← 研踪 PaperTrailarXiv:2604.11791 · v2
A Mechanistic Analysis of Looped Reasoning Language Models
首次提交 2026-04-14 01:55
版本更新 2026-11-16 09:23
雷达观测 2026-11-17 17:12
摘要
Reasoning has become a central capability in large language models. Recent research has shown that reasoning performance can be improved by looping an LLM's layers in the latent dimension, resulting in looped reasoning language models. Despite promising results, few works have investigated how the internal dynamics and computation of looped models differ from that of standard feedforward models. In this paper, we conduct a mechanistic analysis of latent states and layer behavior in looped language models. We analyze cyclic recurrence and show that for many of the studied models each layer in the cycle converges to a distinct fixed point; consequently, the recurrent block follows a consistent cyclic trajectory in the latent space. We provide evidence that as these fixed points are reached, attention-head behavior stabilizes, leading to constant behavior - stages of inference - across recurrences. Empirically, we discover that recurrent blocks tend to organize their computation such that it closely mirrors that of feedforward models, repeating this behavior in depth with each iteration. We study how architectural choices influence the emergence and stability of these cyclic fixed points and stages of inference, providing practical guidance for looped language model design.
作者备注:Accepted at NeurIPS 2026
观测到的版本
相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。
v2 · 2026-11-16 09:23 · 最新观测版本
A Mechanistic Analysis of Looped Reasoning Language Models
Reasoning has become a central capability in large language models. Recent research has shown that reasoning performance can be improved by looping an LLM's layers in the latent dimension, resulting in looped reasoning language models. Despite promising results, few works have investigated how the internal dynamics and computation of looped models differ from that of standard feedforward models. In this paper, we conduct a mechanistic analysis of latent states and layer behavior in looped language models. We analyze cyclic recurrence and show that for many of the studied models each layer in the cycle converges to a distinct fixed point; consequently, the recurrent block follows a consistent cyclic trajectory in the latent space. We provide evidence that as these fixed points are reached, attention-head behavior stabilizes, leading to constant behavior - stages of inference - across recurrences. Empirically, we discover that recurrent blocks tend to organize their computation such that it closely mirrors that of feedforward models, repeating this behavior in depth with each iteration. We study how architectural choices influence the emergence and stability of these cyclic fixed points and stages of inference, providing practical guidance for looped language model design.
查看此版本解读 →查看此版本原文 ↗