Skill Self-Play: Pushing the Frontier of LLM Capability with Co-Evolving Skills
The paper proposes “skills” as the verification layer for open-ended LLM self-training.
Skill Self-Play uses a proposer, solver, and dynamic skill controller in a reinforcement-learning loop. The proposer creates harder tasks around sampled skills, while the solver attempts solutions and the controller updates the skill library from execution feedback. The authors say this keeps tasks varied without giving up reliable checks. They report gains on tool-use and reasoning benchmarks, including turnarounds for initially misaligned models. ArXiv · AI/CL/LG's note
Skill Self-Play uses a proposer, solver, and dynamic skill controller in a reinforcement-learning loop. The proposer creates harder tasks around sampled skills, while the solver attempts solutions and the controller updates the skill library from execution feedback. The authors say this keeps tasks varied without giving up reliable checks. They report gains on tool-use and reasoning benchmarks, including turnarounds for initially misaligned models. ArXiv · AI/CL/LG's note
score 6