AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Beyond Mode Collapse: Distribution Matching for Diverse Reasoning

Xiaozhe Li, Yang Li, Xinyu Fang, Shengyuan Ding, Peiji Li, Yongkang Chen, Yichuan Ma, TianYi Lyu, Linyang Li, Dahua Lin, Qipeng Guo, Qingwen Liu and 1 more

ICML 2026 · Reinforcement learning › RL from feedback · approach: GRPO stability and variants · 0 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 group-level distribution-matching loss to GRPO (DMPO)GRPO on Qwen2.5-VL-7B-Instruct · MM-NP-BenchQuality Ratio: 38.4 -> 43.1Improved
add group-level distribution-matching loss to GRPO (DMPO)GRPO on Qwen2.5-VL-7B-Instruct · MM-NP-BenchSuccess Rate: 55.7 -> 61.9Improved
add DMPO to GRPOGRPO on Qwen2.5-Math-7B · NP-Bench (text)Quality Ratio: 40.1 -> 43.9Improved
add DMPO to GRPOGRPO on Qwen2.5-Math-7B · Math average (6 benchmarks)accuracy: 42.5 -> 44.5Improved
train with DMPO instead of GRPO on MM-NP dataQwen2.5-VL-7B-Instruct · 7 OOD visual reasoning benchmarksaverage accuracy: 36.8 -> 37.1Improved
add MM-16K data with DMPODMPO (MM-NP) · 7 OOD visual reasoning benchmarksaverage accuracy: 37.1 -> 39.8Improved
remove GRPO loss and keep only distribution matching (ablation)DMPO · MM-NP-BenchQuality Ratio: 43.1 -> 38.5Hurt
replace MSE with Jensen-Shannon divergence at group level (ablation)DMPO · MM-NP-BenchQuality Ratio: 43.1 -> 38.6Hurt
pure distribution matching without GRPO versus GRPO (ablation)GRPO · MM-NP-BenchQuality Ratio: 38.4 -> 38.5No clear effect
vary lambda and alpha outside optimal ranges (ablation)DMPO · MM-NP-BenchQuality RatioMixed
learned partition function (FlowRL) instead of group-level Boltzmann targetdiversity-focused RL · MM-NP-BenchQuality Ratio: 43.1 -> 38.9Hurt
optimize pass@k reward instead of DMPOGRPO · MM-NP-BenchQuality Ratio: 38.4 -> 34.8Hurt

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

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