A-SR: Self-Evolving Agentic LLMs for Symbolic Regression via Hierarchical Coordination
A recent preprint on arXiv (2608.04872) presents A-SR, an innovative framework that enhances symbolic regression by transferring control from modifying expressions to utilizing role-conditioned evidence views. In contrast to current LLM-guided approaches, A-SR facilitates formula discovery through a combination of protocol routing, an online evaluator-reward system, and state-routed memory. Feedback from evaluators refines role-level utilities and directs motifs and diagnostics to agents. A-SR evolves autonomously at two distinct timescales: it adjusts the search process during a run and distills the recorded trajectories into open-source LLMs across different runs. This framework aims to boost efficiency and precision in symbolic regression, applicable in domains such as physics, biology, and economics. The paper indicates potential advancements in AI-driven scientific exploration. The release date remains unspecified.
Key facts
- A-SR is a self-evolving agentic framework for symbolic regression.
- It shifts control from expression edits to role-conditioned evidence views.
- It coordinates formula discovery through routing among coordination protocols.
- It uses an online evaluator-reward role policy and state-routed process memory.
- Evaluator feedback characterizes reliability and productivity, updates role-level utilities, and routes elite motifs, failure traces, and validity diagnostics to different agents.
- The framework self-evolves at two timescales: within a run and across runs.
- Within a run, it adapts the search process without updating LLM parameters.
- Across runs, recorded trajectories can be distilled into open-source LLMs as role-conditioned proposal priors.
- The paper is available on arXiv with identifier 2608.04872.
- The announcement type is 'cross'.
Entities
Institutions
- arXiv