PulseAugur
LIVE 10:35:32
research · [1 source] ·
0
research

New method enhances LLM reasoning diversity without sacrificing stability

Researchers have introduced Expert-Sample, a novel training-free method designed to enhance the performance of fine-grained Mixture-of-Experts (MoE) models. This technique addresses the trade-off between diversity and stability in test-time scaling by analyzing the routing scores of MoE layers. Expert-Sample leverages the observation that MoE routers exhibit a high-confidence 'certain head' and a low-confidence 'uncertain tail', selectively injecting stochasticity into the latter to improve generation diversity without compromising output stability. The method has demonstrated consistent improvements in accuracy and pass@n metrics across various reasoning and coding tasks when evaluated on models like Qwen3-30B-A3B-Instruct. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Introduces a training-free method to improve MoE model diversity and accuracy on reasoning and coding tasks.

RANK_REASON This is a research paper detailing a new method for improving MoE model performance.

Read on arXiv cs.LG →

COVERAGE [1]

  1. arXiv cs.LG TIER_1 · Yuanteng Chen, Peisong Wang, Nanxin Zeng, Yuantian Shao, Shuang Qiu, Gang Li, Jing Liu, Jian Cheng ·

    Certain Head, Uncertain Tail: Expert-Sample for Test-Time Scaling in Fine-Grained MoE

    arXiv:2602.02443v2 Announce Type: replace Abstract: Test-time scaling improves LLM performance by generating multiple candidate solutions, yet token-level sampling requires temperature tuning that trades off diversity against stability. Fine-grained MoE, featuring hundreds of wel…