Unveiling the Compositional Ability Gap in Vision-Language Reasoning Model

💡 Motivation

• Can reasoning abilities learned through pure-text training be composed with visual perception to solve multimodal reasoning tasks?

• Can independently acquired visual reasoning skills be integrated to solve tasks that require multiple abilities?

• Can such compositional capabilities generalize to out-of-distribution (OOD) task variants with altered objectives?

📜 Key Observations

• RL-trained models consistently outperform SFT in compositional settings, particularly for cross-task generalization;
• despite high performance on individual tasks, VLMs struggle with compositional reasoning when grounded in multimodal inputs;
• explicitly structuring the reasoning process—by prompting visual description before reasoning (caption-before-thinking) and incorporating intermediate progress rewards leads to significant improvements in compositional performance.

🤗 ComPA Benchmark

• Pure-Text Geometric Reasoning(PT-GR) & Pure-Text Spatial Reasoning(PT-SR)
• Multi-Modal Geometric Reasoning(MM-GR) & Multi-Modal Spatial Reasoning(MM-SR)

🎯 Task Setup Summary

Task Type	Training Subset	Test Subset
Cross-Modal Composition	PT-GR, PT-SR	PT-GR, PT-SR, MM-GR, MM-SR
Cross-Task Composition (Pure-text)	PT-GR, PT-SR	PT-GR, PT-SR, PT-Comp
Cross-Task Composition (Multimodal)	MM-GR, MM-SR	MM-GR, MM-SR, MM-Comp
OOD Composition	MM-GR, MM-SR	MM-GR-OOD, MM-SR-OOD, MM-Comp-OOD

Setup

conda create -n compa python=3.11 
conda activate compa

bash setup.sh

Training

RL-ground for shape area and grid position task

bash src/scripts/run_shape_spatial_rl_ground.sh

RL

bash src/scripts/run_shape_spatial_rl.sh

SFT

bash src/scripts/run_shape_spatial_sft.sh

Evaluation

bash src/scripts/test_grpo_mm_multigpu.sh