AI Paper Atlas

Research map / Large language models / Reasoning and chain of thought

D-FUSEr: Diverse Failure, Unified Success via Error-Distribution Shaping in LLM Reasoning

David D. Baek, Andrew Estornell, Yichi Zhang, Muhammad Faaiz Taufiq, Jean-Francois Ton, Jie Mei, Tao Wang

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Routing, pruning and distillation for reasoning

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes D-FUSEr, which shapes the error distribution across LLM reasoning samples so that errors are diverse while successes are unified, to improve test-time scaling via majority vote and iterative refinement.

Problem

Test-time scaling methods (majority vote aggregation, self-reflection, multi-agent inference) depend on how errors are distributed across samples; concentrated errors degrade aggregation accuracy and refinement effectiveness.

Approach

Shape the error distribution across sampled solutions: encourage diverse failures and unified successes, rather than only raising raw per-sample accuracy.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add diverse failure reward to distributional-reflection training (ablation)Dist-Reflect (Qwen2.5-14B) · MMLU-Pro (OOD)majority-vote accuracy (%): 70.1 -> 73.5Improved
add diverse failure reward to distributional-reflection training (ablation)Dist-Reflect (Qwen2.5-14B) · RBENCH (OOD)majority-vote accuracy (%): 60.9 -> 65.6Improved
add diverse failure reward to distributional-reflection training (ablation)Dist-Reflect (Qwen2.5-14B) · BBH (OOD)majority-vote accuracy (%): 88.2 -> 91.9Improved
add diverse failure reward to distributional-reflection training (ablation)Dist-Reflect (Qwen2.5-14B) · GPQA (OOD)majority-vote accuracy (%): 49.2 -> 48.9No clear effect
add diverse failure reward to distributional-reflection training (ablation)Dist-Reflect (Llama-3.1-8B) · RBENCH (OOD)majority-vote accuracy (%): 37.9 -> 54.9Improved
add diverse failure reward to distributional-reflection training (ablation)Dist-Reflect (Llama-3.1-8B) · GPQA (OOD)majority-vote accuracy (%): 33.8 -> 45.8Improved
use D-FUSEr instead of GRPO trainingGRPO (Gemma-2-2B) · GPQA (OOD)majority-vote accuracy (%): 29.6 -> 37.7Improved
add theoretical dynamic anti-votes at aggregation (ablation)majority vote (Qwen2.5-3B) · RBENCHmajority-vote accuracy (%): 39.3 -> 48.3Improved
add fixed-strength anti-votes at aggregation (ablation)majority vote (Qwen2.5-7B) · RBENCHmajority-vote accuracy (%): 50.1 -> 50.4No clear effect
add diverse failure reward to distributional-reflection training (ablation)Dist-Reflect (Llama-8B) · BBH (OOD)average accuracy (%): 69.8 -> 62.2Hurt
train with distributional reflection instead of Reflect-GRPO / GRPOGRPO (Qwen-14B) · RBENCH (OOD)majority-vote accuracy (%): 51.7 -> 60.9Improved

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