{"id":297,"date":"2026-03-08T23:14:23","date_gmt":"2026-03-08T15:14:23","guid":{"rendered":"https:\/\/blog.zeronely.cn\/?p=297"},"modified":"2026-03-08T23:14:24","modified_gmt":"2026-03-08T15:14:24","slug":"%e5%85%b3%e4%ba%8e%e6%af%95%e8%ae%be%e7%9a%84%e5%88%9d%e6%ac%a1%e6%80%9d%e8%80%83","status":"publish","type":"post","link":"https:\/\/blog.zeronely.cn\/index.php\/2026\/03\/08\/%e5%85%b3%e4%ba%8e%e6%af%95%e8%ae%be%e7%9a%84%e5%88%9d%e6%ac%a1%e6%80%9d%e8%80%83\/","title":{"rendered":"\u5173\u4e8e\u6bd5\u8bbe\u7684\u521d\u6b21\u601d\u8003"},"content":{"rendered":"\n

Cross-Modal Causal Relational Reasoning for Event-Level Visual Question Answering<\/h2>\n\n\n\n

\u63d0\u51fa\u8de8\u6a21\u6001\u56e0\u679c\u5173\u7cfb\u63a8\u7406\u6846\u67b6<\/strong>\uff08CMCIR\uff09\uff0c\u8be5\u6846\u67b6\u5305\u542b\u56e0\u679c\u611f\u77e5\u89c6\u89c9\u8bed\u8a00\u63a8\u7406\uff08CVLR\uff09\u3001\u65f6\u7a7a Transformer\uff08STT\uff09\u3001\u89c6\u89c9\u8bed\u8a00\u7279\u5f81\u878d\u5408\uff08VLFF\uff09\u4e09\u5927\u6838\u5fc3\u6a21\u5757\uff0c\u901a\u8fc7\u524d\u95e8\u548c\u540e\u95e8\u56e0\u679c\u5e72\u9884\u89e3\u5f00\u89c6\u89c9\u4e0e\u8bed\u8a00\u7684\u865a\u5047\u5173\u8054\u3001\u6355\u6349\u7ec6\u7c92\u5ea6\u7684\u6a21\u6001\u4ea4\u4e92\u5e76\u81ea\u9002\u5e94\u878d\u5408\u5168\u5c40\u8bed\u4e49\u7279\u5f81\uff1b\u5728SUTD-TrafficQA\u3001TGIFQA\u3001MSVD-QA\u3001MSRVTT-QA\u56db\u4e2a\u4e8b\u4ef6\u7ea7\u6570\u636e\u96c6\uff08what huo\u7b49\u7684\u95ee\u7b54\u6570\u636e\u96c6\uff09\u4e0a\u7684\u5b9e\u9a8c\u9a8c\u8bc1\u4e86 CMCIR \u7684\u4f18\u8d8a\u6027\uff0c\u5176\u5728\u56e0\u679c\u7ed3\u6784\u53d1\u73b0\u548c\u9c81\u68d2\u7684\u4e8b\u4ef6\u7ea7\u89c6\u89c9\u95ee\u7b54\u4e0a\u8868\u73b0\u7a81\u51fa\uff0c\u540c\u65f6\u6d88\u878d\u5b9e\u9a8c\u3001\u53c2\u6570\u654f\u611f\u6027\u5206\u6790\u7b49\u4e5f\u8bc1\u5b9e\u4e86\u5404\u6a21\u5757\u7684\u6709\u6548\u6027\uff0c\u8be5\u7814\u7a76\u4e5f\u662f\u9996\u6b21\u4e3a\u4e8b\u4ef6\u7ea7\u89c6\u89c9\u95ee\u7b54\u4efb\u52a1\u63a2\u7d22\u8de8\u6a21\u6001\u56e0\u679c\u7ed3\u6784\u3002<\/p>\n\n\n\n

\u4e3b\u8981\u662f\u63d0\u51fa\u4e86\u65b0\u7684\u67b6\u6784\uff0c\u7136\u540e\u5728qa\u6570\u636e\u96c6\u4e0a\u8dd1sota\uff0c\u5728\u56e0\u679c\u7ed3\u6784\u53d1\u73b0\u4e0a\u8868\u73b0\u7a81\u51fa\u3002<\/strong><\/p>\n\n\n\n

