Internal Chain-of-Thought: Empirical Evidence for Layer-wise Subtask Scheduling in LLMs

TL;DR

This paper provides empirical evidence that large language models internally decompose complex tasks into subtasks at different layers, sequentially executing them, which enhances understanding and potential control of LLM behavior.

Contribution

It demonstrates the existence of layer-wise subtask decomposition and execution in LLMs through novel methods and analysis, advancing interpretability and control strategies.

Findings

Distinct subtasks are learned at different network depths.

Subtasks are executed sequentially across layers.

Layer-wise execution pattern is consistent across benchmarks.

Abstract

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.