PAPOGalaxy/PAPO-D-Qwen2.5-VL-3B

VISIONPricing:Input $0.32 / Cached $0.016 / Output $1.6Concurrent Unit Cost:1Model Size:3BQuant:BF16Context Size:32kPublished:Aug 3, 2025License:mitArchitecture:Transformer Open Weights Featherless Exclusive Cold

PAPOGalaxy/PAPO-D-Qwen2.5-VL-3B is a 3 billion parameter multimodal language model based on the Qwen2.5 architecture, specifically designed for multimodal reasoning tasks. It utilizes Perception-Aware Policy Optimization (PAPO), an extension of Reinforcement Learning with Verifiable Rewards (RLVR), to improve visual input perception. This model excels at tasks with high vision dependency by integrating an Implicit Perception Loss, leading to significant reductions in perception errors without requiring additional data curation or external reward models.

Loading preview...

PAPO-D-Qwen2.5-VL-3B: Perception-Aware Multimodal Reasoning

PAPOGalaxy/PAPO-D-Qwen2.5-VL-3B is a 3 billion parameter multimodal language model developed by PAPOGalaxy, specifically designed to enhance reasoning abilities in multimodal contexts. This model addresses the common issue of suboptimal performance in multimodal reasoning tasks due to errors in visual input perception, a limitation often found in traditional Reinforcement Learning with Verifiable Rewards (RLVR) approaches.

Key Capabilities

  • Enhanced Multimodal Reasoning: Improves reasoning by focusing on better perception of visual inputs.
  • Perception-Aware Policy Optimization (PAPO): Extends DAPO with an Implicit Perception Loss, encouraging the model to learn perception and reasoning simultaneously.
  • Reduced Perception Errors: Achieves a substantial reduction (30.5%) in perception errors, particularly in vision-dependent tasks.
  • Internal Supervision: Operates without relying on additional data curation, external reward models, or proprietary models, using internal supervision signals.
  • Robust Performance: Demonstrates significant overall improvements (4.4%) on diverse multimodal benchmarks, with even greater gains (approaching 8.0%) on tasks heavily reliant on vision.

Good for

  • Multimodal AI Applications: Ideal for scenarios requiring robust understanding and reasoning across both text and visual data.
  • Vision-Dependent Tasks: Particularly effective for tasks where accurate visual perception is critical for successful reasoning.
  • Research in RLVR: Provides a foundation for new RL frameworks that integrate visually grounded reasoning, as detailed in the associated research paper: Perception-Aware Policy Optimization for Multimodal Reasoning.
  • Developers seeking efficient multimodal models: Offers strong performance in a 3B parameter model, making it suitable for various applications.

Finetunes
2 models