InterVid<\/h2>\n\n\n\n

1 \u63d0\u51fa\u4e86\u4e00\u4e2a\u6570\u636e\u96c6<\/strong>\uff0c\u9ad8\u8bed\u4e49\u76f8\u5173\u7684\u89c6\u9891 – \u6587\u672c\u6570\u636e\u96c6
2 \u7528\u4e00\u4e2a\u57fa\u4e8eclip\u7684vit-l\u67b6\u6784\u6784\u5efa\u7684viclip\u4f5c\u4e3a\u57fa\u7ebf\uff0c\u7528\u4e0d\u540c\u6570\u636e\u96c6\u8bad\u7ec3\uff0c\u5728\u96f6\u6837\u672c\u52a8\u4f5c\u8bc6\u522b\u548c\u89c6\u9891\u68c0\u7d22\u5927\u5e45\u5ea6\u8d85\u8d8a\uff08\u4ece\u800c\u8bc1\u660e\u6570\u636e\u96c6\u6709\u6548\u6027\uff09
3 \u5237sota\u7684\u6570\u636e\u96c6\u5206\u522b\u662fKinetics \u7cfb\u5217\uff08\u96f6\u6837\u672c\u52a8\u4f5c\u8bc6\u522b\uff09\u548cMSR-VTT\uff08\u516c\u5f00\u89c6\u9891\u63cf\u8ff0\u6570\u636e\u96c6\uff09LSMDC\uff08Large Scale Movie Description Challenge\uff09\uff08\u57fa\u4e8e\u7535\u5f71\u7247\u6bb5\u7684\u89c6\u9891\u68c0\u7d22 \/ \u63cf\u8ff0\u6570\u636e\u96c6\uff09\u7b49\u89c6\u9891\u68c0\u7d22\u4efb\u52a1\u6570\u636e\u96c6\uff0c\u8fd8\u6709\u90e8\u5206SomethingSomethingV2\uff08SthSthV2\uff09\/SomethingSomethingV2\uff08SthSthV2\uff09\u7b49\u8f85\u52a9\u9a8c\u8bc1\u7684\u6570\u636e\u96c6\u3002<\/p>\n\n\n\n

intervid2<\/h2>\n\n\n\n

1 6B \u53c2\u6570\u591a\u6a21\u6001\u89c6\u9891\u7406\u89e3\u57fa\u7840\u6a21\u578b
2 \u6784\u5efa\u4e86\u542b402M \u6761\u6570\u636e\u6761\u76ee\u7684\u5927\u89c4\u6a21\u591a\u6a21\u6001\u89c6\u9891\u8bad\u7ec3\u96c6
3 \u572860 + \u97f3\u89c6\u9891\u4efb\u52a1\u4e0a\u53d6\u5f97SOTA\u6027\u80fd\uff0c\u5c24\u5176\u5728\u957f\u89c6\u9891\u7406\u89e3\u3001\u89c6\u9891\u5bf9\u8bdd\u3001\u65f6\u5e8f\u52a8\u4f5c\u8bc6\u522b\u3001\u89c6\u9891\u68c0\u7d22\u7b49\u4efb\u52a1\u8868\u73b0\u7a81\u51fa
4 \u6838\u5fc3\u65b9\u6cd5\u662f\u4e09\u9636\u6bb5\u6e10\u8fdb\u5f0f\u8bad\u7ec3\u4e0e\u6a21\u578b\u67b6\u6784\uff0c6B \u53c2\u6570\u7684\u89c6\u9891\u7f16\u7801\u5668\uff0c\u57fa\u4e8e Vision Transformer\uff08ViT\uff09\u6784\u5efa\uff0c\u52a0\u5165\u6ce8\u610f\u529b\u6c60\u5316\u548c3D \u4f4d\u7f6e\u5d4c\u5165\uff1b\u8f93\u5165\u89c6\u9891\u91c7\u7528\u7a00\u758f\u91c7\u6837 8 \u5e27\u300114\u00d714 \u7a7a\u95f4\u4e0b\u91c7\u6837\uff0c\u751f\u6210\u65f6\u7a7a token \u540e\u8fdb\u884c\u540e\u7eed\u8bad\u7ec3<\/p>\n\n\n\n

