Research map / Vision-language and multimodal / Multimodal LLMs
CR³: Boosting Compositional Reasoning in MLLMs Through Rule-Based Reinforcement Learning
Shun Qian, Bingquan Liu, Chengjie Sun, Peijin Xie, Zhen Xu, Baoxun Wang
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 CR³, a rule-based reinforcement learning approach to boost compositional reasoning in multimodal large language models (MLLMs).
Problem
Compositional reasoning (structured combinations of objects, attributes, and relations) has been studied mainly for VLMs like CLIP and SigLIP, with limited exploration in MLLMs.
Approach
Applies rule-based reinforcement learning to MLLMs to improve compositional reasoning; the abstract is truncated, so the specific reward design and method details are not stated.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply CR3 rule-based RL | Qwen2.5-VL-7B · Compositional avg | average score: 47.2 -> 57.2 | Improved |
| apply CR3 rather than SFT on same data | Qwen2.5-VL-3B · Compositional avg | average score: 29.9 -> 44.6 | Improved |
| apply CR3 | InternVL3-2B · Compositional avg | average score: 24.9 -> 36.4 | Improved |
| apply SFT on compositional data | InternVL3-8B · Compositional avg | average score: 48.8 -> 50.2 | Improved |
| apply SFT on compositional data (out-of-domain) | InternVL3-2B · MMMU | accuracy: 44.7 -> 42.7 | Hurt |
| apply CR3 (out-of-domain) | Qwen2.5-VL-7B · MMMU | accuracy: 46.7 -> 52.0 | Improved |
| add dynamic task mixing instead of uniform mixing (ablation) | CR3 on Qwen2.5-VL-3B · MMVP | accuracy: 38.0 -> 44.7 | Improved |
| add dynamic task mixing instead of fixed-ratio mixing (ablation) | CR3 on Qwen2.5-VL-3B · Winoground | group score: 20.5 -> 27.0 | Improved |
| train on all three tasks instead of TG-VCR only (ablation) | CR3 on Qwen2.5-VL-3B · Winoground | group score: 20.5 -> 27.0 | Improved |
| train on CITM only (ablation) | Qwen2.5-VL-3B · MMVP | accuracy: 26.0 -> 39.3 | Improved |
| train on VG-TCR only (ablation) | Qwen2.5-VL-3B · Winoground | text score: 61.8 -> 65.8 | Improved |
| set format reward weight to 1.0 instead of 0.0 (ablation) | CR3 on Qwen2.5-VL-3B · Winoground | group score: 19.5 -> 27.0 | Improved |
Limitations
- Only the abstract was available; full text unavailable
Closest papers by meaning
- Unveiling the Compositional Ability Gap in Vision-Language Reasoning Model (NeurIPS 2025 · similarity 0.92)
- LENS: Multi-level Evaluation of Multimodal Reasoning with Large Language Models (ICLR 2026 · similarity 0.90)
- Perception-R1: Advancing Multimodal Reasoning Capabilities of MLLMs via Visual Perception Reward (ICLR 2026 · similarity 0.90)
- Incentivizing Versatile Video Reasoning in MLLMs via Data-Efficient Reinforcement Learning (CVPR 2026 · similarity 0.89)
- Learn to Think: Improving Multimodal Reasoning through Vision-Aware Self-Improvement Training (ICML 2026 · similarity 0.89)
- ConMe: Rethinking Evaluation of Compositional Reasoning for Modern VLMs (NeurIPS 2024 · similarity 0.89)
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.
