a，Astra\u4ecd\u6709\u4f18\u52bf；TotalText\u4e0eDocLayNet\u4e5f\u5206\u522b\u5b58\u5728\u66f4\u5f3a\u7684\u6bd4\u8f83\u7cfb\u7edf\u3002<\/p>\n

\"\u88681\uff1a\u68c0\u6d4b\u4e0e\u6307\u4ee3\u6982\u89c8\"<\/p>\n

\u88681｜\u68c0\u6d4b\u4e0e\u6307\u4ee3\u6982\u89c8\u3002\u661f\u53f7\u3001\u7f3a\u5931\u9879\u4e0e\u534f\u8bae\u4f8b\u5916\u6cbf\u7528\u8bba\u6587，\u9700\u7ed3\u5408\u6b63\u6587\u89e3\u91ca\u3002<\/p>\n

\"\u88682\uff1a\u673a\u5668\u4eba\u7a7a\u95f4\u4e0eGUI\u6982\u89c8\"<\/p>\n

\u88682｜\u673a\u5668\u4eba\u7a7a\u95f4\u4e0eGUI\u6982\u89c8\u3002\u661f\u53f7\u3001\u7f3a\u5931\u9879\u4e0e\u534f\u8bae\u4f8b\u5916\u6cbf\u7528\u8bba\u6587，\u9700\u7ed3\u5408\u6b63\u6587\u89e3\u91ca\u3002<\/p>\n

\"\u88683\uff1aOCR\u4e0e\u6587\u6863\u5e03\u5c40\u6982\u89c8\"<\/p>\n

\u88683｜OCR\u4e0e\u6587\u6863\u5e03\u5c40\u6982\u89c8\u3002\u661f\u53f7\u3001\u7f3a\u5931\u9879\u4e0e\u534f\u8bae\u4f8b\u5916\u6cbf\u7528\u8bba\u6587，\u9700\u7ed3\u5408\u6b63\u6587\u89e3\u91ca\u3002<\/p>\n

\"\u88684\uff1a\u5bf9\u8c61pointing\u6982\u89c8\"<\/p>\n

\u88684｜\u5bf9\u8c61pointing\u6982\u89c8\u3002\u661f\u53f7\u3001\u7f3a\u5931\u9879\u4e0e\u534f\u8bae\u4f8b\u5916\u6cbf\u7528\u8bba\u6587，\u9700\u7ed3\u5408\u6b63\u6587\u89e3\u91ca\u3002<\/p>\n

F1mIoU\u6307\u8de8IoU\u9608\u503c\u7684F1\u6c47\u603b，F1@Point\u5219\u7ed3\u5408\u70b9\u662f\u5426\u843d\u5165\u76ee\u6807\u63a9\u7801\u4e0e\u5b9e\u4f8b\u5339\u914d\u3002\u80fd\u628a\u70b9\u653e\u5728\u7269\u4f53\u5185\u90e8，\u4e0e\u628a\u6574\u4e2a\u8fb9\u754c\u6846\u5b9a\u4f4d\u51c6\u786e，\u662f\u4e0d\u540c\u6311\u6218\u3002Dense200\u4e0a\u70b9\u4e0e\u6846\u7684\u6392\u5e8f\u53d8\u5316，\u6070\u597d\u63d0\u9192\u6211\u4eec\u4e0d\u80fd\u7528\u4e00\u4e2a\u7edf\u4e00\u53e3\u53f7\u8986\u76d6\u6240\u6709\u7a7a\u95f4\u4efb\u52a1\u3002<\/p>\n

\u62a5\u544a\u5bf9\u81ea\u8eab\u6a21\u578b\u91c7\u7528\u4e94\u4e2a\u79cd\u5b50\u3001\u6bcf\u4e2a\u4e24\u6b21\u5171\u5341\u6b21\u8fd0\u884c\u5747\u503c，\u90e8\u5206\u672c\u5730\u57fa\u7ebf\u4e3a\u4e09\u4e2a\u79cd\u5b50\u5404\u4e00\u6b21，\u5176\u4ed6\u884c\u53d6\u5916\u90e8\u516c\u5f00\u503c\u3002\u661f\u53f7\u53ef\u80fd\u8868\u793a\u5916\u90e8\u7ed3\u679c\u6216\u652f\u6301\u4f8b\u5916，N\/A\u4e0e\u7f3a\u5931\u503c\u4e0d\u80fd\u5f53\u96f6\u5206；\u63d0\u793a\u548c\u89e3\u6790\u534f\u8bae\u4e0d\u517c\u5bb9，\u4e5f\u4e0d\u80fd\u76f4\u63a5\u8bc1\u660e\u6a21\u578b\u5185\u5728\u6ca1\u6709\u8be5\u80fd\u529b\u3002<\/p>\n

\u5b8c\u6574\u4efb\u52a1\u8868\u5728\u540e\u6587\u6309\u8bc4\u4ef7\u5bf9\u8c61\u5206\u7ec4\u4fdd\u7559，\u4ee5\u4fbf\u4ece\u53ec\u56de\u3001\u7cbe\u5ea6\u3001\u4e25\u683c\u5b9a\u4f4d\u9608\u503c\u4e0e\u89e3\u6790\u9519\u8bef\u7387\u5206\u522b\u7406\u89e3\u5dee\u8ddd\u3002<\/p>\n

6. \u52a8\u4f5c\u8fc1\u79fb\u5b9e\u9a8c\u63a7\u5236\u4e86\u4ec0\u4e48<\/h3>\n

GroundingPI \u7684\u673a\u5668\u4eba\u6bd4\u8f83\u628a\u4e0d\u540c\u57fa\u7840\u6a21\u578b\u63a5\u5230\u540c\u4e00\u4e2aAction DiT\u3002\u6bcf\u6b21\u89c2\u6d4b\u4e0e\u6307\u4ee4\u53ea\u8fd0\u884c\u4e00\u6b21\u4e3b\u5e72，\u5728\u516b\u4e2a\u5f52\u4e00\u5316\u6df1\u5ea6\u62bd\u53d6\u9690\u85cf\u72b6\u6001，\u6295\u5f71\u3001\u52a0\u6df1\u5ea6\u5d4c\u5165\u5e76\u91cd\u91c7\u6837\u4e3a\u6bcf\u5c4264\u4e2a\u6761\u4ef6token，\u5bbd\u5ea6\u7edf\u4e001024\u3002<\/p>\n

\"\u886814\uff1a\u7edf\u4e00Action<\/p>\n

\u886814｜\u7edf\u4e00Action DiT\u67b6\u6784\u3002\u661f\u53f7\u3001\u7f3a\u5931\u9879\u4e0e\u534f\u8bae\u4f8b\u5916\u6cbf\u7528\u8bba\u6587，\u9700\u7ed3\u5408\u6b63\u6587\u89e3\u91ca\u3002<\/p>\n

Action DiT\u670916\u4e2a\u539f\u5b50block，8\u4e2across-attention\u4e0e8\u4e2aself-attention\u4ea4\u66ff，\u800c\u4e0d\u662f16\u7ec4\u5404\u542b\u4e24\u79cd\u6ce8\u610f\u529b\u7684\u6210\u5bf9block\u3002\u52a8\u4f5chorizon\u4e3a16，\u63a8\u7406\u4f7f\u75284\u6b65Euler；\u611f\u77e5\u7279\u5f81\u5728\u52a8\u4f5c\u53bb\u566a\u8fc7\u7a0b\u4e2d\u7f13\u5b58\u590d\u7528\u3002<\/p>\n

\u8bad\u7ec3\u7528 <\/p>\n

a<\/p>\n

t<\/p>\n

=<\/p>\n

(<\/p>\n

1<\/p>\n

\u2212<\/p>\n

t<\/p>\n

)<\/p>\n

\u03f5<\/p>\n

+<\/p>\n

t<\/p>\n

a<\/p>\n

a_t=(1-t)\\\\epsilon+ta<\/p>\n

<\/span><\/span>a<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>(<\/span>1<\/span><\/span>\u2212<\/span><\/span><\/span><\/span>t<\/span>)<\/span>\u03f5<\/span><\/span>+<\/span><\/span><\/span><\/span>t<\/span>a<\/span><\/span><\/span><\/span><\/span>，\u76ee\u6807\u901f\u5ea6 <\/p>\n

a<\/p>\n

\u2212<\/p>\n

\u03f5<\/p>\n

a-\\\\epsilon<\/p>\n

<\/span><\/span>a<\/span><\/span>\u2212<\/span><\/span><\/span><\/span>\u03f5<\/span><\/span><\/span><\/span><\/span>，\u5bf9\u52a8\u4f5c\u4f4d\u7f6e\u6700\u5c0f\u5316\u901f\u5ea6\u9884\u6d4b\u5e73\u65b9\u8bef\u5dee\u3002\u4e0d\u540c\u57fa\u7840\u6a21\u578b\u5171\u7528\u52a8\u4f5c\u8868\u793a\u3001\u52a8\u4f5c\u6570\u636e\u3001\u4f18\u5316\u9884\u7b97\u548c\u63a8\u7406\u8fc7\u7a0b；\u6309\u4e3b\u5e72\u5bbd\u5ea6\u9002\u914d\u7684\u5c0f\u6295\u5f71\u5668\u662f\u63a5\u53e3\u53c2\u6570\u5dee\u5f02\u4e4b\u4e00\u3002<\/p>\n

\"\u886815\uff1a\u52a8\u4f5c\u5b66\u4e60\u9884\u7b97\"<\/p>\n

\u886815｜\u52a8\u4f5c\u5b66\u4e60\u9884\u7b97\u3002\u661f\u53f7\u3001\u7f3a\u5931\u9879\u4e0e\u534f\u8bae\u4f8b\u5916\u6cbf\u7528\u8bba\u6587，\u9700\u7ed3\u5408\u6b63\u6587\u89e3\u91ca\u3002<\/p>\n

\u7edf\u4e00\u52a8\u4f5c\u5934\u6709\u52a9\u4e8e\u7814\u7a76\u8868\u793a\u8fc1\u79fb，\u4f46\u4e5f\u89c4\u5b9a\u4e86\u7ed3\u8bba\u8303\u56f4\u3002\u5b83\u6ca1\u6709\u6bd4\u8f83\u6bcf\u4e2aVLA\u6216WAM\u7cfb\u7edf\u539f\u751f\u7684\u6700\u4f18\u52a8\u4f5c\u67b6\u6784，\u4e5f\u4e0d\u80fd\u8bc1\u660e\u67d0\u7c7b\u57fa\u7840\u6a21\u578b\u5728\u6240\u6709\u9002\u914d\u914d\u65b9\u4e0b\u66f4\u5f31\u3002<\/p>\n