\u5173\u4e8e\u6bd5\u8bbe\u7684\u4e00\u4e9b\u60f3\u6cd5<\/h2>\n\n\n\n

\u9996\u5148\u662f\u4e3b\u4efb\u52a1<\/strong><\/p>\n\n\n\n

\u4e3b\u4efb\u52a1 A\uff1a\u56e0\u679c\/\u65f6\u5e8f
NExT-QA<\/strong>\uff08NExT-QA: Next Phase of Question-Answering to Explaining Temporal Actions (CVPR’21)\uff09
Causal-VidQA<\/strong>\u6570\u636e\u96c6[CVPR 2022] A large-scale public benchmark dataset for video question-answering, especially about evidence and commonsense reasoning.<\/p>\n\n\n\n

\u4e3b\u4efb\u52a1 B\uff1a\u4e8b\u4ef6\u7ea7\u56e0\u679c\u56fe\/\u4e8b\u4ef6\u94fe\u53d1\u73b0<\/p>\n\n\n\n

MECD \/ MECD+<\/strong><\/p>\n\n\n\n

\u6784\u5efa\u4e8b\u4ef6\u56e0\u679c\u56fe<\/p>\n\n\n\n

CausalVQA\uff08physically grounded\uff09
\u7269\u7406\u56e0\u679c\u95ee\u7b54 \u53cd\u4e8b\u5b9e\/\u5047\u8bbe\/\u9884\u6d4b\u63a8\u7406<\/p>\n\n\n\n

\u6211\u7684\u76ee\u6807\uff1a
\u628a\u89c6\u9891\u62c6\u6210\u4e00\u4e32\u4e8b\u4ef6\uff08event segments\uff09\uff0c\u518d\u5efa\u4e00\u4e2a\u4e8b\u4ef6\u4e4b\u95f4\u7684\u5173\u7cfb\u7ed3\u6784\uff08\u94fe\/\u56fe\uff09\uff0c\u6700\u540e\u7528\u8fd9\u4e2a\u7ed3\u6784\u53bb\u56de\u7b54\u95ee\u9898\uff08QA\uff09\u6216\u505a\u56e0\u679c\/\u65f6\u95f4\u63a8\u7406\u3002<\/strong><\/p>\n\n\n\n

\u90e8\u5206\u8bba\u6587\u8c03\u7814\uff1a<\/p>\n\n\n\n

EC-GNNs\uff082023\uff09<\/strong>\uff1a\u628a\u201cdense caption\uff08\u5bc6\u96c6\u4e8b\u4ef6\u5b57\u5e55\uff09\u201d\u4f5c\u4e3a\u65b0\u6a21\u6001\uff0c\u5f15\u5165\u56fe\u7f51\u7edc\u505a\u8de8\u6a21\u6001\u63a8\u7406\uff08caption\/video\/question \u4e09\u56fe + cross-modal attention + \u591a\u6b65\u63a8\u7406\u878d\u5408\uff09\uff0c\u5728 TGIF-QA \/ MSVD-QA \u4e0a\u9a8c\u8bc1\u4e8b\u4ef6\u4fe1\u606f\u6709\u7528<\/p>\n\n\n\n

\u7528\u4e8b\u4ef6\/\u5b57\u5e55\u505a\u7b2c\u4e09\u6a21\u6001\uff0c\u7136\u540e\u56fe\u63a8\u7406<\/p>\n\n\n\n

CLanG\uff082024\uff09<\/strong>\uff1a\u5f3a\u8c03\u201cmulti-object event graph representation learning\u201d\uff0c\u9762\u5411 NExT-QA \/ TGIF-QA-R \u7b49\uff0c\u6838\u5fc3\u662f\u628a\u89c6\u9891\u91cc\u7684\u5bf9\u8c61-\u4e8b\u4ef6\u7ec4\u7ec7\u6210\u56fe\uff0c\u5e76\u505a\u8868\u793a\u5b66\u4e60\u4e0e\u63a8\u7406<\/p>\n\n\n\n

\u591a\u5bf9\u8c61\u4e8b\u4ef6\u56fe<\/p>\n\n\n\n

ENTER\uff082025\uff09<\/strong>\uff1a\u660e\u786e\u63d0\u51fa event-based interpretable reasoning\uff1a\u7528\u6a21\u578b\u751f\u6210event graphs\uff0c\u518d\u7528 LLM \u57fa\u4e8e\u56fe\u505a\u63a8\u7406\uff0c\u5728 NExT-QA \u8fd9\u7c7b\u57fa\u51c6\u7ed9\u51fa\u5f88\u5f3a\u7684 zero-shot \u7ed3\u679c<\/p>\n\n\n\n

