Think, Compress, Answer: Reinforcement Learning for Multimodal Grounding via Reasoning Compression

Key points

• Extended reasoning has substantially improved multimodal large language models (MLLMs) on tasks requiring compositional and relational inference.1
• However, multimodal grounding imposes a distinct requirement: reasoning must ultimately converge on a target-specific compact representation that supports precise localization.1
• TCA is optimized through a multi-stage Group Relative Policy Optimization (GRPO) with progressive rewards that establish the structured output protocol, minimize reasoning-specific information while preserving target-relevant content, and directly improve localization accuracy.1
• Experiments across six multimodal grounding benchmarks show that TCA attains substantial zero-shot performance gains, demonstrating the effectiveness of reasoning compression.1

Sources (1)

• [1]Think, Compress, Answer: Reinforcement Learning for Multimodal Grounding via Reasoning Compression
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 19, 01:34 AM

  Motivated by the information bottleneck principle, we propose Think-Compress-Answer (TCA), a reinforcement learning framework that decomposes multimodal grounding into three phases: free-form reasoning, target-aware context compression, and answer generation conditioned on the compressed context.

  Extended reasoning has substantially improved multimodal large language models (MLLMs) on tasks requiring compositional and relational inference.

Extractive summary: sentences quoted from the sources.