Proxy Exploration and Reusable Guidance: A Modular LLM Post-Training Paradigm via Proxy-Guided Update Signals
Paper • 2607.11505 • Published • 11
📦 Model Weights
Pre-trained GRPO checkpoints are available on Huggingface:
| Checkpoint | Role | Training |
|---|---|---|
Qwen3-1.7B-Math-GRPO-Steps500 |
Proxy | DeepMath-103K · GRPO · 500 steps |
Qwen3-1.7B-Math-GRPO-Steps800 |
Proxy | DeepMath-103K · GRPO · 800 steps |
Qwen3-8B-Math-GRPO-Steps400 |
Primary | DeepMath-103K · GRPO · 400 steps |
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support