Jinhe/ReflectRL-Qwen2.5-3B-Instruct-GRPO

TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:3.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 4, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

ReflectRL-Qwen2.5-3B-Instruct-GRPO is a 3 billion parameter instruction-tuned causal language model developed by Jinhe and collaborators, based on the Qwen2.5-3B-Instruct architecture. This model implements the ReflectRL framework, which enables learning from Golden Negative Trajectories (GNTs) during on-policy post-training by using them as reflective context. It is specifically designed for tasks where learning from failed expert trajectories can improve performance, transitioning from reflective to direct reasoning for inference.

Loading preview...

ReflectRL-Qwen2.5-3B-Instruct-GRPO: Learning from Golden Negative Trajectories

This model, ReflectRL-Qwen2.5-3B-Instruct-GRPO, is a 3 billion parameter instruction-tuned language model built upon the Qwen2.5-3B-Instruct base. Developed by Jinhe and a team of researchers, it introduces the novel ReflectRL framework, detailed in the paper "ReflectRL: Learning from Golden Negative Trajectories via Reflective-to-Direct Reasoning".

Key Capabilities & Features

  • Golden Negative Trajectory (GNT) Learning: ReflectRL is designed to learn effectively from failed expert trajectories, which are typically underutilized in standard training paradigms.
  • Reflective-to-Direct Reasoning: During training, the model uses GNTs as reflective context. This reflective interface is gradually decayed, allowing the final model to perform direct reasoning during inference without requiring special prompts.
  • Lightweight On-Policy Post-Training: The framework is described as lightweight, focusing on efficient on-policy post-training to integrate insights from GNTs.
  • Reflection Advantage: The core motivation is to leverage a "Reflection Advantage" by allocating part of each rollout group to a reflective interface during training.

When to Use This Model

This model is particularly suited for applications where:

  • Learning from examples of what not to do (failed attempts, incorrect paths) is crucial for improving model performance.
  • You need a model that can benefit from a reflective learning process but ultimately operates with standard, direct prompts for inference.
  • You are exploring advanced reinforcement learning from human feedback (RLHF) or similar on-policy post-training methods.

Finetunes
2 models