Research map / Large language models / Language, safety and interpretability
Identifying and Analyzing Performance-Critical Tokens in Large Language Models
Yu Bai, Heyan Huang, Cesare Spinoso-Di Piano, Sanxing Chen, Marc-Antoine Rondeau, Yang Gao, Jackie Chi Kit Cheung
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 |
|---|---|---|---|
| restrict test-example attention to template-token representations only (vs zero-shot) (ablation) | Llama 2 7B · AGNews/SST2/TREC/DBPedia avg | accuracy relative change: +31.3 | Improved |
| restrict test-example attention to content-token representations only (vs zero-shot) (ablation) | Llama 2 13B and Mistral 7B · AGNews/SST2/TREC/DBPedia avg | accuracy relative change: -14.1 (Llama 2 13B), -10.0 (Mistral 7B) | Hurt |
| mask template-token representations from test example attention (ablation) | Standard ICL · Llama 2 7B avg | accuracy relative change: -32.8 | Hurt |
| mask content-token representations from test example attention (ablation) | Standard ICL · Llama 2 7B avg | accuracy relative change: -3.8 | No clear effect |
| mask stopword-token representations from test example attention (ablation) | Standard ICL · Llama 2 13B avg | accuracy relative change: -2.9 | Hurt |
| drop content tokens from the prompt (token-level) rather than only masking attention (ablation) | Llama 7B · Avg classification | accuracy (%): 72.8 -> 54.9 | Hurt |
| remove all template tokens from the prompt (ablation) | Llama models · classification datasets | accuracy: to 0% | Hurt |
| replace template tokens with fixed random strings (ablation) | Llama 7B · Avg classification (swap/lexical test) | accuracy (%): 74.8 -> 15.5 | Hurt |
| replace template tokens with fixed random strings (ablation) | OpenLlama 3B · Avg classification | accuracy (%): 58.0 -> 52.6 | Hurt |
| use different random strings per demonstration instead of fixed ones (break repetition) (ablation) | Llama 13B · Avg classification | accuracy (%): 49.3 -> 24.3 | Hurt |
| use different random strings per demonstration instead of fixed ones (break repetition) (ablation) | Llama 33B · Avg classification | accuracy (%): 68.2 -> 64.2 | Hurt |
| swap Tin and Tout template tokens (ablation) | Llama 33B · Avg classification | accuracy (%): 81.6 -> 73.2 | Hurt |
Closest papers by meaning
- Capturing In-Context Learning Dynamics with Task Operators (NeurIPS 2026 · similarity 0.89)
- Understanding In-Context Learning from Repetitions (ICLR 2024 · similarity 0.89)
- Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization (ICLR 2025 · similarity 0.89)
- Where does In-context Learning Happen in Large Language Models? (NeurIPS 2024 · similarity 0.89)
- Revisiting In-context Learning Inference Circuit in Large Language Models (ICLR 2025 · similarity 0.89)
- In-context Vectors: Making In Context Learning More Effective and Controllable Through Latent Space Steering (ICML 2024 · similarity 0.88)
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.