\u7528 LLM \u628a\u89c6\u9891\u53d8\u6210\u4e8b\u4ef6\u56fe\uff0c\u518d\u63a8\u7406<\/p>\n\n\n\n

ICE-Chain\uff1aIntervention-Consistent Event Chain Reasoning<\/p>\n\n\n\n

\u628a\u4e8b\u4ef6\u94fe\u5f53\u6210\u201c\u53ef\u5e72\u9884\u7684\u7ed3\u6784\u5316\u4e2d\u95f4\u53d8\u91cf\u201d\uff0c\u5e76\u7528\u201c\u5e72\u9884\u4e00\u81f4\u6027\uff08intervention consistency\uff09\u201d\u53bb\u5b66\u4e60\u54ea\u4e9b\u4e8b\u4ef6\u771f\u7684\u5bfc\u81f4\u7b54\u6848\u3002<\/p>\n\n\n\n

\u5982\u4f55\u5b9e\u73b0\uff1a<\/p>\n\n\n\n

**\u5bf9\u6bcf\u4e2a\u89c6\u9891\uff0c\u6784\u9020\u4e8b\u4ef6\u5e8f\u5217 **
e1,e2,\u2026,eN\uff1a<\/p>\n\n\n\n

\u5728 MECD \u4e0a\uff1a\u76f4\u63a5\u7528\u6570\u636e\u96c6\u81ea\u5e26\u7684\u4e8b\u4ef6\u7247\u6bb5\u4e0e\u4e8b\u4ef6\u6587\u672c<\/p>\n\n\n\n

\u6bcf\u4e2a\u4e8b\u4ef6\u8282\u70b9\u505a\u4e00\u4e2a\u8de8\u6a21\u6001\u8868\u793a\uff1a<\/strong><\/p>\n\n\n\n

\u89c6\u89c9\uff1a\u89c6\u9891\u7f16\u7801\u5668\u8f93\u51fa\uff08\u590d\u7528 NExT-QA\/Causal-VidQA \u5b98\u65b9\/\u793e\u533a\u63d0\u4f9b\u7684\u7279\u5f81\u7ba1\u7ebf\uff0c\u964d\u4f4e\u6210\u672c\uff09<\/p>\n\n\n\n

\u6587\u672c\uff1a\u4e8b\u4ef6\u63cf\u8ff0\uff08\u6765\u81ea\u6570\u636e\u96c6\u6216\u751f\u6210\u5b57\u5e55\uff09\u7f16\u7801<\/p>\n\n\n\n

\u6784\u5efa\u4e24\u7c7b\u8fb9\uff1a<\/strong><\/p>\n\n\n\n

\u65f6\u95f4\u8fb9\uff1a\u6839\u636e\u65f6\u95f4\u6233\/\u5207\u6bb5\u987a\u5e8f\u5f97\u5230 before\/after\uff08\u786e\u5b9a\u6027\uff09<\/p>\n\n\n\n

\u56e0\u679c\u8fb9\uff08\u53ef\u5b66\u4e60\uff09\uff1a\u7528\u201c\u4e8b\u4ef6 Granger \u98ce\u683c\u7684\u5e72\u9884\u8bc4\u5206\u201d\u5b66\u4e60<\/p>\n\n\n\n

\u601d\u8def\u501f\u9274 MECD+ \u7684 Event Granger Test\uff1a\u6bd4\u8f83\u201cmask \u67d0\u4e2a\u524d\u7f6e\u4e8b\u4ef6\u201d\u5bf9\u201c\u9884\u6d4b\u7ed3\u679c\u4e8b\u4ef6\u8868\u793a\/\u7b54\u6848\u6982\u7387\u201d\u7684\u5f71\u54cd\uff0c\u4ece\u800c\u4f30\u8ba1\u56e0\u679c\u8d21\u732e<\/p>\n\n\n\n

\u521b\u65b0\uff1a\u628a\u8fd9\u4e2a\u8bc4\u5206\u7aef\u5230\u7aef\u5730\u63a5\u5230 VideoQA\uff0c\u8ba9\u5b83\u4e3a\u201c\u9009\u9879\u5224\u522b\u201d\u670d\u52a1\uff0c\u800c\u4e0d\u662f\u53ea\u505a\u56e0\u679c\u56fe\u4efb\u52a1\u3002<\/p>\n\n\n\n

