{"id":10730,"date":"2026-08-04T18:29:23","date_gmt":"2026-08-04T10:29:23","guid":{"rendered":"https:\/\/infernews.com\/blog\/deferred-exposure-of-future-trajectories-for-verifiable-reasoning-in-autonomous-2\/"},"modified":"2026-08-04T18:29:23","modified_gmt":"2026-08-04T10:29:23","slug":"deferred-exposure-of-future-trajectories-for-verifiable-reasoning-in-autonomous-2","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/deferred-exposure-of-future-trajectories-for-verifiable-reasoning-in-autonomous-2\/","title":{"rendered":"DEFT-RLVR \u7528\u5ef6\u5f8c\u66dd\u5149\u6e1b\u5c11\u81ea\u52d5\u99d5\u99db VLM \u8aa4\u5224"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/08\/pasted-20e0d133b64d.jpg\" alt=\"Ground-truth trajectory exposure can induce post-hoc rationalization and hallucination.\"><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u81ea\u52d5\u99d5\u99db Vision-language-action models \u4e00\u65e6\u5728\u63a8\u7406\u524d\u5148\u898b\u5230\u771f\u5be6\u672a\u4f86\u8ecc\u8de1\uff0c\u5f88\u5bb9\u6613\u628a\u7b54\u6848\u5408\u7406\u5316\uff0c\u537b\u672a\u5fc5\u771f\u4fc2\u6839\u64da\u5834\u666f\u4f5c\u5224\u65b7\u3002DEFT-RLVR \u5c6c\u65bc\u8a13\u7df4\u8207\u9a57\u8b49\u6846\u67b6\u9805\u76ee\uff0c\u6838\u5fc3\u662f\u628a\u300c\u5148\u770b\u5834\u666f\u4f5c\u6c7a\u5b9a\u300d\u540c\u300c\u4e4b\u5f8c\u518d\u5c0d\u7167\u5019\u9078\u8ecc\u8de1\u300d\u62c6\u958b\uff0c\u6e1b\u5c11 trajectory anchoring bias\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b83\u7684\u505a\u6cd5\u5514\u4fc2\u76f4\u63a5\u751f\u6210\u958b\u653e\u5f0f\u5ea7\u6a19\uff0c\u800c\u662f\u5148\u7528 AD-MCQ \u628a\u898f\u5283\u554f\u984c\u6539\u6210\u591a\u9078\u984c\uff0c\u518d\u7528 DEFT \u5169\u968e\u6bb5\u6d41\u7a0b\u8655\u7406\uff1a\u6a21\u578b\u5148\u505a candidate-blind scene reasoning\uff0c\u4e4b\u5f8c\u5148\u898b\u5230\u5019\u9078\u8ecc\u8de1\u518d\u63c0\u7b54\u6848\u3002\u9019\u7a2e\u8a2d\u8a08\u7684\u597d\u8655\uff0c\u662f\u7b54\u6848\u53ef\u88ab\u7cbe\u78ba\u6838\u5c0d\uff1b\u4ee3\u50f9\u5247\u662f\u4efb\u52d9\u8868\u8ff0\u88ab\u6536\u7a84\u5230\u5019\u9078\u96c6\u5408\u4e4b\u5167\uff0c\u8f03\u63a5\u8fd1\u300c\u53ef\u9a57\u8b49\u6c7a\u7b56\u300d\u800c\u5514\u4fc2\u5b8c\u6574\u8def\u5f91\u751f\u6210\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u91dd\u5c0d\u7684\u4e0d\u662f\u611f\u77e5\u672c\u8eab\uff0c\u800c\u662f\u63a8\u7406\u76e3\u7763\u88ab\u7b54\u6848\u6c61\u67d3\u7684\u554f\u984c<\/li>\n<li>AD-MCQ \u4fdd\u7559 braking\u3001speed \u540c lateral geometry \u7b49\u5dee\u7570\uff0c\u65b9\u4fbf\u7cbe\u78ba\u8a55\u5206<\/li>\n<li>DEFT-RLVR \u7528 GRPO \u806f\u5408\u512a\u5316\uff0c\u4e26\u4e14\u53ef\u9078\u7528 rubric \u76e3\u7763\u63a8\u7406\u904e\u7a0b<\/li>\n<li>RL \u968e\u6bb5\u76f4\u63a5\u7531 base VLM \u958b\u59cb\uff0c\u6bcb\u9808 cold-start SFT<\/li>\n<li>\u6a21\u578b\u63a1\u7528 Qwen3-VL<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u90e8\u7f72\u9580\u6abb\u4e0d\u7b97\u4f4e\u3002\u9805\u76ee\u8981\u6c42\u53e6\u884c\u5b89\u88dd\u652f\u63f4 CUDA \u7684 vLLM\uff0cWaymo codebook reconstruction \u9084\u8981\u984d\u5916\u7528 Python 3.9 \u7684 autovla waymo py39 \u74b0\u5883\uff1b\u9810\u8a2d recipe \u4ea6\u660e\u986f\u504f\u5411\u5927\u578b\u591a GPU \u7bc0\u9ede\uff0c\u5c0f\u578b\u8a2d\u5099\u9700\u8981\u81ea\u884c\u8abf\u6574 tensor parallelism\u3001batch size\u3001sequence length \u540c offloading\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u76ee\u524d\u516c\u958b\u8cc7\u8a0a\u986f\u793a\uff0c\u5b83\u5728 AD-MCQ-500 \u4e0a\u540c\u6642\u63d0\u5347 trajectory selection \u8207 candidate-blind reasoning\uff0c\u8f03\u7a69\u59a5\u7684\u5224\u65b7\u662f\uff1a\u9019\u5957\u65b9\u6cd5\u5c0d\u7814\u7a76\u81ea\u52d5\u99d5\u99db VLM\/VLA \u8a13\u7df4\u53ef\u9760\u6027\u3001\u60f3\u907f\u514d\u300c\u5148\u77e5\u7b54\u6848\u518d\u89e3\u91cb\u300d\u7684\u5718\u968a\u5c24\u5176\u6709\u53c3\u8003\u50f9\u503c\uff1b\u8981\u843d\u5730\u5230\u66f4\u958b\u653e\u7684\u771f\u5be6\u898f\u5283\u6d41\u7a0b\uff0c\u4ecd\u8981\u770b\u5019\u9078\u8ecc\u8de1\u69cb\u5efa\u8207\u7b97\u529b\u6210\u672c\u80fd\u5426\u63a5\u53d7\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/github.com\/hzx122\/DEFT-RLVR\" rel=\"noopener noreferrer\"><strong>GitHub<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u81ea\u52d5\u99d5\u99db Vision-language-action models \u5514\u5c11\u90fd\u6703\u88ab\u6b63\u78ba\u8ecc\u8de1\u300c\u63d0\u793a\u300d\u5230\u7b54\u5571\u3002DEFT-RLVR\u60f3\u8655\u7406\u7684\uff0c\u6b63\u662f\u9019\u7a2e\u770b\u4f3c\u6703\u63a8\u7406\u3001\u5176\u5be6\u5148\u77e5\u9053\u7b54\u6848\u7684\u504f\u5dee\u3002<\/p>\n","protected":false},"author":8,"featured_media":10729,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","footnotes":""},"categories":[133,179,152,119,149,204],"tags":[],"class_list":["post-10730","post","type-post","status-publish","format-standard","hentry","category-133","category-nvidia","category-python","category-119","category-149","category-visionlanguageaction"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/10730","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/users\/8"}],"replies":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/comments?post=10730"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/10730\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/10729"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=10730"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=10730"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=10730"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}