| license: apache-2.0 | |
| base_model: | |
| - Qwen/Qwen2.5-VL-7B-Instruct | |
| pipeline_tag: image-text-to-text | |
| <p align="center"> | |
| <img src="https://github.com/HansenHua/EAPO-ICML26/raw/main/introduction.png" width="90%"></img> | |
| </p> | |
| EAPO (Exploration-Aware Policy Optimization) is a reinforcement learning framework for training agentic large language models to perform adaptive exploration during test-time interaction. Unlike prior methods that apply exploration uniformly across all states, EAPO enables agents to selectively explore only when environmental uncertainty is high, improving long-horizon reasoning and decision making in interactive environments such as GUI control, web navigation, and embodied tasks. | |
| <p align="center"> | |
| <img src="https://github.com/HansenHua/EAPO-ICML26/raw/main/performance.jpg" width="50%"></img> | |
| </p> | |
| Paper: Learning to Explore: Scaling Agentic Reasoning via Exploration-Aware Policy Optimization (https://arxiv.org/abs/2605.08978) | |
| Code: https://github.com/HansenHua/EAPO-ICML26 |