RLVR Fine-Tuning Collapses Solution Diversity; Replay Buffers Fix It

Manoel Ribeiro 团队（Liv Daliberti 领衔）发布 ModeBench 基准，揭示当前 RLVR 微调的系统性缺陷：训练早期遇到的答案被持续强化，形成

2026-10-15 ~ 2026-10-15 · 2 related posts

• RLVR fine-tuning collapses solution space, replay buffer fix boosts diversity and pass@8 — manoelribeiro · 2026-10-15
• ModeBench shows RLVR collapses LLMs into single-solution mode; Re:Max training preserves diversity — manoelribeiro · 2026-10-15