{"id":9737,"date":"2026-06-28T17:04:15","date_gmt":"2026-06-28T09:04:15","guid":{"rendered":"https:\/\/infernews.com\/blog\/official-code-repository-for-the-paper-quot-hallucination-in-world-models-is-pre\/"},"modified":"2026-06-28T17:04:15","modified_gmt":"2026-06-28T09:04:15","slug":"official-code-repository-for-the-paper-quot-hallucination-in-world-models-is-pre","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/official-code-repository-for-the-paper-quot-hallucination-in-world-models-is-pre\/","title":{"rendered":"MMBench2 \u9ede\u6a23\u9810\u6e2c World Model \u5e7b\u89ba"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/06\/1.jpg\" alt=\"walker run\"><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">MMBench2 \u662f\u4e00\u500b\u570d\u7e5e large generative world models \u7684\u7814\u7a76\u578b\u57fa\u6e96\u8207\u958b\u6e90\u9805\u76ee\uff0c\u7d50\u5408\u8cc7\u6599\u96c6\u3001\u6a21\u578b\u3001\u8a13\u7df4\u8207\u8a55\u6e2c\u7a0b\u5f0f\u3002\u5b83\u4e3b\u8981\u8655\u7406 World Models \u5728\u751f\u6210\u672a\u4f86\u8ecc\u8de1\u6642\u51fa\u73fe hallucination \u7684\u554f\u984c\uff0c\u4e5f\u5c31\u662f\u756b\u9762\u770b\u4f3c\u5408\u7406\uff0c\u4f46\u5df2\u7d93\u504f\u96e2\u771f\u5be6\u52d5\u614b\u8207\u52d5\u4f5c\u689d\u4ef6\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u73fe\u6709\u505a\u6cd5\u591a\u6578\u96c6\u4e2d\u5728\u628a world model \u505a\u5f97\u66f4\u5927\uff0c\u6216\u6cbf\u7528\u56fa\u5b9a\u7684 open-loop rollout \u7bc4\u5f0f\u89c0\u5bdf\u751f\u6210\u6548\u679c\uff1b\u4f5c\u8005\u8a8d\u70ba\u9019\u6a23\u5f88\u96e3\u76f4\u63a5\u627e\u51fa\u6a21\u578b\u4f55\u6642\u958b\u59cb\u5931\u771f\u3002\u9019\u500b\u9805\u76ee\u6539\u4ee5\u300c\u53ef\u9810\u6e2c\u3001\u53ef\u9810\u9632\u300d\u70ba\u6838\u5fc3\uff0c\u63d0\u51fa\u4e09\u7a2e runtime hallucination predictors\uff1atokenizer round-trip residual\u3001flow instability\u3001inter-seed denoising variance\uff0c\u4e26\u914d\u5408 motion-normalized \u7248\u672c\u505a\u5373\u6642\u76e3\u6e2c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6a21\u578b\u8a2d\u8a08\u5927\u81f4\u8ddf\u96a8 Dreamer 4 \u8def\u7dda\uff0c\u4f46\u91cd\u9ede\u4e0d\u53ea\u5728\u67b6\u69cb\u672c\u8eab\uff0c\u800c\u662f\u628a coverage-aware training \u8207 targeted data collection \u653e\u5165\u540c\u4e00\u5957\u6d41\u7a0b\u3002\u4f5c\u8005\u628a hallucination \u8996\u70ba data coverage \u554f\u984c\uff0c\u56e0\u6b64\u6703\u91cd\u62bd\u6a23 under-represented \u7684 state-action space\uff0c\u4ea6\u6703\u7528 predictors \u7576 curiosity reward \u505a closed-loop online data collection\uff0c\u9019\u6bd4\u55ae\u7d14\u52a0\u5927\u6a21\u578b\u66f4\u6709\u65b9\u5411\u6027\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u90e8\u7f72\u7406\u89e3\u4e0a\uff0c\u9019\u500b\u9805\u76ee\u5df2\u63d0\u4f9b\u4e92\u52d5\u5f0f\u7db2\u9801\u4ecb\u9762\uff0c\u53ef\u5728 CUDA GPU \u4e0a\u76f4\u63a5\u555f\u52d5\uff0c\u4e26\u7528 live simulators \u7a2e\u51fa rollout\uff0c\u9023\u5b8c\u6574\u8cc7\u6599\u96c6\u90fd\u5514\u4e00\u5b9a\u8981\u5148\u4e0b\u8f09\u3002\u5b98\u65b9\u4ea6\u516c\u958b 350M-parameter pretrained \u8207 finetuned world models\uff0c\u4ee5\u53ca 427 \u5c0f\u6642\u3001\u6db5\u84cb 210 \u500b continuous control tasks\u300110 \u500b domain \u7684 MMBench2 