\"\u56fe6\uff1a\u53d7\u63a7\u52a8\u4f5c\u8fc1\u79fb\u53ca\u5f00\u73af\u9a7e\u9a76\u8bef\u5dee\"<\/p>\n

\u56fe6｜\u53d7\u63a7\u52a8\u4f5c\u8fc1\u79fb\u53ca\u5f00\u73af\u9a7e\u9a76\u8bef\u5dee\u3002\u6765\u81eaGroundingPI\u6280\u672f\u62a5\u544a\u3002<\/p>\n

6.1 OOD\u4f18\u52bf\u662f\u771f\u5b9e\u4fe1\u53f7，\u4f4e\u7edd\u5bf9\u6210\u529f\u7387\u4e5f\u5fc5\u987b\u4fdd\u7559<\/h4>\n

GroundingPI\u5728\u516d\u79cd\u673a\u5668\u4eba\u8bbe\u7f6e\u4e2d\u7684\u4e94\u79cd\u7b2c\u4e00，\u8986\u76d6\u56db\u4e2aOOD\u8bbe\u7f6e\u3002RoboTwin2.0 Full\u4e3a66.20%；Clean2Random\u4ece\u5e72\u51c0\u6f14\u793a\u8fc1\u79fb\u5230\u968f\u673a\u573a\u666f，GroundingPI\u4e3a17.60%，Rex-Omni\u4e3a14.10%\u3002\u76f8\u5bf9\u63d0\u5347\u7ea624.8%，\u5bf9\u5e94\u7edd\u5bf9\u589e\u52a03.50\u4e2a\u767e\u5206\u70b9\u3002<\/p>\n

\u8fd9\u4e24\u4e2a\u8868\u8ff0\u90fd\u91cd\u8981\u300224.8%\u7684\u76f8\u5bf9\u63d0\u5347\u8868\u660e\u8fdb\u6b65\u660e\u663e，17.60%\u7684\u7edd\u5bf9\u6210\u529f\u7387\u5219\u8bf4\u660e\u4efb\u52a1\u4f9d\u7136\u56f0\u96be\u3002\u5c06\u7ed3\u679c\u5199\u6210\u201c\u6210\u529f\u7387\u63d0\u534724.8\u4e2a\u767e\u5206\u70b9\u201d\u4f1a\u5938\u5927\u6539\u5584；\u5c06\u5176\u79f0\u4e3a\u6210\u719f\u673a\u5668\u4eba\u80fd\u529b\u540c\u6837\u7f3a\u4e4f\u4f9d\u636e\u3002<\/p>\n

RoboCasa-GR1 Full\u4e2d，GroundingPI37.75\u4f4e\u4e8eRynnBrain39.00\u3002\u5b83\u7684\u5f3a\u9879\u5c24\u5176\u4f53\u73b0\u5728\u5206\u5e03\u53d8\u5316\u4e0e\u4f4e\u6570\u636e\u9002\u914d，\u4e0d\u662f\u6bcf\u9879ID\u7ed3\u679c\u90fd\u9886\u5148\u3002<\/p>\n

\"\u56fe7\uff1a\u611f\u77e5\u9884\u8bad\u7ec3\u89d2\u8272\u7684\u6982\u5ff5\u793a\u610f\"<\/p>\n

\u56fe7｜\u611f\u77e5\u9884\u8bad\u7ec3\u89d2\u8272\u7684\u6982\u5ff5\u793a\u610f\u3002\u6765\u81eaGroundingPI\u6280\u672f\u62a5\u544a\u3002<\/p>\n

\u56fe7\u63cf\u7ed8\u611f\u77e5\u8bad\u7ec3\u7684\u89d2\u8272\u5206\u5de5，\u56fe\u5f62\u5206\u533a\u53ea\u662f\u6982\u5ff5\u793a\u610f，\u5e76\u975e\u6d4b\u5f97\u7684\u53c2\u6570\u91cf\u6216\u6570\u636e\u91cf\u6bd4\u4f8b\u3002\u89c6\u89c9\u57fa\u5143\u80fd\u4f5c\u4e3a\u7269\u7406\u63d0\u793a\u63a5\u53e3，\u4f46\u73b0\u6709\u8fc1\u79fb\u5b9e\u9a8c\u6ca1\u6709\u5355\u72ec\u9694\u79bb\u201c\u52a0\u5165\u7269\u7406\u63d0\u793a\u201d\u8fd9\u4e00\u5e72\u9884\u7684\u6536\u76ca\u3002<\/p>\n

6.2 \u6f14\u793a\u6570\u636e\u6548\u7387\u4e0d\u7b49\u4e8e\u603b\u8bad\u7ec3\u6210\u672c\u66f4\u4f4e<\/h4>\n

\"\u56fe8\uff1a\u52a8\u4f5c\u6f14\u793a\u6570\u636e\u6548\u7387\"<\/p>\n

\u56fe8｜\u52a8\u4f5c\u6f14\u793a\u6570\u636e\u6548\u7387\u3002\u6765\u81eaGroundingPI\u6280\u672f\u62a5\u544a\u3002<\/p>\n

\u5728RoboCasa-GR1 ID\u4e0a，50%\u6f14\u793a\u6570\u636e\u7684GroundingPI\u8fbe\u523028.75%，\u9ad8\u4e8e\u4f7f\u752875%\u6f14\u793a\u7684\u6700\u5f3a\u57fa\u7ebf27.75%\u3002\u5269\u4f59\u52a8\u4f5c\u5b66\u4e60\u8bbe\u7f6e\u56fa\u5b9a，\u652f\u6301\u5b83\u5728\u8be5\u4efb\u52a1\u4e2d\u63d0\u9ad8\u6f14\u793a\u5229\u7528\u6548\u7387\u3002<\/p>\n

\u4f46\u662f，\u5176\u611f\u77e5\u9884\u8bad\u7ec3\u672c\u8eab\u6d88\u8017\u4e86\u5927\u91cf\u56fe\u6587\u4e0e\u7a7a\u95f4\u6570\u636e\u3002\u4e0d\u80fd\u7531\u201c\u5c11\u7528\u673a\u5668\u4eba\u6f14\u793a\u201d\u63a8\u5bfc\u201c\u7aef\u5230\u7aef\u603b\u6570\u636e\u6216\u603b\u7b97\u529b\u66f4\u5c11\u201d\u3002\u56db\u4e2a\u6570\u636e\u6bd4\u4f8b\u7684\u5b9e\u9a8c\u4e5f\u6ca1\u6709\u5355\u72ec\u5206\u79bb\u611f\u77e5\u5b66\u4e60\u548c\u8fd0\u52a8\u63a7\u5236\u7684\u6837\u672c\u590d\u6742\u5ea6\u3002<\/p>\n

6.3 \u81ea\u52a8\u9a7e\u9a76\u76840.296\u7c73\u5c5e\u4e8e\u5f00\u73af\u8f68\u8ff9\u6307\u6807<\/h4>\n

nuScenes\u4efb\u52a1\u8f93\u5165\u5f53\u524d\u524d\u89c6\u56fe\u548c\u6700\u591a\u4e09\u79d2\u5386\u53f2\u81ea\u8f66\u72b6\u6001，\u8f93\u51fa\u672a\u6765\u4e09\u79d2\u3001\u6bcf0.5\u79d2\u4e00\u4e2a\u7d2f\u8ba1\u4f4d\u7f6e，\u5171\u516d\u70b9\u3002\u6a2a\u7eb5\u5750\u6807\u91c7\u7528\u56fa\u5b9a\u7c73\u5236\u8303\u56f4\u91cf\u5316\u5230\u540c\u4e00\u5750\u6807\u8bcd\u8868；\u8fd9\u4e9b\u662f\u7269\u7406\u5750\u6807，\u4e0d\u662f\u56fe\u50cf\u5f52\u4e00\u5316\u5750\u6807\u3002<\/p>\n

<\/p>\n

q<\/p>\n

z<\/p>\n

=<\/p>\n

round<\/p>\n

\u2061<\/p>\n

e<\/p>\n

v<\/p>\n

e<\/p>\n

n<\/p>\n

\u2009\u2063<\/p>\n

(<\/p>\n

999<\/p>\n

z<\/p>\n

\u2212<\/p>\n

a<\/p>\n

b<\/p>\n

\u2212<\/p>\n

a<\/p>\n

)<\/p>\n

,<\/p>\n

z<\/p>\n

^<\/p>\n

=<\/p>\n

a<\/p>\n

+<\/p>\n

(<\/p>\n

b<\/p>\n

\u2212<\/p>\n

a<\/p>\n

)<\/p>\n

q<\/p>\n

z<\/p>\n

\/<\/p>\n

999.<\/p>\n

q_z=\\\\operatorname{round}_{even}\\\\!\\\\left(999\\\\frac{z-a}{b-a}\\\\right),\\\\qquad \\\\hat z=a+(b-a)q_z\/999. <\/p>\n

<\/span><\/span>q<\/span><\/span>z<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>round<\/span><\/span><\/span>e<\/span>v<\/span>e<\/span>n<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span><\/span>999<\/span><\/span><\/span>b<\/span><\/span>\u2212<\/span><\/span>a<\/span><\/span><\/span><\/span><\/span><\/span><\/span>z<\/span><\/span>\u2212<\/span><\/span>a<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>)<\/span><\/span><\/span><\/span>,<\/span><\/span><\/span><\/span>z<\/span><\/span><\/span>^<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>a<\/span><\/span>+<\/span><\/span><\/span><\/span>(<\/span>b<\/span><\/span>\u2212<\/span><\/span><\/span><\/span>a<\/span>)<\/span>q<\/span><\/span>z<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\/999.<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

\u7eb5\u5411\u8303\u56f4\u4e3a <\/p>\n

[<\/p>\n

\u2212<\/p>\n

60<\/p>\n

,<\/p>\n

60<\/p>\n

]<\/p>\n

[-60,60]<\/p>\n

<\/span><\/span>[<\/span>\u2212<\/span>60<\/span>,<\/span><\/span>60<\/span>]<\/span><\/span><\/span><\/span><\/span> \u7c73，\u6a2a\u5411 <\/p>\n

