Research map / Large language models / Language, safety and interpretability
When to Ensemble: Identifying Token-Level Points for Stable and Fast LLM Ensembling
Heecheol Yun, Kwangmin Ki, Jung Hyun Lee, Eunho Yang
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 |
|---|---|---|---|
| add SAFE (verify-then-ensemble) to UniTE | UniTE · MATH500 CoT (Internlm3+Qwen2.5) | accuracy: 59.6 -> 77.4 | Improved |
| add SAFE to UniTE | UniTE · Average over 5 benchmarks (Internlm3+Qwen2.5) | accuracy: 75.12 -> 84.20 | Improved |
| add SAFE to GaC | GaC · Average over 5 benchmarks (Internlm3+Qwen2.5) | accuracy: 83.09 -> 83.59 | Improved |
| add SAFE to GaC | GaC · GSM8K CoT (Internlm3+Qwen2.5) | accuracy: 91.28 -> 91.36 | No clear effect |
| add SAFE to UniTE for Internlm3+EXAONE3.5 | UniTE · MMLU-redux/BBH/ARC-C CoT | accuracy: MMLU 72.51 -> 76.08; MATH500 73.6 -> 77.0; still below best single on MMLU | Mixed |
| ensemble at every token (UniTE) vs single models under CoT | UniTE · MATH500 CoT | accuracy: 72.4/74.8 -> 59.6 | Hurt |
| add SAFE to GaC on models with similar tokenizers | GaC · MATH500 CoT (Qwen2+Llama3.1) | accuracy: 52.4 -> 59.4 | Improved |
| add probability sharpening (ablation) | UniTE + SAFE · MATH500 CoT | accuracy: 76.6 -> 77.4 | Improved |
| add probability sharpening (ablation) | GaC + SAFE · MATH500 CoT | accuracy: 75.2 -> 76.0 | Improved |
| use geometric mean instead of heuristic sharpening (ablation) | UniTE + SAFE · MMLU-redux/MATH500/GSM8K avg | accuracy: 82.42 -> 82.73 | Improved |
| raise sharpening threshold lambda from 0.1 to 0.3 (ablation) | UniTE + SAFE · Average of 3 benchmarks | accuracy: 82.42 -> 81.90 | Hurt |
| increase drafter sequence length from 3 to 5/8 (ablation) | UniTE + SAFE · MMLU-redux CoT | accuracy: 77.67 -> 77.81 (n=5) -> 78.31 (n=8) | Improved |
Closest papers by meaning
- Rethinking LLM Ensembling from the Perspective of Mixture Models (ICML 2026 · similarity 0.94)
- Determine-Then-Ensemble: Necessity of Top-k Union for Large Language Model Ensembling (ICLR 2025 · similarity 0.93)
- Ensembling Language Models with Sequential Monte Carlo (NeurIPS 2026 · similarity 0.92)
- Ensemble Learning for Heterogeneous Large Language Models with Deep Parallel Collaboration (NeurIPS 2024 · similarity 0.89)
- SpecEM: Training-Free LLM Ensembling via Iterative Drafting, Verification, and Online Feedback (NeurIPS 2025 · similarity 0.89)
- A Probabilistic Perspective on Unlearning and Alignment for Large Language Models (ICLR 2025 · similarity 0.89)
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.
