Meta论文：量化推理模型因过度怀疑而失败
Paper from Meta shows Quantized reasoning models often lose because they keep do…
Paper from Meta shows Quantized reasoning models often lose because they keep doubting a correct answer instead of finishing.
Meta的论文显示，量化推理模型常常因为对正确答案反复怀疑而不去完成，从而失败。
Many of them reason well enough, but compression makes them hesitate at the wrong time.
其中许多模型推理能力足够，但压缩使它们在错误时机犹豫不决。
The problem is that post-training quantization, a way to shrink models after training, can make reasoning models cheaper to run but worse at finishing cleanly.
问题在于训练后量化，一种在训练后缩小模型的方法，可以使推理模型运行成本更低，但完成度更差。
The authors found that strong quantization does not only make models less capable, since in many failures the model already reached the right answer but then second-guessed itself.
作者发现，强量化不仅降低模型能力，因为在许多失败案例中，模型已经得出正确答案，但随后自我怀疑。
Their core idea is that quantization adds noise at uncertain word choices, so the model becomes more likely to pick words like “wait,” “but,” or “alternatively” that reopen the problem.
他们的核心观点是，量化在不确定的词汇选择上增加噪声，使模型更可能选择“等等”、“但是”或“或者”等重新打开问题的词汇。
They tested this across math, coding, and science tasks using 5 reasoning models, several quantization methods, and model sizes from 1.5B to 32B.
他们使用5个推理模型、多种量化方法以及从1.5B到32B的模型规模，在数学、编程和科学任务上进行了测试。
The main result is that aggressive quantization raised overthinking failures up to 52%, while a small penalty on 50 hesitation words cut reasoning length by 12% to 23% and often kept or improved accuracy.
主要结果是，激进量化将过度思考失败率提高到52%，而对50个犹豫词汇的小惩罚将推理长度缩短12%至23%，并经常保持或提高准确性。
Given compressed models are widely used to save memory and cost, very important to know that a very small decoding fix can stop many of them from wasting tokens and losing answers they already had.
鉴于压缩模型广泛用于节省内存和成本，非常重要的是，一个非常小的解码修复可以阻止许多模型浪费令牌并丢失已有的答案。
– arxiv. org/abs/2606.00206
– arxiv.org/abs/2606.00206
Title: "Quantized Reasoning Models Think They Need to Think Longer, but They Do Not"
标题：“量化推理模型认为它们需要思考更久，但实际上不需要”
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力