[<\/p>\n

\u2212<\/p>\n

20<\/p>\n

,<\/p>\n

20<\/p>\n

]<\/p>\n

[-20,20]<\/p>\n

<\/span><\/span>[<\/span>\u2212<\/span>20<\/span>,<\/span><\/span>20<\/span>]<\/span><\/span><\/span><\/span><\/span> \u7c73\u3002\u6bcf\u70b9\u5df2\u7ecf\u662f\u76f8\u5bf9\u5f53\u524d\u8f66\u4f53\u7684\u7d2f\u8ba1\u4f4d\u7f6e，\u4e0d\u80fd\u518d\u7d2f\u52a0\u4e00\u6b21\u3002\u8bc4\u4ef7\u4f7f\u7528\u672a\u91cf\u5316\u771f\u503c\u3002<\/p>\n

<\/p>\n

L<\/p>\n

2<\/p>\n

@<\/p>\n

h<\/p>\n

=<\/p>\n

1<\/p>\n

2<\/p>\n

h<\/p>\n

\u2211<\/p>\n

k<\/p>\n

=<\/p>\n

1<\/p>\n

2<\/p>\n

h<\/p>\n

d<\/p>\n

\u02c9<\/p>\n

k<\/p>\n

,<\/p>\n

L<\/p>\n

2<\/p>\n

A<\/p>\n

v<\/p>\n

g<\/p>\n

=<\/p>\n

1<\/p>\n

3<\/p>\n

\u2211<\/p>\n

h<\/p>\n

=<\/p>\n

1<\/p>\n

3<\/p>\n

L<\/p>\n

2<\/p>\n

@<\/p>\n

h<\/p>\n

.<\/p>\n

\\\\mathrm{L2}@h=\\\\frac{1}{2h}\\\\sum_{k=1}^{2h}\\\\bar d_k,\\\\qquad \\\\mathrm{L2}_{Avg}=\\\\frac13\\\\sum_{h=1}^3\\\\mathrm{L2}@h. <\/p>\n

<\/span><\/span>L2<\/span><\/span>@<\/span>h<\/span><\/span>=<\/span><\/span><\/span><\/span><\/span><\/span>2<\/span>h<\/span><\/span><\/span><\/span><\/span><\/span><\/span>1<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>k<\/span>=<\/span>1<\/span><\/span><\/span><\/span><\/span>\u2211<\/span><\/span><\/span><\/span>2<\/span>h<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>d<\/span><\/span><\/span>\u02c9<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>k<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>,<\/span><\/span><\/span>L2<\/span><\/span><\/span>A<\/span>v<\/span>g<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span><\/span><\/span>3<\/span><\/span><\/span><\/span><\/span><\/span><\/span>1<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>h<\/span>=<\/span>1<\/span><\/span><\/span><\/span><\/span>\u2211<\/span><\/span><\/span><\/span>3<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>L2<\/span><\/span>@<\/span>h<\/span>.<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

\u8fd9\u91cc\u6309\u5404\u65f6\u95f4\u8303\u56f4\u5185\u7684\u7d2f\u8ba1\u70b9\u8bef\u5dee\u5e73\u5747，\u518d\u5e73\u57471\/2\/3\u79d2\u7ed3\u679c，\u533a\u522b\u4e8e\u53ea\u6bd4\u8f83\u5404\u7aef\u70b9\u3002GroundingPI0.296\u7c73\u76f8\u5bf9Qwen3-VL-4B0.301\u7c73，\u5dee0.005\u7c73\u3002\u79bb\u7ebf\u8f68\u8ff9\u66f4\u63a5\u8fd1\u8bb0\u5f55\u4e0d\u4ee3\u8868\u5728\u7ebf\u78b0\u649e\u7387\u4e0b\u964d，\u4e5f\u6ca1\u6709\u6d4b\u51fa\u8bfb\u53d6\u8def\u724c\u7684\u72ec\u7acb\u8d21\u732e\u6216\u95ed\u73af\u5ef6\u8fdf\u3002<\/p>\n

7. \u9884\u8bad\u7ec3\u89c4\u6a21\u4e0e\u4efb\u52a1\u6df7\u5408，\u5206\u522b\u8bf4\u660e\u4e86\u4ec0\u4e48<\/h3>\n

\"\u56fe9\uff1a\u9884\u8bad\u7ec3\u66b4\u9732\u91cf\u4e0e\u8fc1\u79fb\"<\/p>\n

\u56fe9｜\u9884\u8bad\u7ec3\u66b4\u9732\u91cf\u4e0e\u8fc1\u79fb\u3002\u6765\u81eaGroundingPI\u6280\u672f\u62a5\u544a\u3002<\/p>\n

\"\u886816\uff1a\u9010\u9636\u6bb5\u9884\u8bad\u7ec3\u66b4\u9732\u7ed3\u679c\"<\/p>\n

\u886816｜\u9010\u9636\u6bb5\u9884\u8bad\u7ec3\u66b4\u9732\u7ed3\u679c\u3002\u661f\u53f7\u3001\u7f3a\u5931\u9879\u4e0e\u534f\u8bae\u4f8b\u5916\u6cbf\u7528\u8bba\u6587，\u9700\u7ed3\u5408\u6b63\u6587\u89e3\u91ca\u3002<\/p>\n

grounding\u5747\u5206\u4ece61.69\u5347\u81f373.68，OOD\u6210\u529f\u7387\u4ece26.59\u5347\u81f333.38；\u4f46ID\u6210\u529f\u7387\u5728\u6700\u540e\u4e00\u6b65\u4ece38.75\u964d\u81f337.75\u3002\u611f\u77e5\u66f4\u5f3a\u4e0e\u8fc1\u79fb\u6539\u5584\u5b58\u5728\u8054\u7cfb，\u4e0d\u80fd\u6982\u62ec\u4e3a\u6240\u6709\u52a8\u4f5c\u6307\u6807\u90fd\u5355\u8c03\u968f\u89c4\u6a21\u4e0a\u5347\u3002\u56db\u4e2a\u8bbe\u7f6e\u4e14\u7f3a\u5c11\u5145\u5206\u91cd\u590d\u4e0d\u786e\u5b9a\u6027，\u8fd8\u4e0d\u8db3\u4ee5\u5efa\u7acb\u666e\u9002scaling law\u3002<\/p>\n

OOD\u805a\u5408\u6309Container\u3001Appearance\u3001Type\u5404\u8bc4\u4f30\u5957\u4ef6\u6570\u91cf\u52a0\u6743：<\/p>\n

\u8bba\u6587\u5f0f（2）：<\/p>\n

<\/p>\n

S<\/p>\n

R<\/p>\n

O<\/p>\n

O<\/p>\n

D<\/p>\n

=<\/p>\n

14
\n \u2009<\/p>\n

S<\/p>\n

R<\/p>\n

C<\/p>\n

o<\/p>\n

n<\/p>\n

t<\/p>\n

a<\/p>\n

i<\/p>\n

n<\/p>\n

e<\/p>\n

r<\/p>\n

+<\/p>\n

18
\n \u2009<\/p>\n

S<\/p>\n

R<\/p>\n

A<\/p>\n

p<\/p>\n

p<\/p>\n

e<\/p>\n

a<\/p>\n

r<\/p>\n

a<\/p>\n

n<\/p>\n

c<\/p>\n

e<\/p>\n

+<\/p>\n

32
\n \u2009<\/p>\n

S<\/p>\n

R<\/p>\n

T<\/p>\n

y<\/p>\n

p<\/p>\n

e<\/p>\n

64<\/p>\n

.<\/p>\n

\\\\mathrm{SR}_{\\\\mathrm{OOD}}= \\\\frac{14\\\\,\\\\mathrm{SR}_{\\\\mathrm{Container}}+18\\\\,\\\\mathrm{SR}_{\\\\mathrm{Appearance}}+32\\\\,\\\\mathrm{SR}_{\\\\mathrm{Type}}}{64}. <\/p>\n

<\/span><\/span>SR<\/span><\/span><\/span>OOD<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span><\/span><\/span>64<\/span><\/span><\/span><\/span><\/span><\/span><\/span>14<\/span><\/span>SR<\/span><\/span><\/span>Container<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>+<\/span><\/span>18<\/span><\/span>SR<\/span><\/span><\/span>Appearance<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>+<\/span><\/span>32<\/span><\/span>SR<\/span><\/span><\/span>Type<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>.<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

14\u300118\u300132\u662f\u8bc4\u4ef7\u6743\u91cd，\u4e0d\u662f\u8bad\u7ec3\u6df7\u5408\u6bd4\u4f8b\u3002\u5e73\u5747\u5206\u53d8\u5316\u65f6，\u5e94\u540c\u65f6\u67e5\u770b\u6bcf\u4e2a\u5957\u4ef6\u7684\u7ed3\u679c\u3002<\/p>\n

\"\u56fe10\uff1a\u516d\u7ec4\u4efb\u52a1\u7ec4\u6210\u6d88\u878d\"<\/p>\n

\u56fe10｜\u516d\u7ec4\u4efb\u52a1\u7ec4\u6210\u6d88\u878d\u3002\u6765\u81eaGroundingPI\u6280\u672f\u62a5\u544a\u3002<\/p>\n

\"\u886817\uff1a\u516d\u7ec4\u4efb\u52a1\u6df7\u5408\"<\/p>\n

\u886817｜\u516d\u7ec4\u4efb\u52a1\u6df7\u5408\u3002\u661f\u53f7\u3001\u7f3a\u5931\u9879\u4e0e\u534f\u8bae\u4f8b\u5916\u6cbf\u7528\u8bba\u6587，\u9700\u7ed3\u5408\u6b63\u6587\u89e3\u91ca\u3002<\/p>\n

\u516d\u7ec4\u5206\u522b\u4e3a\u57fa\u7840grounding\u3001\u5bc6\u96c6grounding\u3001\u6307\u4ee3\u3001\u57fa\u7840pointing\u3001\u673a\u5668\u4ebapointing，\u4ee5\u53caOCR\/\u5e03\u5c40\/GUI\u8f85\u52a9\u7ec4\u3002\u53bb\u6389\u5bc6\u96c6\u7ec4，ID\u4e0eOOD\u5206\u522b\u4e0b\u964d4.67\u548c3.47\u4e2a\u767e\u5206\u70b9，\u9a7e\u9a76L2\u589e\u52a00.015\u7c73，\u652f\u6301\u5bc6\u96c6\u7a7a\u95f4\u76d1\u7763\u7684\u5e7f\u6cdb\u8fc1\u79fb\u4ef7\u503c\u3002<\/p>\n

