arXiv：cs.AI· Pawel Olszowiec, Paulina Tomaszewska, Maciej \.Zelaszczyk, Michal Byra, Neo Christopher Chung·· 16 小时前
循环语言模型的潜在推理解读：八款模型五视角机制可解释性分析
Monitoring latent reasoning in looped language models
AI 导读
研究者构建了一套基于机制可解释性的检查协议，用五个视角审视来自四个家族的八款循环语言模型，区分提示词仅在循环状态内传递的 carry loops（Ouro、LoopUS）与每次迭代重新输入的 re-read loops（Parcae，以及 Raven 家族的 Rec-OLMo 与 Huginn）。
正文
Abstract:Looped language models (LMs) reuse one block of layers several times per token generation, promising more latent computation at no additional storage cost. Despite their rising popularity, the analysis of looped LMs is still in its infancy: little is known about whether the iterative transformations actually improve their capabilities, or whether these models remain stable under prompt perturbation. We construct an inspection protocol rooted in Mechanistic Interpretability comprising five diverse lenses. We apply it to examine eight looped LMs from four families. Models differ in many aspects but the key one is whether the prompt is transmitted only within the looped state (carry loops: Ouro, LoopUS) or re-supplied at every iteration (re-read loops: Parcae, and Rec-OLMo & Huginn from the Raven family). The results of our study provide valuable insights concerning safety and efficient use of looped language models.
| Comments: | 39 pages, 25 figures, 11 tables |
| Subjects: | Artificial Intelligence (cs.AI); Computation and Language (cs.CL) |
| ACM classes: | I.2.0; I.2.7 |
| Cite as: | arXiv:2610.18127 [cs.AI] |
| (or arXiv:2610.18127v1 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2610.18127 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Pawel Olszowiec [view email]
[v1]
Tue, 13 Oct 2026 10:39:05 UTC (1,501 KB)
来源：arXiv：cs.AI · arxiv.org