\u5173\u952e\u521b\u65b0\uff1a<\/p>\n\n\n\n

Option-aware \u94fe\u9a8c\u8bc1\uff08\u4e13\u95e8\u5bf9\u4ed8\u591a\u9009\u504f\u7f6e\uff09<\/p>\n\n\n\n

\u591a\u9009\u9898\u7684\u4e00\u4e2a\u73b0\u5b9e\u95ee\u9898\u662f\uff1a\u6a21\u578b\u5f88\u5bb9\u6613\u8d70\u201c\u9009\u9879-\u89c6\u89c9\u76f8\u4f3c\u5ea6\u6377\u5f84\u201d\uff08EOB\uff09
\u4f60\u7684\u505a\u6cd5\u662f\uff1a\u5bf9\u6bcf\u4e2a\u5019\u9009\u9009\u9879 o_k \u5355\u72ec\u6784\u9020\u4e00\u6761\u6700\u652f\u6301\u5b83\u7684\u4e8b\u4ef6\u94fe\uff0c\u518d\u505a\u201c\u94fe\u9a8c\u8bc1\u201d\uff1a<\/p>\n\n\n\n

\u5bf9\u6bcf\u4e2a\u9009\u9879 k\uff1a\u4ece\u4e8b\u4ef6\u56fe\u91cc\u68c0\u7d22\/\u63a8\u7406\u5f97\u5230 top-1 \u6216 top-K \u6761\u94fe<\/p>\n\n\n\n

\u5173\u952e\u521b\u65b0\uff1aOption-aware \u94fe\u9a8c\u8bc1<\/strong>\uff08\u4e13\u95e8\u5bf9\u4ed8\u591a\u9009\u504f\u7f6e\uff09<\/p>\n\n\n\n

\u591a\u9009\u9898\u7684\u4e00\u4e2a\u73b0\u5b9e\u95ee\u9898\u662f\uff1a\u6a21\u578b\u5f88\u5bb9\u6613\u8d70\u201c\u9009\u9879-\u89c6\u89c9\u76f8\u4f3c\u5ea6\u6377\u5f84\u201d\uff08EOB\uff09([arXiv][11])
\u4f60\u7684\u505a\u6cd5\u662f\uff1a\u5bf9\u6bcf\u4e2a\u5019\u9009\u9009\u9879 o_k \u5355\u72ec\u6784\u9020\u4e00\u6761\u6700\u652f\u6301\u5b83\u7684\u4e8b\u4ef6\u94fe<\/strong>\uff0c\u518d\u505a\u201c\u94fe\u9a8c\u8bc1\u201d\uff1a<\/p>\n\n\n\n

    \n
  • \u5bf9\u6bcf\u4e2a\u9009\u9879 k\uff1a\u4ece\u4e8b\u4ef6\u56fe\u91cc\u68c0\u7d22\/\u63a8\u7406\u5f97\u5230 top-1 \u6216 top-K \u6761\u94fe
    [
    \\text{chain}k = (e<\/em>{a}\\rightarrow e_{b}\\rightarrow \\dots \\rightarrow e_{t})
    ]<\/li>\n\n\n\n
  • \u628a\uff08\u89c6\u9891\u3001\u95ee\u9898\u3001\u9009\u9879\u3001\u94fe\uff09\u4e00\u8d77\u8f93\u5165\u5230\u7b54\u6848\u6253\u5206\u5668\uff0c\u5f97\u5230 score_k<\/li>\n\n\n\n
  • \u6700\u7ec8\u9009 argmax(score_k)<\/li>\n<\/ul>\n\n\n\n

    \u8fd9\u6837\u505a\u7684\u610f\u4e49\uff1a<\/p>\n\n\n\n

      \n
    • \u5982\u679c\u8d1f\u9009\u9879\u88ab GroundAttack \u53d8\u5f97\u201c\u89c6\u89c9\u4e0a\u4e5f\u5408\u7406\u201d\uff0c\u4f60\u4f9d\u7136\u53ef\u4ee5\u9760\u56e0\u679c\u94fe\u4e00\u81f4\u6027<\/strong>\u7b5b\u6389\u5b83\uff0c\u800c\u4e0d\u662f\u9760\u201c\u8fd9\u9009\u9879\u592a\u79bb\u8c31\u4e86\u201d<\/li>\n<\/ul>\n\n\n\n