\u5176\u51e0\u4f55\u52a8\u673a\u5f88\u76f4\u89c2：\u7b49\u5bbd <\/p>\n

w<\/p>\n

w<\/p>\n

<\/span><\/span>w<\/span><\/span><\/span><\/span><\/span>\u3001\u7b49\u9ad8\u7684\u4e24\u4e2a\u6846\u53ea\u5728\u6c34\u5e73\u65b9\u5411\u9519\u5f00 <\/p>\n

\u03b4<\/p>\n

\\\\delta<\/p>\n

<\/span><\/span>\u03b4<\/span><\/span><\/span><\/span><\/span>，\u4e14 <\/p>\n

\u2223<\/p>\n

\u03b4<\/p>\n

\u2223<\/p>\n

<<\/p>\n

w<\/p>\n

|\\\\delta|<w<\/p>\n

<\/span><\/span>\u2223<\/span>\u03b4<\/span>\u2223<\/span><\/span><<\/span><\/span><\/span><\/span>w<\/span><\/span><\/span><\/span><\/span> \u65f6：<\/p>\n

\u8bba\u6587\u5f0f（3）：<\/p>\n

<\/p>\n

IoU<\/p>\n

\u2061<\/p>\n

(<\/p>\n

\u03b4<\/p>\n

)<\/p>\n

=<\/p>\n

w<\/p>\n

\u2212<\/p>\n

\u2223<\/p>\n

\u03b4<\/p>\n

\u2223<\/p>\n

w<\/p>\n

+<\/p>\n

\u2223<\/p>\n

\u03b4<\/p>\n

\u2223<\/p>\n

,<\/p>\n

IoU<\/p>\n

\u2061<\/p>\n

(<\/p>\n

\u03b4<\/p>\n

)<\/p>\n

\u2265<\/p>\n

\u03c4
\n \u2005\u200a<\/p>\n

\u27fa
\n \u2005\u200a<\/p>\n

\u2223<\/p>\n

\u03b4<\/p>\n

\u2223<\/p>\n

\u2264<\/p>\n

w<\/p>\n

1<\/p>\n

\u2212<\/p>\n

\u03c4<\/p>\n

1<\/p>\n

+<\/p>\n

\u03c4<\/p>\n

.<\/p>\n

\\\\operatorname{IoU}(\\\\delta)=\\\\frac{w-|\\\\delta|}{w+|\\\\delta|}, \\\\qquad \\\\operatorname{IoU}(\\\\delta)\\\\geq\\\\tau \\\\;\\\\Longleftrightarrow\\\\; |\\\\delta|\\\\leq w\\\\frac{1-\\\\tau}{1+\\\\tau}. <\/p>\n

<\/span><\/span>IoU<\/span><\/span>(<\/span>\u03b4<\/span>)<\/span><\/span>=<\/span><\/span><\/span><\/span><\/span><\/span>w<\/span><\/span>+<\/span><\/span>\u2223<\/span>\u03b4<\/span>\u2223<\/span><\/span><\/span><\/span><\/span><\/span><\/span>w<\/span><\/span>\u2212<\/span><\/span>\u2223<\/span>\u03b4<\/span>\u2223<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>,<\/span><\/span><\/span>IoU<\/span><\/span>(<\/span>\u03b4<\/span>)<\/span><\/span>\u2265<\/span><\/span><\/span><\/span>\u03c4<\/span><\/span><\/span>\u27fa<\/span><\/span><\/span><\/span><\/span>\u2223<\/span>\u03b4<\/span>\u2223<\/span><\/span>\u2264<\/span><\/span><\/span><\/span>w<\/span><\/span><\/span>1<\/span><\/span>+<\/span><\/span>\u03c4<\/span><\/span><\/span><\/span><\/span><\/span><\/span>1<\/span><\/span>\u2212<\/span><\/span>\u03c4<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>.<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

\u5047\u8bbe\u9608\u503c0.9\u3001\u76ee\u6807\u5bbd10\u50cf\u7d20，\u5141\u8bb8\u4f4d\u79fb\u4ec5\u7ea60.526\u50cf\u7d20\u3002\u8fd9\u4e2a\u8ba1\u7b97\u53ea\u662f\u7b49\u5c3a\u5bf8\u6c34\u5e73\u5e73\u79fb\u7684\u7b80\u5316\u6a21\u578b，\u5374\u6e05\u695a\u8bf4\u660e\u5c0f\u76ee\u6807\u5bf9\u7edd\u5bf9\u8bef\u5dee\u66f4\u654f\u611f\u3002<\/p>\n

\u53bb\u6389OCR\/\u5e03\u5c40\/GUI\u8f85\u52a9\u7ec4，\u673a\u5668\u4ebaID\/OOD\u4e0b\u964d5.33\/3.63\u4e2a\u767e\u5206\u70b9；\u4f46\u53ea\u4fdd\u7559\u8be5\u7ec4\u53c8\u6700\u5dee，\u8bf4\u660e\u5b83\u4e0e\u5176\u4ed6\u4efb\u52a1\u4e92\u8865\u3002\u8fd9\u4e2a\u7ec4\u5305\u542b\u4e09\u7c7b\u76d1\u7763，\u4e0d\u80fd\u628a\u6574\u4f53\u5220\u7ec4\u6548\u5e94\u5168\u90e8\u5f52\u56e0\u4e8eOCR，\u4e5f\u4e0d\u80fd\u628a\u5404\u4e2a\u7559\u4e00\u6d88\u878d\u5dee\u503c\u76f8\u52a0\u5f53\u4f5c\u72ec\u7acb\u8d21\u732e\u3002<\/p>\n

7.1 OCR\u201c\u50ac\u5316\u611f\u77e5\u201d\u4ecd\u7136\u662f\u5f85\u9a8c\u8bc1\u673a\u5236<\/h4>\n

\"\u886818\uff1aOCR\u6df7\u5408\u65f6\u5e8f\"<\/p>\n

\u886818｜OCR\u6df7\u5408\u65f6\u5e8f\u3002\u661f\u53f7\u3001\u7f3a\u5931\u9879\u4e0e\u534f\u8bae\u4f8b\u5916\u6cbf\u7528\u8bba\u6587，\u9700\u7ed3\u5408\u6b63\u6587\u89e3\u91ca\u3002<\/p>\n

\u51b7\u542f\u52a8\u52a0\u5165OCR\u4e0e\u8bad\u7ec3\u6279\u6b21\u6301\u7eed\u6df7\u5408\u7ed3\u5408\u65f6，\u673a\u5668\u4eba\u8fc1\u79fb\u6700\u597d；\u9a7e\u9a76L2\u5728\u663e\u793a\u7cbe\u5ea6\u4e0b\u90fd\u4e3a0.296\u3002OCR\u662f\u5426\u901a\u8fc7\u6587\u5b57\u8f6c\u5199\u68af\u5ea6\u6539\u5584\u975e\u6587\u5b57\u7a7a\u95f4\u8bc6\u522b，\u4ecd\u9700\u66f4\u7ec6\u5bf9\u7167\u3002<\/p>\n

\u628aOCR\u54cd\u5e94\u4f4d\u7f6e\u62c6\u6210\u6846\u3001\u8f6c\u5199\u548c\u534f\u8bae\u4e09\u7c7b，\u635f\u5931\u53ef\u5199\u4e3a：<\/p>\n

\u8bba\u6587\u5f0f（4）：<\/p>\n

<\/p>\n

L<\/p>\n

O<\/p>\n

C<\/p>\n

R<\/p>\n

=<\/p>\n

L<\/p>\n

b<\/p>\n

+<\/p>\n

L<\/p>\n

t<\/p>\n

+<\/p>\n

L<\/p>\n

p<\/p>\n

,<\/p>\n

L<\/p>\n

r<\/p>\n

=<\/p>\n

\u2212<\/p>\n

1<\/p>\n

\u2223<\/p>\n

T<\/p>\n

\u2223<\/p>\n

\u2211<\/p>\n

i<\/p>\n

\u2208<\/p>\n

T<\/p>\n

r<\/p>\n

log<\/p>\n

\u2061<\/p>\n

p<\/p>\n

\u03b8<\/p>\n

(<\/p>\n

y<\/p>\n

i<\/p>\n

\u2223<\/p>\n

I<\/p>\n

,<\/p>\n

Q<\/p>\n

,<\/p>\n

y<\/p>\n

<<\/p>\n

i<\/p>\n

)<\/p>\n

,<\/p>\n

r<\/p>\n

\u2208<\/p>\n

{<\/p>\n

b<\/p>\n

,<\/p>\n

t<\/p>\n

,<\/p>\n

p<\/p>\n

}<\/p>\n

.<\/p>\n

\\\\mathcal L_{\\\\mathrm{OCR}}=\\\\mathcal L_b+\\\\mathcal L_t+\\\\mathcal L_p, \\\\qquad \\\\mathcal L_r=-\\\\frac{1}{|\\\\mathcal T|}\\\\sum_{i\\\\in\\\\mathcal T_r} \\\\log p_\\\\theta(y_i\\\\mid I,Q,y_{<i}),\\\\quad r\\\\in\\\\{b,t,p\\\\}. <\/p>\n

<\/span><\/span>L<\/span><\/span>OCR<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>L<\/span><\/span>b<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>+<\/span><\/span><\/span><\/span>L<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>+<\/span><\/span><\/span><\/span>L<\/span><\/span>p<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>,<\/span><\/span><\/span>L<\/span><\/span>r<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>\u2212<\/span><\/span><\/span>\u2223<\/span>T<\/span>\u2223<\/span><\/span><\/span><\/span><\/span><\/span><\/span>1<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>i<\/span>\u2208<\/span>T<\/span><\/span>r<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u2211<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>log<\/span><\/span><\/span>p<\/span><\/span>\u03b8<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>y<\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u2223<\/span><\/span><\/span><\/span>I<\/span>,<\/span><\/span>Q<\/span>,<\/span><\/span>y<\/span><\/span><<\/span>i<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>)<\/span>,<\/span><\/span><\/span>r<\/span><\/span>\u2208<\/span><\/span><\/span><\/span>{<\/span>b<\/span>,<\/span><\/span>t<\/span>,<\/span><\/span>p<\/span>}<\/span>.<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

