Dimensions overall score 7.0

GitHub Code Pulse

No public code linked for this paper yet.

• Existing RLVR approaches for MLLMs rely on outcome-driven optimization that updates both perception and reasoning using a shared reward based solely on the final answer.limitation 95%
• The shared reward signal in existing RLVR methods blurs credit assignment, frequently improving reasoning patterns while failing to reliably enhance the accuracy of upstream visual evidence extraction.limitation 90%
• A diagnostic analysis of GRPO on WeMath shows it reduces reasoning errors (from 24.3% to 7.6%) much more than perception errors.result 95%
• The proposed PRCO framework introduces a dual-role RLVR approach with a shared policy, consisting of an Observer role for evidence captioning and a Solver role for answer prediction.method 90%
• PRCO employs role-specific reward signals: the Solver is optimized using verifiable outcome rewards on the final answer, while the Observer receives a utility reward derived from the Solver's downstream success.method 90%
• PRCO yields consistent improvements across model scales by over 7 points on average accuracy compared to the base model.result 85%
• PRCO outperforms prior open-source RL-tuned baselines on eight challenging multimodal reasoning benchmarks.result 85%
• The training does not rely on any existing chain-of-thought data and initiates RL training directly without supervised fine-tuning.method 90%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card