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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| deactivate keystone neurons (0.0072% of neurons) (ablation) | Llama-3.1-8B-Instruct · MMLU | accuracy: 0.713 -> 0 | Hurt |
| deactivate an equal number of random neurons (ablation) | Llama-3.1-8B-Instruct · MMLU | accuracy: 0.713 -> 0.706 | No clear effect |
| deactivate keystone neurons (ablation) | Qwen3-8B · C4 | perplexity: 15.74 -> 19.53 | Hurt |
| deactivate keystone neurons (ablation) | Llama-3.1-8B-Instruct · C4 | perplexity: 11.78 -> 522 | Hurt |
| deactivate random neurons (ablation) | Qwen3-8B · Math500 | accuracy: 0.942 -> 0.638 | Hurt |
| replace keystone activations with response-mean constants (ablation) | Llama-3.1-8B-Instruct · MMLU | accuracy: 0.713 -> 0 | Hurt |
| replace keystone activations with first-token constants (ablation) | Llama-3.1-8B-Instruct · Math500 | accuracy: 0.496 -> 0 | Hurt |
| restrict SFT updates to keystone neurons instead of full fine-tuning | Llama-3.1-8B-Instruct · MATH500 | accuracy: 31.40 -> 51.20 | Improved |
| restrict SFT updates to keystone neurons instead of full fine-tuning | Llama-3.1-8B-Instruct · MMLU | accuracy: 62.5 -> 69.8 | Improved |
| restrict SFT updates to keystone neurons instead of full fine-tuning | Llama-3.2-1B-Instruct · GSM8K | accuracy: 45.03 -> 50.27 | Improved |
| restrict SFT updates to keystone neurons instead of full fine-tuning | Llama-3.1-8B-Instruct · WildGuard | accuracy: 75.65 -> 82.82 | Improved |
| restrict SFT updates to keystone neurons instead of full fine-tuning | Llama-3.2-1B-Instruct · WildGuard | accuracy: 69.22 -> 73.40 | Improved |
Closest papers by meaning
- How do Large Language Models Handle Multilingualism? (NeurIPS 2024 · similarity 0.89)
- The Emergence of Abstract Thought in Large Language Models Beyond Any Language (NeurIPS 2025 · similarity 0.89)
- LensLLM: Unveiling Fine-Tuning Dynamics for LLM Selection (ICML 2025 · similarity 0.88)
- Remarkable Robustness of LLMs: Stages of Inference? (NeurIPS 2025 · similarity 0.88)
- PaceLLM: Brain-Inspired Large Language Models for Long-Context Understanding (NeurIPS 2025 · similarity 0.88)
- Towards Understanding Fine-Tuning Mechanisms of LLMs via Circuit Analysis (ICML 2025 · similarity 0.87)
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.