\u8981\u9694\u79bb\u8f6c\u5199\u68af\u5ea6，\u5e94\u79fb\u9664 <\/p>\n

L<\/p>\n

t<\/p>\n

\\\\mathcal L_t<\/p>\n

<\/span><\/span>L<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>，\u4f46\u4fdd\u7559\u8f6c\u5199token\u4f5c\u4e3ateacher-forcing\u6761\u4ef6，\u5e76\u4fdd\u6301\u603b\u5206\u6bcd\u3002\u5220\u6389\u8f6c\u5199token\u4f1a\u6539\u53d8\u6846\u9884\u6d4b\u6761\u4ef6，\u91cd\u65b0\u5f52\u4e00\u5316\u5269\u4f59token\u53c8\u4f1a\u6539\u53d8\u5176\u4ed6\u635f\u5931\u6743\u91cd，\u90fd\u4e0d\u518d\u662f\u5355\u7eaf\u7684\u201c\u79fb\u9664\u8f6c\u5199\u76d1\u7763\u201d\u3002<\/p>\n

\u5bf9\u4e8e\u89c6\u89c9\u53c2\u6570 <\/p>\n

\u03b8<\/p>\n

v<\/p>\n

\\\\theta_v<\/p>\n

<\/span><\/span>\u03b8<\/span><\/span>v<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>，\u4ee4\u8f6c\u5199\u68af\u5ea6\u4e3a <\/p>\n

g<\/p>\n

t<\/p>\n

g_t<\/p>\n

<\/span><\/span>g<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>，\u975e\u6587\u5b57grounding\u68af\u5ea6\u4e3a <\/p>\n

g<\/p>\n

g<\/p>\n

g_g<\/p>\n

<\/span><\/span>g<\/span><\/span>g<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u3002\u5728\u5c40\u90e8Hessian\u8303\u6570\u6709\u754c\u6761\u4ef6\u4e0b：<\/p>\n

\u8bba\u6587\u5f0f（5）：<\/p>\n

<\/p>\n

L<\/p>\n

g<\/p>\n

(<\/p>\n

\u03b8<\/p>\n

v<\/p>\n

\u2212<\/p>\n

\u03b7<\/p>\n

g<\/p>\n

t<\/p>\n

)<\/p>\n

\u2212<\/p>\n

L<\/p>\n

g<\/p>\n

(<\/p>\n

\u03b8<\/p>\n

v<\/p>\n

)<\/p>\n

\u2264<\/p>\n

\u2212<\/p>\n

\u03b7<\/p>\n

\u27e8<\/p>\n

g<\/p>\n

g<\/p>\n

,<\/p>\n

g<\/p>\n

t<\/p>\n

\u27e9<\/p>\n

+<\/p>\n

H<\/p>\n

\u03b7<\/p>\n

2<\/p>\n

2<\/p>\n

\u2225<\/p>\n

g<\/p>\n

t<\/p>\n

\u2225<\/p>\n

2<\/p>\n

2<\/p>\n

.<\/p>\n

\\\\mathcal L_g(\\\\theta_v-\\\\eta g_t)-\\\\mathcal L_g(\\\\theta_v) \\\\leq-\\\\eta\\\\langle g_g,g_t\\\\rangle +\\\\frac{H\\\\eta^2}{2}\\\\lVert g_t\\\\rVert_2^2. <\/p>\n

<\/span><\/span>L<\/span><\/span>g<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>\u03b8<\/span><\/span>v<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u2212<\/span><\/span><\/span><\/span>\u03b7<\/span>g<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>)<\/span><\/span>\u2212<\/span><\/span><\/span><\/span>L<\/span><\/span>g<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>\u03b8<\/span><\/span>v<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>)<\/span><\/span>\u2264<\/span><\/span><\/span><\/span>\u2212<\/span>\u03b7<\/span>\u27e8<\/span>g<\/span><\/span>g<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>,<\/span><\/span>g<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u27e9<\/span><\/span>+<\/span><\/span><\/span><\/span><\/span><\/span>2<\/span><\/span><\/span><\/span><\/span><\/span><\/span>H<\/span>\u03b7<\/span><\/span>2<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u2225<\/span>g<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u2225<\/span><\/span>2<\/span><\/span><\/span><\/span>2<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>.<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

\u5f53\u4e24\u79cd\u68af\u5ea6\u65b9\u5411\u4e00\u81f4\u4e14\u5b66\u4e60\u7387\u8db3\u591f\u5c0f\u65f6，\u8fd9\u4e2a\u4e0a\u754c\u5141\u8bb8\u975e\u6587\u5b57\u635f\u5931\u4e0b\u964d\u3002\u5b83\u662f\u6761\u4ef6\u6027\u7684\u5c40\u90e8\u5206\u6790，\u4e0d\u662f\u5bf9\u771f\u5b9e\u8bad\u7ec3\u5168\u8fc7\u7a0b\u7684\u8bc1\u660e；AdamW\u8fd8\u5e94\u68c0\u67e5\u5b9e\u9645\u53c2\u6570\u66f4\u65b0\u65b9\u5411\u3002\u8bba\u6587\u63d0\u51fa\u4e86\u53ef\u68c0\u9a8c\u5047\u8bbe，\u5b8c\u6574\u673a\u5236\u9a8c\u8bc1\u4ecd\u672a\u5b8c\u6210\u3002<\/p>\n

8. \u8f93\u51fa\u66f4\u77ed，\u79bb\u5b9e\u65f6\u52a8\u4f5c\u8fd8\u5dee\u4e00\u6b21\u7aef\u5230\u7aef\u6d4b\u91cf<\/h3>\n

\"\u56fe11\uff1a\u5750\u6807\u3001\u89c6\u89c9\u7f16\u7801\u5668\u4e0eRL\u6d88\u878d\"<\/p>\n

\u56fe11｜\u5750\u6807\u3001\u89c6\u89c9\u7f16\u7801\u5668\u4e0eRL\u6d88\u878d\u3002\u6765\u81eaGroundingPI\u6280\u672f\u62a5\u544a\u3002<\/p>\n

\u91cf\u5316\u5750\u6807\u76f8\u5bf9\u6587\u672c\u5750\u6807\u4f7fgrounding\u5747\u5206\u4ece71.21\u63d0\u9ad8\u523073.68；\u66ff\u6362\u89c6\u89c9\u7f16\u7801\u5668\u6216\u79fb\u9664\u5f3a\u5316\u540e\u8bad\u7ec3\u4e5f\u4f1a\u964d\u4f4e\u8be5\u914d\u7f6e\u4e0b\u7684\u8868\u73b0\u3002\u5176\u4e2d\u53bb\u6389RL\u4e3a72.86，\u5bf9\u5e940.82\u4e2a\u767e\u5206\u70b9\u5dee\u5f02\u3002<\/p>\n

\"\u88685\uff1a\u5750\u6807\u5e8f\u5217\u5f00\u9500\"<\/p>\n

\u88685｜\u5750\u6807\u5e8f\u5217\u5f00\u9500\u3002\u661f\u53f7\u3001\u7f3a\u5931\u9879\u4e0e\u534f\u8bae\u4f8b\u5916\u6cbf\u7528\u8bba\u6587，\u9700\u7ed3\u5408\u6b63\u6587\u89e3\u91ca\u3002<\/p>\n

COCO\u4e0a\u5e73\u5747\u6bcf\u68467.6token，\u5bf9\u7167\u5916\u90e8SEED1.5-VL\u7edf\u8ba1148.8；Dense200\u4e3a5.1\u4e0e74.5\u3002\u8fd9\u5bf9\u5e94\u7ea619.6\u500d\u548c14.6\u500d\u7684\u5e8f\u5217\u538b\u7f29，\u4e0d\u662f\u76f8\u540c\u786c\u4ef6\u3001\u76f8\u540c\u8f93\u5165\u3001\u76f8\u540c\u89e3\u7801\u5b9e\u73b0\u4e0b\u7684\u63a8\u7406\u52a0\u901f\u500d\u6570\u3002<\/p>\n

\"\u56fe12\uff1a\u5b9e\u4f8b\u6570\u91cf\u4e0e\u751f\u6210\u6210\u672c\"<\/p>\n

\u56fe12｜\u5b9e\u4f8b\u6570\u91cf\u4e0e\u751f\u6210\u6210\u672c\u3002\u6765\u81eaGroundingPI\u6280\u672f\u62a5\u544a\u3002<\/p>\n

\u540c\u7c7b\u591a\u4e2a\u6846\u5171\u4eab\u6807\u7b7e，\u76ee\u6807\u591a\u65f6\u53ef\u4ee5\u644a\u8584\u5305\u88c5\u5f00\u9500；\u4f46\u603b\u8f93\u51fa\u4ecd\u968f\u76ee\u6807\u6570\u589e\u52a0\u3002\u5b9e\u9645\u7cfb\u7edf\u9700\u8981\u5206\u5f00\u6d4b\u91cf\u89c6\u89c9\u524d\u5411\u3001prefill\u3001\u81ea\u56de\u5f52\u8f93\u51fa\u3001\u89e3\u6790\u3001\u5750\u6807\u8f6c\u6362\u4e0e\u52a8\u4f5c\u6267\u884c，\u8bb0\u5f55\u5c3e\u5ef6\u8fdf，\u800c\u4e0d\u4ec5\u662f\u5e73\u5747token\u901f\u5ea6\u3002<\/p>\n

\u5982\u679c\u4e0b\u6e38\u53ea\u9700\u8981\u5f53\u524d\u64cd\u4f5c\u5bf9\u8c61，\u76f4\u63a5\u8bf7\u6c42\u4e0e\u76ee\u6807\u76f8\u5173\u7684\u533a\u57df\u53ef\u80fd\u6bd4\u6bcf\u5e27\u8f93\u51fa\u6240\u6709\u7269\u4f53\u66f4\u5408\u9002；\u82e5\u4efb\u52a1\u8981\u6c42\u76d8\u70b9\u5168\u90e8\u5b9e\u4f8b，\u5219\u4e0d\u80fd\u4ee5\u5c11\u8f93\u51fa\u6362\u53d6\u8868\u9762\u901f\u5ea6\u3002\u8fd9\u5c5e\u4e8e\u4efb\u52a1\u534f\u8bae\u4e0e\u670d\u52a1\u9884\u7b97\u7684\u5171\u540c\u8bbe\u8ba1\u3002<\/p>\n

