Research map / Vision-language and multimodal / Multimodal LLMs

Decouple to Generalize: Context-First Self-Evolving Learning for Data-Scarce Vision-Language Reasoning

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual reasoning · 0 citations (OpenAlex)

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.

Proposes DoGe, a two-stage RL loop where a question-masked Thinker generates context analyses scored by a frozen Solver's pass rate, followed by GRPO annealing on original problems, evaluated on Qwen2.5-VL models in chemistry, earth science, and multimodal math.

Problem

RL post-training of VLMs needs abundant high-quality multimodal data, which is scarce in specialized domains (chemistry, earth science, multimodal math); synthetic data and self-rewarding methods tend to converge to narrow distributions and suffer reward hacking and entropy collapse.

Approach

Decouple a single base VLM into a learnable Thinker that analyzes question-masked context and a frozen Solver that answers the original question given that analysis; the Solver's pass rate is the reward for the Thinker, followed by standard GRPO on the original problems, forming an iterative learning-application loop.

Key results

• DoGe-7B (Iter3) reaches an average of 53.2 across seven benchmarks versus 50.9 for the Qwen2.5VL-7B baseline (+2.3 average).
• DoGe-3B (Iter3) averages 48.9 versus 43.2 for Qwen2.5VL-3B (+5.7 average); DoGe-3B Iter1 gains +9.1 on MathVista.
• On the 7B ablation, naive GRPO at Iter2 drops to 46.6 average versus 50.1 for DoGe, with MathVision 18.0 vs 25.3.

Surprising findings

• Naive GRPO outperforms DoGe on MathVista at Iter1 (66.1 vs 62.1) in the ablation table.
• Vision-R1-7B beats DoGe-7B on MathVision (29.0 vs 25.2 best DoGe).

Limitations

• Gains are modest on the 7B model (+2.3 average) and some benchmarks regress relative to baselines or naive GRPO.
• Synthetic question pool is limited and relies on proprietary Gemini models for generation and analysis reports.
• Results from a single run per setting with no variance reporting; claims of stability rest on qualitative curves.

Experimental setup

Datasets: Geometry3k geoqa-plus ChartQA SMolInstruct PubChem NASA EPIC

Benchmarks: MMMU MMStar HallusionBench MathVision MathVista ChemBench MSEarthMCQ

Metrics: accuracy mean@4

Baselines: Qwen2.5VL-7B Qwen2.5VL-3B Vision-R1-7B Visionary-3B InternVL2.5-8B InternVL2.5-2B

Builds on: GRPO DAPO verl Vision-R1

Closest papers by meaning

• MathSE: Improving Multimodal Mathematical Reasoning via Self-Evolving Iterative Reflection and Reward-Guided Fine-Tuning (AAAI 2026 · similarity 0.90)
• VisPlay: Self-Evolving Vision-Language Models (CVPR 2026 · similarity 0.89)
• VL-Rethinker: Incentivizing Self-Reflection of Vision-Language Models with Reinforcement Learning (NeurIPS 2025 · similarity 0.89)
• Agent0-VL: Exploring Self-Evolving Agent for Tool-Integrated Vision-Language Reasoning (ICML 2026 · similarity 0.89)
• Latent Chain-of-Thought for Visual Reasoning (NeurIPS 2025 · similarity 0.88)
• All Roads Lead to Rome: Incentivizing Divergent Thinking in Vision-Language Models (CVPR 2026 · similarity 0.88)