AI Paper Atlas

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

ICLR 2026 · Large language models › Language, safety and interpretability · approach: Parallel and early-exit decoding · 2 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
add SAFE (verify-then-ensemble) to UniTEUniTE · MATH500 CoT (Internlm3+Qwen2.5)accuracy: 59.6 -> 77.4Improved
add SAFE to UniTEUniTE · Average over 5 benchmarks (Internlm3+Qwen2.5)accuracy: 75.12 -> 84.20Improved
add SAFE to GaCGaC · Average over 5 benchmarks (Internlm3+Qwen2.5)accuracy: 83.09 -> 83.59Improved
add SAFE to GaCGaC · GSM8K CoT (Internlm3+Qwen2.5)accuracy: 91.28 -> 91.36No clear effect
add SAFE to UniTE for Internlm3+EXAONE3.5UniTE · MMLU-redux/BBH/ARC-C CoTaccuracy: MMLU 72.51 -> 76.08; MATH500 73.6 -> 77.0; still below best single on MMLUMixed
ensemble at every token (UniTE) vs single models under CoTUniTE · MATH500 CoTaccuracy: 72.4/74.8 -> 59.6Hurt
add SAFE to GaC on models with similar tokenizersGaC · MATH500 CoT (Qwen2+Llama3.1)accuracy: 52.4 -> 59.4Improved
add probability sharpening (ablation)UniTE + SAFE · MATH500 CoTaccuracy: 76.6 -> 77.4Improved
add probability sharpening (ablation)GaC + SAFE · MATH500 CoTaccuracy: 75.2 -> 76.0Improved
use geometric mean instead of heuristic sharpening (ablation)UniTE + SAFE · MMLU-redux/MATH500/GSM8K avgaccuracy: 82.42 -> 82.73Improved
raise sharpening threshold lambda from 0.1 to 0.3 (ablation)UniTE + SAFE · Average of 3 benchmarksaccuracy: 82.42 -> 81.90Hurt
increase drafter sequence length from 3 to 5/8 (ablation)UniTE + SAFE · MMLU-redux CoTaccuracy: 77.67 -> 77.81 (n=5) -> 78.31 (n=8)Improved

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