9. \u4e3a\u4ec0\u4e48\u67b6\u6784\u8ba8\u8bba\u4e0d\u80fd\u88ab\u7b80\u5316\u6210\u201c\u7ebf\u6027\u6ce8\u610f\u529b\u4e0d\u9002\u5408\u673a\u5668\u4eba\u201d<\/h3>\n

\u8bba\u6587\u6bd4\u8f83\u7684\u57fa\u7840\u6a21\u578b\u5728\u89c6\u89c9\u7f16\u7801\u5668\u3001\u8bed\u8a00\u67b6\u6784\u3001\u9884\u8bad\u7ec3\u3001\u8fde\u63a5\u5668\u548c\u89c4\u6a21\u4e0a\u90fd\u4e0d\u540c\u3002RynnBrain1.1\u7684Qwen3.5\u57fa\u7840\u5305\u542b\u7ebf\u6027\u6ce8\u610f\u529b\u4e0e\u95e8\u63a7，RynnBrain\u4f7f\u7528\u5168\u6ce8\u610f\u529bQwen3-VL；\u524d\u8005\u673a\u5668\u4eba\u7ed3\u679c\u66f4\u5f31\u3001\u9a7e\u9a76\u7ed3\u679c\u5374\u66f4\u597d\u3002\u8fd9\u4e00\u73b0\u8c61\u63d0\u51fa\u7814\u7a76\u95ee\u9898，\u4f46\u6ca1\u6709\u9694\u79bb\u67d0\u79cd\u6ce8\u610f\u529b\u7ed3\u6784\u7684\u56e0\u679c\u4f5c\u7528\u3002<\/p>\n

9.1 \u884c\u4e3a\u635f\u5931\u80fd\u5426\u8bbf\u95ee\u5230\u7a7a\u95f4\u7279\u5f81<\/h4>\n

\u628a\u516b\u4e2a\u6df1\u5ea6\u7684\u52a8\u4f5c\u63a5\u53e3\u5199\u4e3a：<\/p>\n

\u8bba\u6587\u5f0f（6）：<\/p>\n

<\/p>\n

C<\/p>\n

s<\/p>\n

=<\/p>\n

R
\n \u2009\u2063<\/p>\n

(<\/p>\n

LN<\/p>\n

\u2061<\/p>\n

(<\/p>\n

H<\/p>\n

\u2113<\/p>\n

s<\/p>\n

)<\/p>\n

W<\/p>\n

B<\/p>\n

+<\/p>\n

e<\/p>\n

s<\/p>\n

)<\/p>\n

,<\/p>\n

s<\/p>\n

=<\/p>\n

0<\/p>\n

,<\/p>\n

\u2026<\/p>\n

,<\/p>\n

7<\/p>\n

,<\/p>\n

y<\/p>\n

^<\/p>\n

=<\/p>\n

A<\/p>\n

\u03d5<\/p>\n

(<\/p>\n

\u03be<\/p>\n

;<\/p>\n

C<\/p>\n

0<\/p>\n

,<\/p>\n

\u2026<\/p>\n

,<\/p>\n

C<\/p>\n

7<\/p>\n

)<\/p>\n

.<\/p>\n

C_s=\\\\mathcal R\\\\!\\\\left(\\\\operatorname{LN}(H_{\\\\ell_s})W_B+e_s\\\\right),\\\\quad s=0,\\\\ldots,7, \\\\qquad \\\\hat y=\\\\mathcal A_\\\\phi(\\\\xi;C_0,\\\\ldots,C_7). <\/p>\n

<\/span><\/span>C<\/span><\/span>s<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>R<\/span><\/span><\/span>(<\/span>LN<\/span><\/span>(<\/span>H<\/span><\/span>\u2113<\/span><\/span>s<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>)<\/span>W<\/span><\/span>B<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>+<\/span><\/span>e<\/span><\/span>s<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>)<\/span><\/span><\/span>,<\/span><\/span><\/span>s<\/span><\/span>=<\/span><\/span><\/span><\/span>0<\/span>,<\/span><\/span>\u2026<\/span><\/span>,<\/span><\/span>7<\/span>,<\/span><\/span><\/span><\/span>y<\/span><\/span><\/span>^<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>A<\/span><\/span>\u03d5<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>\u03be<\/span>;<\/span><\/span>C<\/span><\/span>0<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>,<\/span><\/span>\u2026<\/span><\/span>,<\/span><\/span>C<\/span><\/span>7<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>)<\/span>.<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

<\/p>\n

R<\/p>\n

\\\\mathcal R<\/p>\n

<\/span><\/span>R<\/span><\/span><\/span><\/span><\/span>\u4e3a\u91cd\u91c7\u6837\u5668，<\/p>\n

W<\/p>\n

B<\/p>\n

W_B<\/p>\n

<\/span><\/span>W<\/span><\/span>B<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u4e3a\u5171\u4eab\u6295\u5f71，<\/p>\n

e<\/p>\n

s<\/p>\n

e_s<\/p>\n

<\/span><\/span>e<\/span><\/span>s<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u4e3a\u6df1\u5ea6\u5d4c\u5165，<\/p>\n

A<\/p>\n

\u03d5<\/p>\n

\\\\mathcal A_\\\\phi<\/p>\n

<\/span><\/span>A<\/span><\/span>\u03d5<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u662f\u9884\u6d4bflow\u901f\u5ea6\u7684\u52a8\u4f5c\u6a21\u578b\u3002\u4e00\u4e2a\u89c6\u89c9\u4fe1\u606f\u5373\u4f7f\u4fdd\u7559\u5728\u4e3b\u5e72\u91cc，\u4e5f\u4e0d\u4ee3\u8868\u52a8\u4f5c\u63a5\u53e3\u80fd\u5728\u6709\u9650\u8bad\u7ec3\u9884\u7b97\u5185\u6709\u6548\u8bfb\u51fa；\u5173\u952e\u662f\u6574\u6761\u7279\u5f81\u5230\u52a8\u4f5c\u7684\u8def\u5f84\u3002<\/p>\n

\u5bf9\u4e8e\u4e00\u4e2aGated DeltaNet\u5934，\u8bba\u6587\u7528\u5982\u4e0b\u72b6\u6001\u66f4\u65b0\u8ba8\u8bba\u6761\u4ef6\u6027\u4f20\u9012：<\/p>\n

\u8bba\u6587\u5f0f（7）：<\/p>\n

<\/p>\n

S<\/p>\n

t<\/p>\n

=<\/p>\n

S<\/p>\n

t<\/p>\n

\u2212<\/p>\n

1<\/p>\n

A<\/p>\n

t<\/p>\n

+<\/p>\n

\u03b2<\/p>\n

t<\/p>\n

v<\/p>\n

t<\/p>\n

k<\/p>\n

t<\/p>\n

\u22a4<\/p>\n

,<\/p>\n

A<\/p>\n

t<\/p>\n

=<\/p>\n

\u03b1<\/p>\n

t<\/p>\n

(<\/p>\n

I<\/p>\n

\u2212<\/p>\n

\u03b2<\/p>\n

t<\/p>\n

k<\/p>\n

t<\/p>\n

k<\/p>\n

t<\/p>\n

\u22a4<\/p>\n

)<\/p>\n

,<\/p>\n

o<\/p>\n

t<\/p>\n

=<\/p>\n

S<\/p>\n

t<\/p>\n

q<\/p>\n

t<\/p>\n

.<\/p>\n

S_t=S_{t-1}A_t+\\\\beta_t v_tk_t^\\\\top, \\\\qquad A_t=\\\\alpha_t(I-\\\\beta_t k_tk_t^\\\\top), \\\\qquad o_t=S_tq_t. <\/p>\n

<\/span><\/span>S<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>S<\/span><\/span>t<\/span>\u2212<\/span>1<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>A<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>+<\/span><\/span><\/span><\/span>\u03b2<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>v<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>k<\/span><\/span>t<\/span><\/span><\/span><\/span>\u22a4<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>,<\/span><\/span><\/span>A<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>\u03b1<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>I<\/span><\/span>\u2212<\/span><\/span><\/span><\/span>\u03b2<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>k<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>k<\/span><\/span>t<\/span><\/span><\/span><\/span>\u22a4<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>)<\/span>,<\/span><\/span><\/span>o<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>S<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>q<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>.<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

\u56fa\u5b9akey\u3001query\u3001gate\u4e0e\u5176\u4ed6value，\u4ec5\u6270\u52a8 <\/p>\n

v<\/p>\n

j<\/p>\n

v_j<\/p>\n

<\/span><\/span>v<\/span><\/span>j<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>，\u5728\u952e\u5f52\u4e00\u5316\u3001\u95e8\u503c\u6709\u754c\u7684\u5047\u8bbe\u4e0b\u6709：<\/p>\n

\u8bba\u6587\u5f0f（8\u30019）：<\/p>\n

<\/p>\n

\u0394<\/p>\n

o<\/p>\n

t<\/p>\n

=<\/p>\n

\u03ba<\/p>\n

j<\/p>\n

t<\/p>\n

\u0394<\/p>\n

v<\/p>\n

j<\/p>\n

,<\/p>\n

\u03ba<\/p>\n

j<\/p>\n

t<\/p>\n

=<\/p>\n

\u03b2<\/p>\n

j<\/p>\n

k<\/p>\n

j<\/p>\n

\u22a4<\/p>\n

A<\/p>\n

j<\/p>\n

+<\/p>\n

1<\/p>\n

\u22ef<\/p>\n

A<\/p>\n

t<\/p>\n

q<\/p>\n

t<\/p>\n

,<\/p>\n

t<\/p>\n

\u2265<\/p>\n

j<\/p>\n

,<\/p>\n

\u2223<\/p>\n

\u03ba<\/p>\n

j<\/p>\n

t<\/p>\n

\u2223<\/p>\n

\u2264<\/p>\n

\u03b2<\/p>\n

j<\/p>\n

Q<\/p>\n

\u220f<\/p>\n

r<\/p>\n

=<\/p>\n

j<\/p>\n

+<\/p>\n

1<\/p>\n

t<\/p>\n

\u03b1<\/p>\n

r<\/p>\n

.<\/p>\n

\\\\begin{aligned} \\\\Delta o_t&=\\\\kappa_{jt}\\\\Delta v_j, \\\\qquad \\\\kappa_{jt}=\\\\beta_j k_j^\\\\top A_{j+1}\\\\cdots A_tq_t,\\\\quad t\\\\geq j, \\\\\\\\ |\\\\kappa_{jt}|&\\\\leq\\\\beta_j Q\\\\prod_{r=j+1}^{t}\\\\alpha_r. \\\\end{aligned} <\/p>\n

