The Year in AI Papers
The year/Independent research

Paper 2604.26904

ClawGym: A Scalable Framework for Building Effective Claw Agents

Fei Bai, Huatong Song, Shuang Sun, Daixuan Cheng, Yike Yang, Chuan Hao, Renyuan Li, Feng Chang, Yuan Wei, Ran Tao, Bryan Dai, Jian Yang, et al.

Published
Apr 2026
Research lab
Independent
Citations
3
GitHub
Not linked

01 In brief

Summary

ClawGym is a scalable framework for developing Claw-style personal agents, addressing the lack of systematic data synthesis, training, and evaluation in such environments.

It introduces ClawGym-SynData, a dataset of 13.5K filtered tasks generated via persona-driven top-down and skill-grounded bottom-up pipelines, with realistic mock workspaces and hybrid verification (code-based and rubric-based).

ClawGym-Agents are trained via supervised fine-tuning on 24.5K high-quality black-box rollout trajectories from OpenClaw, with optional reinforcement learning using sandbox-parallel rollouts.

ClawGym-Bench provides 200 rigorously filtered and human-reviewed benchmark instances.

Experiments show significant improvements: Qwen3-8B improves by 38.90% on PinchBench and 43.46% on ClawGym-Bench, while Qwen3-30B-A3B improves by 54.68% and 25.96%, respectively.

Behavioral analyses reveal key capabilities like tool-use appropriateness, long-horizon robustness, and fine-grained instruction following.

02 From the paper

Abstract

Claw-style environments support multi-step workflows over local files, tools, and persistent workspace states. However, scalable development around these environments remains constrained by the absence of a systematic framework, especially one for synthesizing verifiable training data and integrating it with agent training and diagnostic evaluation. To address this challenge, we present ClawGym, a scalable framework that supports the full lifecycle of Claw-style personal agent development. Concretely, we construct ClawGym-SynData, a diverse dataset of 13.5K filtered tasks synthesized from persona-driven intents and skill-grounded operations, paired with realistic mock workspaces and hybrid verification mechanisms. We then train a family of capable Claw-style models, termed ClawGym-Agents, through supervised fine-tuning on black-box rollout trajectories, and further explore reinforcement learning via a lightweight pipeline that parallelizes rollouts across per-task sandboxes. To support reliable evaluation, we further construct ClawGym-Bench, a benchmark of 200 instances calibrated through automated filtering and human-LLM review. Relevant resources have been released at https://github.com/ClawGym.