dataset\uff0c\u65b9\u4fbf\u7814\u7a76\u5718\u968a\u91cd\u505a\u8a13\u7df4\u3001\u6bd4\u8f03\u4e0d\u540c\u8b8a\u9ad4\uff0c\u6216\u8005\u5148\u7528 checkpoint \u6aa2\u67e5 hallucination predictor \u7684\u8868\u73fe\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u9805\u76ee\u6027\u8cea<\/strong>\uff1a\u7814\u7a76\u578b benchmark \u52a0\u5de5\u5177\u93c8\uff0c\u4e0d\u53ea\u662f\u55ae\u4e00\u6a21\u578b<\/li>\n<li><strong>\u6838\u5fc3\u5dee\u7570<\/strong>\uff1a\u628a hallucination \u7576\u6210 coverage \u554f\u984c\uff0c\u800c\u975e\u55ae\u9760\u66f4\u5927\u6a21\u578b\u786c\u63a8<\/li>\n<li><strong>\u53ef\u6e2c\u5167\u5bb9<\/strong>\uff1a\u5373\u6642 predictor \u758a\u5716\u3001\u4e0d\u540c\u6a21\u578b\u8b8a\u9ad4\u3001\u4e92\u52d5 rollout \u5c0d\u7167<\/li>\n<li><strong>\u76f8\u95dc\u6a21\u578b<\/strong>\uff1abase\u3001coverage_aware\u3001combined \u4e09\u985e\u8b8a\u9ad4\uff0c\u4ee5\u53ca 350M-parameter world models<\/li>\n<li><strong>\u9069\u5408\u60c5\u5883<\/strong>\uff1aworld modeling\u3001planning\u3001policy learning\u3001\u6a21\u578b\u5b89\u5168\u6aa2\u67e5<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u9019\u500b\u9805\u76ee\u8f03\u9069\u5408\u7814\u7a76 world models\u3001Robotic \u63a7\u5236\u3001\u6a21\u578b\u53ef\u9760\u6027\u8207\u5b89\u5168\u7684\u5718\u968a\u95b1\u8b80\u548c\u8a66\u9a57\u3002\u5b83\u672a\u5fc5\u662f\u4e00\u822c\u958b\u767c\u8005\u5373\u88dd\u5373\u7528\u7684\u61c9\u7528\u5de5\u5177\uff0c\u4f46\u4f5c\u70ba benchmark\u3001\u5206\u6790\u6846\u67b6\u8207\u8cc7\u6599\u57fa\u790e\u8a2d\u65bd\uff0c\u8fa8\u8b58 hallucination \u6210\u56e0\u8207\u6539\u5584\u65b9\u5411\u90fd\u76f8\u7576\u6e05\u695a\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/www.nicklashansen.com\/mmbench2\/\" rel=\"noopener noreferrer\" target=\"_blank\"><strong>\u9805\u76ee\u4e3b\u9801<\/strong><\/a> \u00b7 <a href=\"https:\/\/github.com\/nicklashansen\/mmbench2\" rel=\"noopener noreferrer\" target=\"_blank\"><strong>GitHub<\/strong><\/a> \u00b7 <a href=\"https:\/\/huggingface.co\/nicklashansen\/mmbench2-models\" rel=\"noopener noreferrer\" target=\"_blank\"><strong>\u6a21\u578b<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>MMBench2\u5514\u53ea\u4fc2\u8cc7\u6599\u96c6\uff0c\u4ea6\u9023\u540c\u6a21\u578b\u3001\u8a55\u6e2c\u540c\u4e92\u52d5\u4ecb\u9762\u4e00\u9f4a\u516c\u958b\u3002\u91cd\u9ede\u5728\u65bc\u9810\u6e2c\u540c\u6e1b\u5c11 World Models \u51fa\u73fe hallucination\u3002<\/p>\n","protected":false},"author":8,"featured_media":9736,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","footnotes":""},"categories":[133,170,76,127,184,186,197,199],"tags":[],"class_list":["post-9737","post","type-post","status-publish","format-standard","hentry","category-133","category-170","category-76","category-127","category-robotic","category-186","category-framework","category-dataset-"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9737","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=9737"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9737\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/9736"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=9737"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=9737"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=9737"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}