Enregistré dans:
Détails bibliographiques
Auteurs principaux: Yang, Zhipeng, Li, Junzhuo, Xia, Siyu, Hu, Xuming
Format: Preprint
Publié: 2025
Sujets:
Accès en ligne:https://arxiv.org/abs/2505.14530
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866916973897580544
author Yang, Zhipeng
Li, Junzhuo
Xia, Siyu
Hu, Xuming
author_facet Yang, Zhipeng
Li, Junzhuo
Xia, Siyu
Hu, Xuming
contents We show that large language models (LLMs) exhibit an $\textit{internal chain-of-thought}$: they sequentially decompose and execute composite tasks layer-by-layer. Two claims ground our study: (i) distinct subtasks are learned at different network depths, and (ii) these subtasks are executed sequentially across layers. On a benchmark of 15 two-step composite tasks, we employ layer-from context-masking and propose a novel cross-task patching method, confirming (i). To examine claim (ii), we apply LogitLens to decode hidden states, revealing a consistent layerwise execution pattern. We further replicate our analysis on the real-world $\text{TRACE}$ benchmark, observing the same stepwise dynamics. Together, our results enhance LLMs transparency by showing their capacity to internally plan and execute subtasks (or instructions), opening avenues for fine-grained, instruction-level activation steering.
format Preprint
id arxiv_https___arxiv_org_abs_2505_14530
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Internal Chain-of-Thought: Empirical Evidence for Layer-wise Subtask Scheduling in LLMs
Yang, Zhipeng
Li, Junzhuo
Xia, Siyu
Hu, Xuming
Computation and Language
Machine Learning
We show that large language models (LLMs) exhibit an $\textit{internal chain-of-thought}$: they sequentially decompose and execute composite tasks layer-by-layer. Two claims ground our study: (i) distinct subtasks are learned at different network depths, and (ii) these subtasks are executed sequentially across layers. On a benchmark of 15 two-step composite tasks, we employ layer-from context-masking and propose a novel cross-task patching method, confirming (i). To examine claim (ii), we apply LogitLens to decode hidden states, revealing a consistent layerwise execution pattern. We further replicate our analysis on the real-world $\text{TRACE}$ benchmark, observing the same stepwise dynamics. Together, our results enhance LLMs transparency by showing their capacity to internally plan and execute subtasks (or instructions), opening avenues for fine-grained, instruction-level activation steering.
title Internal Chain-of-Thought: Empirical Evidence for Layer-wise Subtask Scheduling in LLMs
topic Computation and Language
Machine Learning
url https://arxiv.org/abs/2505.14530