SESA: Self-Evolving Search Agents with Procedural Memory
A recent study published on arXiv (2607.29468) presents SESA (Self-Evolving Skill-Augmented Agent), a novel framework that incorporates procedural memory into self-play with tool-enhanced search. Unlike conventional self-play agents that create training scenarios without a lasting state, SESA features a dynamic skill memory that influences subsequent training. The framework includes a challenger that generates problems and a distinct solver that accesses skills. Failures are transformed into reusable skills and added back to memory, modifying the solver's performance and success, which in turn affects the challenger’s rewards and the types of problems presented. This reciprocal relationship fosters the co-evolution of task generation and skill memory, paving the way for new challenges that continuously update memory. The authors of the paper are researchers, highlighting advancements in AI and machine learning.
Key facts
- SESA stands for Self-Evolving Skill-Augmented Agent.
- It combines self-play with procedural memory.
- The challenger poses problems, the solver retrieves skills.
- Failures are distilled into reusable skills and stored in memory.
- Memory updates affect solver behavior and problem distribution.
- The loop allows co-evolution of task generation and skill memory.
- The paper is available on arXiv with ID 2607.29468.
- It addresses limitations of fixed task distributions in skill learning.
Entities
Institutions
- arXiv