Research map / Vision-language and multimodal / Multimodal LLMs

Play to Generalize: Learning to Reason Through Game Play

ICLR 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.

Post-trains Qwen2.5-VL-7B with rule-based RL (RLOO) on synthetic visual games (Snake, Rotation) and shows zero-shot gains on multimodal math, geometry, CLEVR+, multi-discipline, and VSI-Bench without in-domain data.

Problem

Improving multimodal reasoning in MLLMs without domain-specific reasoning datasets; whether skills learned from gameplay transfer to math, multi-discipline, and spatial benchmarks.

Approach

Use arcade-style visual games (a dual-snake game and a 3D object rotation puzzle) as surrogate RL tasks with rule-based binary rewards; the reasoning skills learned in these games transfer zero-shot to multimodal reasoning benchmarks while preserving general visual performance.

Key results

• ViGaL Snake+Rotation reaches 53.9 average on math benchmarks vs 39.3-52.1 for other 7B RL models, and 64.7 on CLEVR+/multi-discipline average, outperforming MM-Eureka-Qwen-7B (62.8).
• On VSI-Bench, ViGaL reaches 38.7 average, up from 36.7 for the base Qwen2.5-VL-7B and above Visual Jigsaw (38.5).
• On Atari-GPT games, ViGaL achieves 2251 cumulative reward vs 1253 for Qwen2.5-VL-7B; RL on games improves average over base by 12.3 points while SFT decreases it by 1.9 points in ablation.

Surprising findings

• SFT on the same game data degraded math and geometry performance relative to the base model, while RL improved it.
• Random reward labels gave no significant gains over the base model, contrasting with prior findings on text-only math data.

Limitations

• Gains on some benchmarks are modest (e.g., OCR and general perception roughly flat; Breakout and Pong not improved on Atari).
• Pass@k analysis indicates base models overtake RL-trained models at large k, suggesting narrowed reasoning coverage.

Experimental setup

Datasets: SnakeBench (game engine, synthetic) Hunyuan3D-generated 3D meshes rendered to image pairs MMK12 (stage 2 ablation) Atari-GPT benchmark MathVista MathVerse MathVision GeoMath Geometry3K CLEVR-Math Super-CLEVR MMMU

Benchmarks: MathVista MathVerse MathVision GeoMath Geometry3K CLEVR-Math Super-CLEVR MMMU MMMU-Pro VSI-Bench Atari-GPT games MuirBench

Metrics: accuracy win count out of 10 cumulative reward average accuracy pass@k

Baselines: MM-Eureka-Qwen-7B R1-Onevision-7B R1-VL-7B Reason-RFT-Zero-7B VLAA-Thinker-7B OpenVLThinker-7B Video-R1 Visual Jigsaw GPT-4o Gemini-2.0-Flash Qwen2.5-VL-7B

Builds on: DeepSeek-R1 RLOO MM-Eureka SnakeBench Atari-GPT Group Policy Gradient

Closest papers by meaning

• Synthesizing Multimodal Verifiable Game Data to Boost VLMs' General Reasoning (ICLR 2026 · similarity 0.92)
• Revisual-R1: Advancing Multimodal Reasoning From Optimized Cold Start to Staged Reinforcement Learning (ICLR 2026 · similarity 0.88)
• SRPO: Enhancing Multimodal LLM Reasoning via Reflection-Aware Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
• Improving Rationality in the Reasoning Process of Language Models through Self-playing Game (ICML 2025 · similarity 0.88)
• Vision-R1: Incentivizing Reasoning Capability in Multimodal Large Language Models (ICLR 2026 · similarity 0.88)
• Open Vision Reasoner: Transferring Linguistic Cognitive Behavior for Visual Reasoning (NeurIPS 2025 · similarity 0.88)