      \u7528\u4e09\u4e2a loss\uff08\u4ece\u6613\u5230\u96be\uff09\uff1a<\/p>\n\n\n\n

      L_QA\uff1aVideoQA \u4ea4\u53c9\u71b5\uff08\u4e3b\u4efb\u52a1\uff09<\/p>\n\n\n\n

      L_graph\uff1a\u56e0\u679c\u8fb9\u76d1\u7763\/\u5f31\u76d1\u7763<\/p>\n\n\n\n

      \u7528 MECD \u7684\u201ccausality \/ complete causality\u201d\u5f53\u76d1\u7763\u4fe1\u53f7\uff0c\u8bad\u7ec3\u4f60\u7684\u56e0\u679c\u8fb9\u9884\u6d4b\u5668<\/p>\n\n\n\n

      L_intervene\uff1a\u5e72\u9884\u4e00\u81f4\u6027\u6b63\u5219\uff08\u8bba\u6587\u4eae\u70b9\uff09<\/p>\n\n\n\n

      \u968f\u673a mask \u6389\u94fe\u4e0a\u7684\u4e8b\u4ef6 vs mask \u975e\u94fe\u4e8b\u4ef6<\/p>\n\n\n\n

      \u7ea6\u675f\uff1amask \u94fe\u4e8b\u4ef6\u5e94\u663e\u8457\u964d\u4f4e\u6b63\u786e\u9009\u9879\u5f97\u5206\uff1bmask \u975e\u94fe\u4e8b\u4ef6\u5f71\u54cd\u66f4\u5c0f<\/p>\n\n\n\n

      \u8fd9\u4f1a\u8ba9\u4f60\u5b66\u5230\u201c\u5fc5\u8981\u4e8b\u4ef6\u94fe\u201d\uff0c\u5e76\u63d0\u5347\u53ef\u89e3\u91ca\u6027\u4e0e\u9c81\u68d2\u6027\uff08\u4e5f\u66f4\u50cf\u201c\u56e0\u679c\u201d\u800c\u4e0d\u662f\u201c\u6ce8\u610f\u529b\u70ed\u529b\u56fe\u201d\uff09\u3002<\/p>\n\n\n\n

      MECD \/ MECD+\uff1a\u7528\u6765\u5b66\u201c\u4e8b\u4ef6\u7ea7\u56e0\u679c\u7ed3\u6784\u201d<\/p>\n\n\n\n

      NExT-QA\uff1a\u7528\u6765\u505a\u201c\u901a\u7528\u56e0\u679c\/\u65f6\u95f4 VideoQA\u201d\u7684\u4e3b\u5bf9\u6bd4
      \u5728 EOB-free\uff08GroundAttack \u751f\u6210\uff09 \u6216\u90a3\u4e2avtuber\u6570\u636e\u751f\u6210\u7684 hard negatives \u4e0a\u518d\u8bc4\u6d4b\u4e00\u6b21<\/p>\n\n\n\n

      baselines<\/strong><\/p>\n\n\n\n

      HCRN \/ HGA \/ HME \/ B2A \u8fd9\u7c7b VideoQA backbone<\/p>\n\n\n\n

      NExT-QA repo<\/p>\n\n\n\n

      \u56e0\u679c\/\u9c81\u68d2\u65b9\u5411 baselines<\/strong><\/p>\n\n\n\n

      MCR\uff08CVPR 2023\uff09\uff1a\u5178\u578b\u201c\u591a\u6a21\u6001\u56e0\u679c\u63a8\u7406\u201d\u6846\u67b6\uff0c\u660e\u786e\u5728 Causal-VidQA \u4e0e NExT-QA \u4e0a\u505a\u5bf9\u6bd4\u5e76\u62a5\u544a\u63d0\u5347<\/p>\n\n\n\n

      \u4e8b\u4ef6\u56fe\/LLM-based baselines<\/strong><\/p>\n\n\n\n

      ENTER\uff08event graph + LLM reasoning\uff09\uff1a\u53ef\u4ee5\u4f5c\u4e3a\u201c\u7ed3\u6784\u5316\u4e8b\u4ef6\u56fe + LLM \u63a8\u7406\u201d\u7684\u5f3a\u57fa\u7ebf\/\u4e0a\u754c\u53c2\u8003<\/p>\n\n\n\n

