# Decomposing Attention To Find Context-Sensitive Neurons

arXiv: https://arxiv.org/abs/2510.03315 · HTML: https://mlsanity.com/paper/2510.03315/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 10 of its 20 references that are in the map.
- First posted on arXiv 1 Oct 2025.

## Builds on (most cited first, 10 of 10)

- [Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://mlsanity.com/paper/1910.10683/) (2019)
- [Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling](https://mlsanity.com/paper/2304.01373/) (2023)
- [Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small](https://mlsanity.com/paper/2211.00593/) (2022)
- [Train Short, Test Long: Attention with Linear Biases Enables Input Length Extrapolation](https://mlsanity.com/paper/2108.12409/) (2021)
- [What Does BERT Look At? An Analysis of BERT's Attention](https://mlsanity.com/paper/1906.04341/) (2019)
- [Datasets: A Community Library for Natural Language Processing](https://mlsanity.com/paper/2109.02846/) (2021)
- [TinyStories: How Small Can Language Models Be and Still Speak Coherent English?](https://mlsanity.com/paper/2305.07759/) (2023)
- [Finding Neurons in a Haystack: Case Studies with Sparse Probing](https://mlsanity.com/paper/2305.01610/) (2023)
- [An Interpretability Illusion for BERT](https://mlsanity.com/paper/2104.07143/) (2021)
- [Analyzing Transformer Dynamics as Movement through Embedding Space](https://mlsanity.com/paper/2308.10874/) (2023)

## Abstract (arXiv)

We study transformer language models, analyzing attention heads whose attention patterns are spread out, and whose attention scores depend weakly on content. We argue that the softmax denominators of these heads are stable when the underlying token distribution is fixed. By sampling softmax denominators from a "calibration text", we can combine together the outputs of multiple such stable heads in the first layer of GPT2-Small, approximating their combined output by a linear summary of the surrounding text. This approximation enables a procedure where from the weights alone - and a single calibration text - we can uncover hundreds of first layer neurons that respond to high-level contextual properties of the surrounding text, including neurons that didn't activate on the calibration text.