SkillRise

SkillRise is an end-to-end reinforcement learning framework for cross-task skill learning. A single policy plays an ordered sequence of related tasks, alternating between solving the current task with an evolving skill document and curating that document from the resulting trajectory before moving to the next task. Evaluated on ALFWorld, WebShop, and ScienceWorld.

Method

SkillRise (a) constructs progressively challenging task sequences from the same task family, (b) rolls out a single policy that alternates Solve and Curate so the skill document evolves and transfers across tasks, and (c) uses decoupled cross-task credit assignment with group-relative advantages computed only within the same role and position.

Setup

pip install -r requirements.txt   # verl is vendored in verl/, no separate install

Environment backends:

• ALFWorld and WebShop: follow the verl-agent setup.
• ScienceWorld: follow the BEACON setup.

Running

Edit examples/<env>/env.sh (set SKILLRISE_MODEL_PATH, data paths, WANDB_API_KEY), then:

bash examples/skillrise_alfworld/skillrise_alfworld_qwen3_4b.sh   # SkillRise
bash examples/skillrise_alfworld/grpo_alfworld_qwen3_4b.sh        # GRPO baseline

Same layout for skillrise_webshop/ and skillrise_sciworld/. Task sequences are bundled under data/groups/.

Results

SkillRise achieves the best overall results across ALFWorld, WebShop, and ScienceWorld, outperforming prompting-based methods and RL baselines.

License

Apache License 2.0 (derivative of verl / verl-agent). See LICENSE and NOTICE.

🤝 Acknowledgement

This project builds on veRL, verl-agent, BEACON, ALFWorld, WebShop, and ScienceWorld. We thank the authors of those projects.