      \"alt<\/div><\/figure>\n\n\n\n
        \n
      • \u5148\u8dd1\u901a NExT-QA \/ Causal-VidQA \u7684\u4e00\u4e2a\u7ecf\u5178 baseline\uff08\u786e\u4fdd\u7ba1\u7ebf\u6ca1\u95ee\u9898\uff09<\/li>\n\n\n\n
      • \u5b9e\u73b0 ICE-Chain \u7684\u6700\u7b80\u7248\u672c\uff1a\u5747\u5300\u5207\u6bb5=\u4e8b\u4ef6\uff1b\u53ea\u505a\u65f6\u95f4\u8fb9\uff1b\u94fe\u68c0\u7d22\u7528 attention\/Top-K<\/li>\n\n\n\n
      • \u52a0 MECD \u76d1\u7763\uff1a\u8bad\u7ec3\u56e0\u679c\u8fb9\u9884\u6d4b\u5668\uff08\u5148\u5728 MECD \u4e0a\u628a\u8fb9\u9884\u6d4b\u505a\u597d\uff09<\/li>\n\n\n\n
      • \u628a\u56e0\u679c\u8fb9\u8fc1\u79fb\u5230 QA\uff1a\u52a0\u5165 L_intervene \u5e72\u9884\u4e00\u81f4\u6027\uff0c\u8ba9\u94fe\u771f\u6b63\u201c\u5f71\u54cd\u7b54\u6848\u201d<\/li>\n\n\n\n
      • \u505a\u53bb\u504f\u7f6e\u9a8c\u8bc1\uff1a\u62ff GroundAttack\/EOB-free\uff08\u82e5\u62ff\u5f97\u5230\u6807\u6ce8\u6216\u81ea\u5df1\u751f\u6210 hard negatives\uff09\u8bc4\u6d4b\u4e00\u6b21\uff0c\u5199\u6210\u4eae\u70b9<\/li>\n\n\n\n
      • \u89c6\u9891\u7f16\u7801\u5668\uff1a\u7528\u73b0\u6210\u9884\u8bad\u7ec3\/\u9884\u63d0\u53d6\u7279\u5f81\uff08?\uff09<\/li>\n\n\n\n
      • \u6587\u672c\u7f16\u7801\u5668\uff1aBERT \u7c7b\uff08\u6216\u8f7b\u91cf Transformer\uff09<\/li>\n\n\n\n
      • \u4e8b\u4ef6\u56fe\u63a8\u7406\uff1a2\uff5e4 \u5c42 Graph Transformer \/ GAT \/ GCN<\/li>\n\n\n\n
      • \u9009\u9879\u6253\u5206\uff1aMLP \u6216 cross-attention head<\/li>\n<\/ul>\n\n\n\n

        or<\/p>\n\n\n\n

          \n
        • \u4e8b\u4ef6\u94fe\/\u4e8b\u4ef6\u56fe\u88ab\u5e8f\u5217\u5316\u6210\u7ed3\u6784\u5316\u6587\u672c<\/strong>\uff08\u4f8b\u5982 JSON\/YAML \u98ce\u683c\u7684 event list + edges\uff09<\/li>\n\n\n\n
        • \u628a\u5b83\u4f5c\u4e3a\u989d\u5916\u63d0\u793a\u8f93\u5165\u5927\u6a21\u578b\uff0c\u6216\u8005\u7528 adapter \u628a\u56fe\u8868\u5f81\u6ce8\u5165 LMM<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"

          Cross-Modal Causal Relational Reasoning for Event-Level […]<\/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":[],"class_list":["post-297","post","type-post","status-publish","format-standard","hentry","category-uncategorized"],"_links":{"self":[{"href":"https:\/\/blog.zeronely.cn\/index.php\/wp-json\/wp\/v2\/posts\/297","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/blog.zeronely.cn\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blog.zeronely.cn\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blog.zeronely.cn\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/blog.zeronely.cn\/index.php\/wp-json\/wp\/v2\/comments?post=297"}],"version-history":[{"count":1,"href":"https:\/\/blog.zeronely.cn\/index.php\/wp-json\/wp\/v2\/posts\/297\/revisions"}],"predecessor-version":[{"id":298,"href":"https:\/\/blog.zeronely.cn\/index.php\/wp-json\/wp\/v2\/posts\/297\/revisions\/298"}],"wp:attachment":[{"href":"https:\/\/blog.zeronely.cn\/index.php\/wp-json\/wp\/v2\/media?parent=297"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blog.zeronely.cn\/index.php\/wp-json\/wp\/v2\/categories?post=297"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blog.zeronely.cn\/index.php\/wp-json\/wp\/v2\/tags?post=297"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}