<\/span><\/span><\/span>\u0394<\/span>o<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u2223<\/span>\u03ba<\/span><\/span>j<\/span>t<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u2223<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span>\u03ba<\/span><\/span>j<\/span>t<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u0394<\/span>v<\/span><\/span>j<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>,<\/span><\/span><\/span>\u03ba<\/span><\/span>j<\/span>t<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span>\u03b2<\/span><\/span>j<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>k<\/span><\/span>j<\/span><\/span><\/span><\/span>\u22a4<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>A<\/span><\/span>j<\/span>+<\/span>1<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u22ef<\/span><\/span>A<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>q<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>,<\/span><\/span><\/span>t<\/span><\/span>\u2265<\/span><\/span>j<\/span>,<\/span><\/span><\/span><\/span><\/span><\/span>\u2264<\/span><\/span>\u03b2<\/span><\/span>j<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>Q<\/span><\/span><\/span>r<\/span>=<\/span>j<\/span>+<\/span>1<\/span><\/span><\/span><\/span><\/span>\u220f<\/span><\/span><\/span><\/span>t<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u03b1<\/span><\/span>r<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>.<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

\u5f53\u6cbf\u9014\u9057\u5fd8\u95e8\u6709\u7edf\u4e00\u5c0f\u4e8e1\u7684\u4e0a\u754c\u65f6，\u8fd9\u6761value\u901a\u8def\u7684\u4e0a\u754c\u53ef\u968f\u5e8f\u5217\u8ddd\u79bb\u8870\u51cf\u3002\u5e8f\u5217\u8ddd\u79bb\u5e76\u4e0d\u662f\u56fe\u50cf\u7269\u7406\u8ddd\u79bb，\u4e5f\u4e0d\u4ee3\u8868\u6574\u4e2a\u6a21\u578b\u7684\u5168\u90e8\u4fe1\u606f\u901a\u8def；\u6b8b\u5dee\u4e0e\u5168\u6ce8\u610f\u529b\u5c42\u4ecd\u53ef\u80fd\u4f20\u9012\u76f8\u5173\u4fe1\u606f\u3002<\/p>\n

\u8003\u8651\u6240\u6709\u6df1\u5ea6\u8bfb\u51fa\u3001\u6295\u5f71\u3001\u91cd\u91c7\u6837\u548c\u52a8\u4f5c\u6a21\u578b，\u6761\u4ef6value\u901a\u8def\u7684Jacobian\u4e3a：<\/p>\n

\u8bba\u6587\u5f0f（10）：<\/p>\n

<\/p>\n

J<\/p>\n

j<\/p>\n

r<\/p>\n

e<\/p>\n

c<\/p>\n

=<\/p>\n

\u2211<\/p>\n

t<\/p>\n

=<\/p>\n

j<\/p>\n

T<\/p>\n

\u03ba<\/p>\n

j<\/p>\n

t<\/p>\n

R<\/p>\n

t<\/p>\n

,<\/p>\n

\u2207<\/p>\n

v<\/p>\n

j<\/p>\n

r<\/p>\n

e<\/p>\n

c<\/p>\n

L<\/p>\n

a<\/p>\n

c<\/p>\n

t<\/p>\n

=<\/p>\n

(<\/p>\n

J<\/p>\n

j<\/p>\n

r<\/p>\n

e<\/p>\n

c<\/p>\n

)<\/p>\n

\u22a4<\/p>\n

\u2207<\/p>\n

y<\/p>\n

^<\/p>\n

L<\/p>\n

a<\/p>\n

c<\/p>\n

t<\/p>\n

.<\/p>\n

J_j^{\\\\mathrm{rec}}=\\\\sum_{t=j}^{T}\\\\kappa_{jt}R_t, \\\\qquad \\\\nabla_{v_j}^{\\\\mathrm{rec}}\\\\mathcal L_{\\\\mathrm{act}} =(J_j^{\\\\mathrm{rec}})^\\\\top\\\\nabla_{\\\\hat y}\\\\mathcal L_{\\\\mathrm{act}}. <\/p>\n

<\/span><\/span>J<\/span><\/span>j<\/span><\/span><\/span><\/span>rec<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span><\/span>t<\/span>=<\/span>j<\/span><\/span><\/span><\/span><\/span>\u2211<\/span><\/span><\/span><\/span>T<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u03ba<\/span><\/span>j<\/span>t<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>R<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>,<\/span><\/span><\/span>\u2207<\/span><\/span>v<\/span><\/span>j<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>rec<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>L<\/span><\/span>act<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>(<\/span>J<\/span><\/span>j<\/span><\/span><\/span><\/span>rec<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>)<\/span><\/span>\u22a4<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u2207<\/span><\/span><\/span>y<\/span><\/span><\/span>^<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>L<\/span><\/span>act<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>.<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

\u5176\u4e2d <\/p>\n

R<\/p>\n

t<\/p>\n

R_t<\/p>\n

<\/span><\/span>R<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span> \u5305\u542b\u540e\u7eed\u7f51\u7edc\u5bf9\u4f4d\u7f6e <\/p>\n

t<\/p>\n

t<\/p>\n

<\/span><\/span>t<\/span><\/span><\/span><\/span><\/span> \u7684\u52a8\u4f5c\u654f\u611f\u5ea6\u3002<\/p>\n

t<\/p>\n

=<\/p>\n

j<\/p>\n

t=j<\/p>\n

<\/span><\/span>t<\/span><\/span>=<\/span><\/span><\/span><\/span>j<\/span><\/span><\/span><\/span><\/span> \u7684\u9879\u4ecd\u5141\u8bb8\u8bbf\u95ee\u5f53\u524dtoken\u81ea\u5df1\u7684\u72b6\u6001，\u540e\u7eed\u901a\u8def\u4e5f\u53ef\u80fd\u653e\u5927\u6216\u62b5\u6d88\u5f71\u54cd，\u56e0\u6b64\u4e0d\u80fd\u4ec5\u51ed\u67d0\u4e00\u9057\u5fd8\u7cfb\u6570\u5c31\u5224\u65ad\u52a8\u4f5c\u5931\u8d25\u3002<\/p>\n

9.2 \u95e8\u63a7\u4e0e\u591a\u5c42\u89c6\u89c9\u6ce8\u5165\u9700\u8981\u53d7\u63a7\u5b9e\u9a8c<\/h4>\n

\u5bf9 <\/p>\n

b<\/p>\n

=<\/p>\n

W<\/p>\n

o<\/p>\n

(<\/p>\n

g<\/p>\n

\u2299<\/p>\n

u<\/p>\n

)<\/p>\n

b=W_o(g\\\\odot u)<\/p>\n

<\/span><\/span>b<\/span><\/span>=<\/span><\/span><\/span><\/span>W<\/span><\/span>o<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>g<\/span><\/span>\u2299<\/span><\/span><\/span><\/span>u<\/span>)<\/span><\/span><\/span><\/span><\/span>\u3001<\/p>\n

g<\/p>\n

=<\/p>\n

\u03c3<\/p>\n

(<\/p>\n

z<\/p>\n

)<\/p>\n

g=\\\\sigma(z)<\/p>\n

<\/span><\/span>g<\/span><\/span>=<\/span><\/span><\/span><\/span>\u03c3<\/span>(<\/span>z<\/span>)<\/span><\/span><\/span><\/span><\/span>，\u6709：<\/p>\n

\u8bba\u6587\u5f0f（11）：<\/p>\n

<\/p>\n

\u2207<\/p>\n

u<\/p>\n

L<\/p>\n

a<\/p>\n

c<\/p>\n

t<\/p>\n

=<\/p>\n

g<\/p>\n

\u2299<\/p>\n

W<\/p>\n

o<\/p>\n

\u22a4<\/p>\n

\u03bb<\/p>\n

,<\/p>\n

\u2207<\/p>\n

z<\/p>\n

L<\/p>\n

a<\/p>\n

c<\/p>\n

t<\/p>\n

=<\/p>\n

g<\/p>\n

\u2299<\/p>\n

(<\/p>\n

1<\/p>\n

\u2212<\/p>\n

g<\/p>\n

)<\/p>\n

\u2299<\/p>\n

u<\/p>\n

\u2299<\/p>\n

W<\/p>\n

o<\/p>\n

\u22a4<\/p>\n

\u03bb<\/p>\n

.<\/p>\n

\\\\nabla_u\\\\mathcal L_{\\\\mathrm{act}}=g\\\\odot W_o^\\\\top\\\\lambda, \\\\qquad \\\\nabla_z\\\\mathcal L_{\\\\mathrm{act}} =g\\\\odot(1-g)\\\\odot u\\\\odot W_o^\\\\top\\\\lambda. <\/p>\n

<\/span><\/span>\u2207<\/span><\/span>u<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>L<\/span><\/span>act<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>g<\/span><\/span>\u2299<\/span><\/span><\/span><\/span>W<\/span><\/span>o<\/span><\/span><\/span><\/span>\u22a4<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u03bb<\/span>,<\/span><\/span><\/span>\u2207<\/span><\/span>z<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>L<\/span><\/span>act<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>g<\/span><\/span>\u2299<\/span><\/span><\/span><\/span>(<\/span>1<\/span><\/span>\u2212<\/span><\/span><\/span><\/span>g<\/span>)<\/span><\/span>\u2299<\/span><\/span><\/span><\/span>u<\/span><\/span>\u2299<\/span><\/span><\/span><\/span>W<\/span><\/span>o<\/span><\/span><\/span><\/span>\u22a4<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u03bb<\/span>.<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

\u5c0f\u95e8\u503c\u4f1a\u8870\u51cf\u8be5\u5206\u652f\u7684\u7279\u5f81\u68af\u5ea6，\u9971\u548c\u4f1a\u8870\u51cf\u95e8logit\u68af\u5ea6；\u4f46\u8fd9\u6ca1\u6709\u8003\u8651\u6240\u6709\u6b8b\u5dee\u901a\u8def\u4e0e\u4f18\u5316\u5668\u7684\u5b9e\u9645\u66f4\u65b0，\u4e5f\u4e0d\u8bc1\u660e\u95e8\u63a7\u5bf9\u6240\u6709\u4efb\u52a1\u6709\u5bb3\u3002<\/p>\n

