AI Paper Atlas

Research map / Large language models / Instruction and fine-tuning

Tiny Brains, Giant Impact: Uncovering the Keystone Neurons of LLM with Just a Few Prompts

Xiangtian Ji, Yuxin Chen, Zhengzhou Cai, Xiang Wang, An Zhang, Tat-Seng Chua

ICML 2026 · Large language models › Instruction and fine-tuning · approach: Efficient LLM fine-tuning systems · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
deactivate keystone neurons (0.0072% of neurons) (ablation)Llama-3.1-8B-Instruct · MMLUaccuracy: 0.713 -> 0Hurt
deactivate an equal number of random neurons (ablation)Llama-3.1-8B-Instruct · MMLUaccuracy: 0.713 -> 0.706No clear effect
deactivate keystone neurons (ablation)Qwen3-8B · C4perplexity: 15.74 -> 19.53Hurt
deactivate keystone neurons (ablation)Llama-3.1-8B-Instruct · C4perplexity: 11.78 -> 522Hurt
deactivate random neurons (ablation)Qwen3-8B · Math500accuracy: 0.942 -> 0.638Hurt
replace keystone activations with response-mean constants (ablation)Llama-3.1-8B-Instruct · MMLUaccuracy: 0.713 -> 0Hurt
replace keystone activations with first-token constants (ablation)Llama-3.1-8B-Instruct · Math500accuracy: 0.496 -> 0Hurt
restrict SFT updates to keystone neurons instead of full fine-tuningLlama-3.1-8B-Instruct · MATH500accuracy: 31.40 -> 51.20Improved
restrict SFT updates to keystone neurons instead of full fine-tuningLlama-3.1-8B-Instruct · MMLUaccuracy: 62.5 -> 69.8Improved
restrict SFT updates to keystone neurons instead of full fine-tuningLlama-3.2-1B-Instruct · GSM8Kaccuracy: 45.03 -> 50.27Improved
restrict SFT updates to keystone neurons instead of full fine-tuningLlama-3.1-8B-Instruct · WildGuardaccuracy: 75.65 -> 82.82Improved
restrict SFT updates to keystone neurons instead of full fine-tuningLlama-3.2-1B-Instruct · WildGuardaccuracy: 69.22 -> 73.40Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy