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

arXiv: https://arxiv.org/abs/2505.14530 · HTML: https://mlsanity.com/paper/2505.14530/

## Key facts

- Cited by 9 papers in the map (data through 14 Oct 2026).
- Placed in Prompting and chain-of-thought reasoning, one of 17 directions in LLM reasoning.
- Builds on 18 of its 54 references that are in the map.
- First posted on arXiv 20 May 2025.

## Placed in

- LLM reasoning / [Prompting and chain-of-thought reasoning](https://mlsanity.com/lr/prompting-and-chain-of-thought-reasoning/)

## Builds on (most cited first, 18 of 18)

- [The Llama 3 Herd of Models](https://mlsanity.com/paper/2407.21783/) (2024)
- [Qwen2.5 Technical Report](https://mlsanity.com/paper/2412.15115/) (2024)
- [Mistral 7B](https://mlsanity.com/paper/2310.06825/) (2023)
- [Toy Models of Superposition](https://mlsanity.com/paper/2209.10652/) (2022)
- [In-context Learning and Induction Heads](https://mlsanity.com/paper/2209.11895/) (2022)
- [DeepSeek-V3 Technical Report](https://mlsanity.com/paper/2412.19437/) (2024)
- [Scaling and evaluating sparse autoencoders](https://mlsanity.com/paper/2406.04093/) (2024)
- [Word Translation Without Parallel Data](https://mlsanity.com/paper/1710.04087/) (2017)
- [Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models](https://mlsanity.com/paper/2403.19647/) (2024)
- [Do I Know This Entity? Knowledge Awareness and Hallucinations in Language Models](https://mlsanity.com/paper/2411.14257/) (2024)
- [Polysemanticity and Capacity in Neural Networks](https://mlsanity.com/paper/2210.01892/) (2022)
- [Separating Tongue from Thought: Activation Patching Reveals Language-Agnostic Concept Representations in Transformers](https://mlsanity.com/paper/2411.08745/) (2024)
- [Answer, Assemble, Ace: Understanding How LMs Answer Multiple Choice Questions](https://mlsanity.com/paper/2407.15018/) (2024)
- [IOPO: Empowering LLMs with Complex Instruction Following via Input-Output Preference Optimization](https://mlsanity.com/paper/2411.06208/) (2024)
- [Do Large Language Models Perform Latent Multi-Hop Reasoning without Exploiting Shortcuts?](https://mlsanity.com/paper/2411.16679/) (2024)
- [Back Attention: Understanding and Enhancing Multi-Hop Reasoning in Large Language Models](https://mlsanity.com/paper/2502.10835/) (2025)
- [Task Vectors in In-Context Learning: Emergence, Formation, and Benefit](https://mlsanity.com/paper/2501.09240/) (2025)
- [Learning Task Representations from In-Context Learning](https://mlsanity.com/paper/2502.05390/) (2025)

## Built on by (most cited first, 9 of 9)

- [Reasoning Beyond Language: A Comprehensive Survey on Latent Chain-of-Thought Reasoning](https://mlsanity.com/paper/2505.16782/) (2025)
- [Implicit Reasoning in Large Language Models: A Comprehensive Survey](https://mlsanity.com/paper/2509.02350/) (2025)
- [How Do Language Models Compose Functions?](https://mlsanity.com/paper/2510.01685/) (2025)
- [Catching rationalization in the act: detecting motivated reasoning before and after CoT via activation probing](https://mlsanity.com/paper/2603.17199/) (2026)
- [Punctuation and Predicates in Language Models](https://mlsanity.com/paper/2508.14067/) (2025)
- [Word Recovery in Large Language Models Enables Character-Level Tokenization Robustness](https://mlsanity.com/paper/2603.10771/) (2026)
- [MTA: Multi-Granular Trajectory Alignment for Large Language Model Distillation](https://mlsanity.com/paper/2605.01374/) (2026)
- [Opening the Black Box: A Survey on the Mechanisms of Multi-Step Reasoning in Large Language Models](https://mlsanity.com/paper/2601.14270/) (2026)
- [Transformers Stop Thinking Too Early, and a Tiny LoRA Fixes It](https://mlsanity.com/paper/2609.36585/) (2026)

## Abstract (arXiv)

We show that large language models (LLMs) exhibit an textitinternal 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 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.