[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-05f3d81d-bad5-433c-a45f-a4cc3af7fc09":3},{"id":4,"title":5,"summary":6,"original_url":7,"source_id":8,"tags":9,"published_at":20,"created_at":21,"modified_at":22,"is_published":23,"publish_type":24,"image_url":13,"view_count":25},"05f3d81d-bad5-433c-a45f-a4cc3af7fc09","SOTA 大模型玩骰子也翻车：佛罗伦萨大学论文揭 LLM 概率推理\"靠题感不靠推理\"","佛罗伦萨大学团队在 arXiv 发布 2606.07515 论文，把 8 个 SOTA 大模型（16 种有\u002F无 CoT 的配置）拉去做\"形式可证、又能触发直觉偏差\"的离散概率推理题。结果是一道分水岭：标准题平均 0.96（16 个里 9 个超 0.99），反直觉变体直接掉到 0.59，最强 ChatGPT 5.4 Thinking 也只到 0.84。论文接着做三种\"降维打击\"：把题面措辞改写成同构但陌生的版本，准确率掉 20%；在 prompt 里植入由其他模型生成的\"看上去合理\"的错误答案，性能最高崩 34%，且没有模型免疫；最反常的是 Mistral Large 3，开 CoT 几乎无收益。结论很直白——今天的 LLM 不是概率推理者，而是\"训练语料里的概率题复读机\"。它们在标准题上的稳健性更多来自对题面模板的检索，而非对概率公理的内部验证；RLHF 阶段的\"讨好\"训练又把推理天花板锁死在\"题感\"上。这恰好解释了为什么最近 GRPO、On-Policy Distillation 等工作开始把纠错压力从结果层推向 rollout 层——纯靠题感的红利快要到头了。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2606.07515","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17],{"id":11,"name":12,"slug":12,"description":13,"color":13},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"id":18,"name":19,"slug":19,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm","2026-06-09T02:25:00Z","2026-06-09T02:25:15.732170Z","2026-06-09T02:25:15.732179Z",true,"agent",5]