arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2604.06374cs.CLcs.LG
收藏

叠加的幻觉?语言模型中潜在思维的原理性分析

The Illusion of Superposition? A Principled Analysis of Latent Thinking in Language Models

Mila – Quebec AI Institute(Mila – 魁北克人工智能研究所);Université de Montréal(蒙特利尔大学);McGill University(麦吉尔大学)

Michael Rizvi-Martel, Guillaume Rabusseau, Marius Mosbach

AI总结:

本文探讨了连续链式思维(Latent CoT)中叠加现象的出现条件,发现仅从头训练的模型表现出叠加特性,而其他训练方式下模型倾向于使用捷径解法。

摘要:

通过连续链式思维(Latent CoT)进行潜在推理已成为替代离散Co T推理的有前景方法。在连续空间中操作增强了表达能力,并被假设能够实现叠加:在单一表示中同时维持多个候选解决方案的能力。尽管有理论论证,仍不清楚语言模型在使用Latent CoT推理时是否实际利用叠加。我们跨三个领域进行研究:无训练领域通过令牌嵌入的凸组合构建潜在思维,微调领域中基模型被调整以生成潜在思维,以及从头训练领域中模型完全使用潜在思维解决给定任务。通过Logit Lens和实体级探测分析内部表示,发现仅从头训练的模型显示出使用叠加的迹象。在无训练和微调领域,叠加要么崩溃要么未被使用,模型倾向于发现捷径解法。我们认为这归因于两个互补现象:i)在自然语言数据上预训练使模型倾向于最后层的令牌;ii)容量对模型偏好解法有巨大影响。共同,我们的结果为连续链式思维推理中何时以及为何出现叠加提供了统一解释,并识别了叠加崩溃的条件。

英文摘要:

Latent reasoning via continuous chain-of-thoughts (Latent CoT) has emerged as a promising alternative to discrete CoT reasoning. Operating in continuous space increases expressivity and has been hypothesized to enable superposition: the ability to maintain multiple candidate solutions simultaneously within a single representation. Despite theoretical arguments, it remains unclear whether language models actually leverage superposition when reasoning using latent CoTs. We investigate this question across three regimes: a training-free regime that constructs latent thoughts as convex combinations of token embeddings, a fine-tuned regime where a base model is adapted to produce latent thoughts, and a from-scratch regime where a model is trained entirely with latent thoughts to solve a given task. Using Logit Lens and entity-level probing to analyze internal representations, we find that only models trained from scratch exhibit signs of using superposition. In the training-free and fine-tuned regimes, we find that the superposition either collapses or is not used at all, with models discovering shortcut solutions instead. We argue that this is due to two complementary phenomena: i) pretraining on natural language data biases models to commit to a token in the last layers ii) capacity has a huge effect on which solutions a model favors. Together, our results offer a unified explanation for when and why superposition arises in continuous chain-of-thought reasoning, and identify the conditions under which it collapses.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