Research map / Vision-language and multimodal / Multimodal LLMs
Unveiling the Compositional Ability Gap in Vision-Language Reasoning Model
Read the paperCheck your idea against 108,967 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.
Probes whether RL- and SFT-post-trained Qwen2.5-VL models compose skills across modalities and tasks using a synthetic benchmark (ComPABench), and proposes RL-Ground (caption-before-think plus progress reward) to narrow the gap.
Problem
Whether VLMs can inherit LLM-style compositional reasoning from RL with verifiable rewards or SFT, i.e., compose skills learned from text or independent visual tasks across modalities and to OOD task objectives.
Approach
A diagnostic benchmark (ComPABench) with paired pure-text and multimodal synthetic tasks (shape area, grid position, and their composition) plus OOD variants, used to compare SFT, RL, and SFT-init RL; and RL-Ground, which adds a caption block before reasoning and a progress reward on intermediate vision-grounded steps.
Key results
- RL on Qwen2.5-VL-7B reaches 31.2% on multimodal compositional task vs 7.2% for SFT; text-only SFT drops compositional accuracy to 2.2% (7B).
- RL-Ground achieves 52.8% on the 7B compositional task, vs 31.2% for plain multimodal RL.
- Multimodal SFT-init RL fails on compositional task (1.0% for 7B), while RL on OOD largest-area task reaches 95% (7B).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train on individual geometric and spatial skills with RL rather than SFT | Qwen2.5-VL-3B-Instruct · ComPABench pure-text cross-task composition | accuracy: 0.6% -> 93% | Helped |
| Train on individual geometric and spatial skills with RL rather than SFT | Qwen2.5-VL-7B-Instruct · ComPABench pure-text cross-task composition | accuracy: 2.2% -> 81.2% | Helped |
| Use RL rather than SFT on individual skills | Qwen2.5-VL-3B-Instruct · ComPABench multimodal cross-task composition | accuracy: 2.2% -> 17.4% | Helped |
| Initialize RL from an SFT checkpoint | Qwen2.5-VL-7B-Instruct · ComPABench multimodal cross-task composition | accuracy: 31.2% -> 1.0% | Hurt |
| Add caption format to RL (ablation) | Qwen2.5-VL-Instruct-7B · Shape Area | accuracy: 74.6 -> 87.4 | Helped |
| Add caption format to RL (ablation) | Qwen2.5-VL-Instruct-7B · Grid Position | accuracy: 83.2 -> 84.6 | Helped |
| Add caption format to RL (ablation) | Qwen2.5-VL-Instruct-7B · Compositional | accuracy: 31.2 -> 28.0 | Hurt |
| Add progress reward to RL (ablation) | Qwen2.5-VL-Instruct-7B · Compositional | accuracy: 31.2 -> 39.6 | Helped |
| Add progress reward to RL (ablation) | Qwen2.5-VL-Instruct-7B · Grid Position | accuracy: 83.2 -> 85.8 | Helped |
| Add progress reward to RL (ablation) | Qwen2.5-VL-Instruct-7B · Shape Area | accuracy: 74.6 -> 76.0 | Helped |
| Add caption format and progress reward to RL (ablation) | Qwen2.5-VL-Instruct-7B · Compositional | accuracy: 31.2 -> 52.8 | Helped |
| Add caption format and progress reward to RL (ablation) | Qwen2.5-VL-Instruct-7B · Grid Position | accuracy: 83.2 -> 88.4 | Helped |
Surprising findings
- Text-only SFT on individual skills severely degrades compositional accuracy (0.6% for 3B, 2.2% for 7B) below the untrained base model (49.4%, 46.4%).
- Initializing multimodal RL from SFT checkpoints hurts compositional generalization compared with RL from the base model.
Limitations
- Benchmark uses synthetic shape and grid images, leaving a domain gap to natural images.
- Only two reasoning types (geometric and spatial) are covered.
- RL-Ground relies on structured captions and intermediate progress signals that may not be available in open-ended tasks.
Experimental setup
Datasets: ComPABench (synthetic shape area, grid position, area-position composition, pure-text and multimodal variants)
Benchmarks: ComPABench
Metrics: accuracy
Baselines: SFT RL (GRPO) SFT-init RL Qwen2.5-VL-3B-Instruct (no post-training) Qwen2.5-VL-7B-Instruct (no post-training)
Builds on: GRPO DeepSeek-R1 R1-V
Closest papers by meaning
- On Robustness and Chain-of-Thought Consistency of RL-Finetuned VLMs (ICML 2026 · similarity 0.92)
- From Seeing to Thinking: Decoupling Perception and Reasoning Improves Post-Training of Vision-Language Models (ICML 2026 · similarity 0.92)
- Activating Visual Context and Commonsense Reasoning Through Masked Prediction in VLMs (AAAI 2026 · similarity 0.91)
- Diagnosing the Compositional Knowledge of Vision Language Models from a Game-Theoretic View (ICML 2024 · similarity 0.91)
- Compositional Reasoning in Language Models under Reinforcement Learning Post-Training (NeurIPS 2026 · similarity 0.91)
- ConMe: Rethinking Evaluation of Compositional Reasoning for Modern VLMs (NeurIPS 2024 · similarity 0.91)
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.