PAPOGalaxy/PAPO-G-H-Qwen2.5-VL-3B

VISIONPricing:Input $0.32 / Cached $0.016 / Output $1.6Concurrent Unit Cost:1Model Size:3BQuant:BF16Context Size:32kPublished:Jul 8, 2025License:mitArchitecture:Transformer0.0K Open Weights Featherless Exclusive Cold

PAPOGalaxy/PAPO-G-H-Qwen2.5-VL-3B is a 3 billion parameter vision-language model based on Qwen2.5, developed by PAPOGalaxy. This model integrates Perception-Aware Policy Optimization (PAPO), an extension of GRPO, to enhance multimodal reasoning by improving visual perception. It is specifically designed to reduce perception errors in multimodal tasks, achieving significant improvements on benchmarks with high vision dependency without relying on additional data or external reward models. Its primary use case is robust multi-step reasoning in multimodal contexts, particularly where accurate visual understanding is critical.

Loading preview...

PAPO Model: Perception-Aware Policy Optimization for Multimodal Reasoning

PAPOGalaxy/PAPO-G-H-Qwen2.5-VL-3B is a 3 billion parameter vision-language model that extends the Qwen2.5 architecture with Perception-Aware Policy Optimization (PAPO). This model addresses the limitations of traditional Reinforcement Learning with Verifiable Rewards (RLVR) in multimodal reasoning by focusing on improving visual perception, a common source of error in such tasks. PAPO integrates an Implicit Perception Loss, a KL divergence term, into the GRPO objective, enabling the model to learn perception and reasoning simultaneously from internal supervision.

Key Capabilities

  • Enhanced Multimodal Reasoning: Significantly improves multi-step reasoning abilities in multimodal contexts.
  • Reduced Perception Errors: Achieves a substantial reduction (30.5%) in visual perception errors.
  • Improved Performance on Vision-Dependent Tasks: Yields overall improvements of 4.4% on diverse multimodal benchmarks, with up to 8.0% improvement on tasks highly dependent on visual input.
  • Self-Supervised Perception Learning: Learns to perceive without requiring additional data curation, external reward models, or proprietary models.
  • Robust RL Framework: Introduces a deeper integration of perception-aware supervision into RLVR learning objectives.

Good for

  • Applications requiring robust multimodal reasoning, especially those with complex visual inputs.
  • Research and development in vision-language models and reinforcement learning for multimodal tasks.
  • Scenarios where accurate visual perception is critical for downstream reasoning and decision-making.

Finetunes
2 models