ClawGym II, a new black-box RL framework, improves Pass@1 by 10-15 points on ClawGym-Bench with Qwen3-30A3B. The unified approach targets stable, scalable training on complex harnesses like OpenClaw and Claude Code.
Key facts
- Improves Pass@1 by 10-15 points on ClawGym-Bench
- Trained with Qwen3-30A3B (30B total, 3B active)
- Targets black-box RL on complex harnesses like OpenClaw and Claude Code
- Framework claims stable and scalable RL for general agents
ClawGym II, announced via @HuggingPapers, presents a unified framework for black-box reinforcement learning of general agents. The key claim: improving Pass@1 by 10-15 points on ClawGym-Bench when training Qwen3-30A3B, a 30-billion-parameter model with 3B active parameters (MoE). The framework is designed to handle complex agent harnesses such as OpenClaw and Claude Code, which are notoriously difficult to train against due to their non-differentiable, tool-heavy environments.
What Black-Box RL Means Here
Black-box RL treats the harness as an opaque function, optimizing the policy purely through reward signals without gradient access to the environment or the harness's internal logic. This contrasts with white-box approaches that rely on differentiable simulators. For agent harnesses like Claude Code, which execute arbitrary code and interact with external APIs, black-box methods are the only viable path. ClawGym II's contribution appears to be in stabilizing this training—addressing the high variance and reward hacking that typically plague RL in such settings.
The 10-15 Point Jump in Context
The reported improvement of 10-15 Pass@1 points on ClawGym-Bench is substantial. Prior RL methods on similar agent benchmarks often struggle to gain more than 5 points due to sparse rewards and unstable policy updates. If replicated, this would place ClawGym II among the top-performing RL frameworks for agentic tasks. The source does not specify the baseline for this improvement, nor does it disclose the training compute or the exact reward shaping used—details that will be critical for independent verification.
Why This Matters
This is not just another RL tweak. The focus on harnesses like Claude Code signals a shift toward training agents directly on production-grade tools, rather than simplified toy environments. If ClawGym II delivers on its stability claims, it could accelerate the development of agents that can be trained on real-world workflows—potentially reducing the need for extensive prompt engineering or fine-tuning on curated datasets.
Key Takeaways
- ClawGym II claims 10-15 Pass@1 point gains on ClawGym-Bench via black-box RL with Qwen3-30A3B, targeting complex agent harnesses.
- Details on baselines and compute remain undisclosed.
What to watch
Watch for a detailed technical report or code release from the ClawGym II authors, which would disclose the baseline, reward design, and compute. The next update could be a public benchmark run on ClawGym-Bench with open weights, allowing independent replication. Also monitor whether OpenAI or Anthropic adopt similar black-box RL for their own agent training pipelines.
