SkillRise is a reinforcement learning framework that trains a single policy to solve a sequence of related tasks by alternating between solving tasks with an evolving skill document and curating that document from trajectories.
5
No data
0
0
Apache-2.0
2026-07-29
It introduces decoupled cross-task credit assignment and achieves state-of-the-art results across three distinct benchmarks (ALFWorld, WebShop, ScienceWorld), outperforming prompting methods and RL baselines.
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.
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.
pip install -r requirements.txt # verl is vendored in verl/, no separate install
Environment backends:
Edit examples//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/.
SkillRise achieves the best overall results across ALFWorld, WebShop, and ScienceWorld, outperforming prompting-based methods and RL baselines.
Apache License 2.0 (derivative of verl / verl-agent). See LICENSE and
NOTICE.
This project builds on veRL, verl-agent, BEACON, ALFWorld, WebShop, and ScienceWorld. We thank the authors of those projects.
Similar projects matched by category, topics, and programming language.
Harness Engineering is a methodology for improving coding agent outputs by carefully crafting the environment around them—providing curated context, tools, and executable constraints that encode an organization’s nonfunctional requirements and cumulative lessons.
A 28.9 million parameter language model runs on an $8 ESP32-S3 microcontroller entirely on-device, generating simple stories at about 9.5 tokens per second.
mimic captures traffic from any iOS or web app and automatically generates a Python client library that lets you call the app's API like a regular library.