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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add group-level distribution-matching loss to GRPO (DMPO) | GRPO on Qwen2.5-VL-7B-Instruct · MM-NP-Bench | Quality Ratio: 38.4 -> 43.1 | Improved |
| add group-level distribution-matching loss to GRPO (DMPO) | GRPO on Qwen2.5-VL-7B-Instruct · MM-NP-Bench | Success Rate: 55.7 -> 61.9 | Improved |
| add DMPO to GRPO | GRPO on Qwen2.5-Math-7B · NP-Bench (text) | Quality Ratio: 40.1 -> 43.9 | Improved |
| add DMPO to GRPO | GRPO on Qwen2.5-Math-7B · Math average (6 benchmarks) | accuracy: 42.5 -> 44.5 | Improved |
| train with DMPO instead of GRPO on MM-NP data | Qwen2.5-VL-7B-Instruct · 7 OOD visual reasoning benchmarks | average accuracy: 36.8 -> 37.1 | Improved |
| add MM-16K data with DMPO | DMPO (MM-NP) · 7 OOD visual reasoning benchmarks | average accuracy: 37.1 -> 39.8 | Improved |
| remove GRPO loss and keep only distribution matching (ablation) | DMPO · MM-NP-Bench | Quality Ratio: 43.1 -> 38.5 | Hurt |
| replace MSE with Jensen-Shannon divergence at group level (ablation) | DMPO · MM-NP-Bench | Quality Ratio: 43.1 -> 38.6 | Hurt |
| pure distribution matching without GRPO versus GRPO (ablation) | GRPO · MM-NP-Bench | Quality Ratio: 38.4 -> 38.5 | No clear effect |
| vary lambda and alpha outside optimal ranges (ablation) | DMPO · MM-NP-Bench | Quality Ratio | Mixed |
| learned partition function (FlowRL) instead of group-level Boltzmann target | diversity-focused RL · MM-NP-Bench | Quality Ratio: 43.1 -> 38.9 | Hurt |
| optimize pass@k reward instead of DMPO | GRPO · MM-NP-Bench | Quality Ratio: 38.4 -> 34.8 | Hurt |
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.
- FlowRL — FlowRL: Matching Reward Distributions for LLM Reasoning (ICLR 2026) · better on 1 of 1 shared result
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
- Whatever Remains Must Be True: Filtering Drives Reasoning in LLMs, Shaping Diversity (ICLR 2026 · similarity 0.91)
- DGPO: Discovering Multiple Strategies with Diversity-Guided Policy Optimization (AAAI 2024 · similarity 0.90)
- SetPO: Set-Level Policy Optimization for Diversity-Preserving LLM Reasoning (ICML 2026 · similarity 0.90)
- TMPO: Trajectory Matching Policy Optimization for Diverse and Efficient Diffusion Alignment (NeurIPS 2026 · similarity 0.89)
- Mode-conditioning unlocks superior test-time compute scaling (ICLR 2026 · similarity 0.89)
- STARE: Surprisal-Guided Token-Level Advantage Reweighting for Policy Entropy Stability (NeurIPS 2026 · similarity 0.88)
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.