DeepStack\u901a\u8fc7\u591a\u5c42\u6ce8\u5165\u89c6\u89c9\u7279\u5f81\u4e30\u5bcc\u8868\u793a\u3002\u5c40\u90e8\u6270\u52a8\u4e0b，\u5404\u6ce8\u5165\u5c42\u5bf9\u52a8\u4f5c\u8f93\u51fa\u7684\u8d21\u732e\u4e3a：<\/p>\n

\u8bba\u6587\u5f0f（12）：<\/p>\n

<\/p>\n

B<\/p>\n

\u2113<\/p>\n

=<\/p>\n

\u2211<\/p>\n

s<\/p>\n

:<\/p>\n

m<\/p>\n

s<\/p>\n

><\/p>\n

\u2113<\/p>\n

\u2202<\/p>\n

y<\/p>\n

^<\/p>\n

\u2202<\/p>\n

c<\/p>\n

s<\/p>\n

\u2202<\/p>\n

c<\/p>\n

s<\/p>\n

\u2202<\/p>\n

h<\/p>\n

m<\/p>\n

s<\/p>\n

J<\/p>\n

m<\/p>\n

s<\/p>\n

\u2212<\/p>\n

1<\/p>\n

\u22ef<\/p>\n

J<\/p>\n

\u2113<\/p>\n

+<\/p>\n

1<\/p>\n

U<\/p>\n

\u2113<\/p>\n

,<\/p>\n

\u03b4<\/p>\n

y<\/p>\n

^<\/p>\n

=<\/p>\n

\u2211<\/p>\n

\u2113<\/p>\n

\u2208<\/p>\n

I<\/p>\n

B<\/p>\n

\u2113<\/p>\n

\u2009<\/p>\n

\u03b4<\/p>\n

f<\/p>\n

\u2113<\/p>\n

+<\/p>\n

O<\/p>\n

(<\/p>\n

\u2225<\/p>\n

\u03b4<\/p>\n

f<\/p>\n

\u2225<\/p>\n

2<\/p>\n

2<\/p>\n

)<\/p>\n

.<\/p>\n

\\\\begin{aligned} B_\\\\ell&=\\\\sum_{s:m_s>\\\\ell} \\\\frac{\\\\partial\\\\hat y}{\\\\partial c_s} \\\\frac{\\\\partial c_s}{\\\\partial h_{m_s}} \\\\mathcal J_{m_s-1}\\\\cdots\\\\mathcal J_{\\\\ell+1}U_\\\\ell,\\\\\\\\ \\\\delta\\\\hat y&=\\\\sum_{\\\\ell\\\\in\\\\mathcal I}B_\\\\ell\\\\,\\\\delta f_\\\\ell +O(\\\\lVert\\\\delta f\\\\rVert_2^2). \\\\end{aligned} <\/p>\n

<\/span><\/span><\/span>B<\/span><\/span>\u2113<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u03b4<\/span><\/span>y<\/span><\/span><\/span>^<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span>s<\/span>:<\/span>m<\/span><\/span>s<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>><\/span>\u2113<\/span><\/span><\/span><\/span><\/span>\u2211<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u2202<\/span>c<\/span><\/span>s<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u2202<\/span><\/span>y<\/span><\/span><\/span>^<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u2202<\/span>h<\/span><\/span>m<\/span><span class=\\"\\" style=\\"<\/p>\n","protected":false},"excerpt":{"rendered":"

\u5199\u5728\u524d\u9762 \u6b22\u8fce\u5927\u5bb6\u5173\u6ce8Rocky\u7684\u77e5\u4e4e：Rocky Ding \u300a\u4e09\u5e74\u9762\u8bd5\u4e94\u5e74\u6a21\u62df\u300bAIGC\/LLM\/AI Agent\u7b97\u6cd5\u5de5\u7a0b\u5e08\/\u5f00\u53d1\u5de5\u7a0b\u5e08\u6c42\u804c\u9762\u8bd5\u79d8\u7c4d\u72ec\u5bb6\u8d44\u6e90：\u3010\u4e09\u5e74\u9762\u8bd5\u4e94\u5e74\u6a21\u62df\u3011WeThinkIn\/AIGC-I […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"topic":[],"class_list":["post-120419","post","type-post","status-publish","format-standard","hentry","category-server"],"yoast_head":"\n\u6df1\u5165\u6d45\u51faGroundingPI\u6280\u672f\u89e3\u6790\uff1a\u7cbe\u786e\u7a7a\u95f4\u611f\u77e5\u5982\u4f55\u8fc1\u79fb\u5230\u673a\u5668\u4eba\u52a8\u4f5c - \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\/120419.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u6df1\u5165\u6d45\u51faGroundingPI\u6280\u672f\u89e3\u6790\uff1a\u7cbe\u786e\u7a7a\u95f4\u611f\u77e5\u5982\u4f55\u8fc1\u79fb\u5230\u673a\u5668\u4eba\u52a8\u4f5c - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u5199\u5728\u524d\u9762 \u6b22\u8fce\u5927\u5bb6\u5173\u6ce8Rocky\u7684\u77e5\u4e4e：Rocky Ding \u300a\u4e09\u5e74\u9762\u8bd5\u4e94\u5e74\u6a21\u62df\u300bAIGC\/LLM\/AI Agent\u7b97\u6cd5\u5de5\u7a0b\u5e08\/\u5f00\u53d1\u5de5\u7a0b\u5e08\u6c42\u804c\u9762\u8bd5\u79d8\u7c4d\u72ec\u5bb6\u8d44\u6e90：\u3010\u4e09\u5e74\u9762\u8bd5\u4e94\u5e74\u6a21\u62df\u3011WeThinkIn\/AIGC-I […]\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/120419.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-10-31T09:19:49+00:00\" \/>\n<meta name=\"author\" content=\"\u7f51\u7855\u5c0fA\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"\u7f51\u7855\u5c0fA\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"10 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/120419.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/120419.html\",\"name\":\"\u6df1\u5165\u6d45\u51faGroundingPI\u6280\u672f\u89e3\u6790\uff1a\u7cbe\u786e\u7a7a\u95f4\u611f\u77e5\u5982\u4f55\u8fc1\u79fb\u5230\u673a\u5668\u4eba\u52a8\u4f5c - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-10-31T09:19:49+00:00\",\"dateModified\":\"2026-10-31T09:19:49+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/5c8055c670c9d136620185663d0813ab\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/120419.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/120419.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/120419.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u6df1\u5165\u6d45\u51faGroundingPI\u6280\u672f\u89e3\u6790\uff1a\u7cbe\u786e\u7a7a\u95f4\u611f\u77e5\u5982\u4f55\u8fc1\u79fb\u5230\u673a\u5668\u4eba\u52a8\u4f5c\"}]},{\"@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\/5c8055c670c9d136620185663d0813ab\",\"name\":\"\u7f51\u7855\u5c0fA\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/0812f751cb127092dc3a9e310765b8b68fc3fd388d0ccef020588dbb4c61de91?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/0812f751cb127092dc3a9e310765b8b68fc3fd388d0ccef020588dbb4c61de91?s=96&d=mystery\",\"caption\":\"\u7f51\u7855\u5c0fA\"},\"sameAs\":[\"https:\/\/www.wsisp.com\/helps\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/wsisphelp\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u6df1\u5165\u6d45\u51faGroundingPI\u6280\u672f\u89e3\u6790\uff1a\u7cbe\u786e\u7a7a\u95f4\u611f\u77e5\u5982\u4f55\u8fc1\u79fb\u5230\u673a\u5668\u4eba\u52a8\u4f5c - \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\/120419.html","og_locale":"zh_CN","og_type":"article","og_title":"\u6df1\u5165\u6d45\u51faGroundingPI\u6280\u672f\u89e3\u6790\uff1a\u7cbe\u786e\u7a7a\u95f4\u611f\u77e5\u5982\u4f55\u8fc1\u79fb\u5230\u673a\u5668\u4eba\u52a8\u4f5c - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u5199\u5728\u524d\u9762 \u6b22\u8fce\u5927\u5bb6\u5173\u6ce8Rocky\u7684\u77e5\u4e4e：Rocky Ding \u300a\u4e09\u5e74\u9762\u8bd5\u4e94\u5e74\u6a21\u62df\u300bAIGC\/LLM\/AI Agent\u7b97\u6cd5\u5de5\u7a0b\u5e08\/\u5f00\u53d1\u5de5\u7a0b\u5e08\u6c42\u804c\u9762\u8bd5\u79d8\u7c4d\u72ec\u5bb6\u8d44\u6e90：\u3010\u4e09\u5e74\u9762\u8bd5\u4e94\u5e74\u6a21\u62df\u3011WeThinkIn\/AIGC-I […]","og_url":"https:\/\/www.wsisp.com\/helps\/120419.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-10-31T09:19:49+00:00","author":"\u7f51\u7855\u5c0fA","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"\u7f51\u7855\u5c0fA","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"10 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/120419.html","url":"https:\/\/www.wsisp.com\/helps\/120419.html","name":"\u6df1\u5165\u6d45\u51faGroundingPI\u6280\u672f\u89e3\u6790\uff1a\u7cbe\u786e\u7a7a\u95f4\u611f\u77e5\u5982\u4f55\u8fc1\u79fb\u5230\u673a\u5668\u4eba\u52a8\u4f5c - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-10-31T09:19:49+00:00","dateModified":"2026-10-31T09:19:49+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/5c8055c670c9d136620185663d0813ab"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/120419.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/120419.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/120419.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"\u6df1\u5165\u6d45\u51faGroundingPI\u6280\u672f\u89e3\u6790\uff1a\u7cbe\u786e\u7a7a\u95f4\u611f\u77e5\u5982\u4f55\u8fc1\u79fb\u5230\u673a\u5668\u4eba\u52a8\u4f5c"}]},{"@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\/5c8055c670c9d136620185663d0813ab","name":"\u7f51\u7855\u5c0fA","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/0812f751cb127092dc3a9e310765b8b68fc3fd388d0ccef020588dbb4c61de91?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/0812f751cb127092dc3a9e310765b8b68fc3fd388d0ccef020588dbb4c61de91?s=96&d=mystery","caption":"\u7f51\u7855\u5c0fA"},"sameAs":["https:\/\/www.wsisp.com\/helps"],"url":"https:\/\/www.wsisp.com\/helps\/author\/wsisphelp"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/120419","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\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=120419"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/120419\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=120419"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=120419"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=120419"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=120419"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}