跳转至

文章背景与核心概要

符号回归(Symbolic Regression)是一项从数据中发现闭式数学方程的经典科学与机器学习任务。传统由大语言模型(LLM)引导的方法通常采用僵化的提议循环,将异构的搜索失败压缩为简单的标量分数和单一提示词,从而限制了探索的深度与鲁棒性。

为了突破这些局限,本文作者推出了 A-SR(Agentic Symbolic Regression),这是一个自引化的智能体框架。该框架将主要的控制单元从简单的表达式编辑转变为基于角色的证据视图(role-conditioned evidence views)。A-SR 通过分层协同机制,在专用协同协议、在线评估器-奖励角色策略以及状态路由的进程内存之间动态分配任务;利用细粒度的评估器反馈刻画可靠性与生产力,并动态更新角色级效用;同时实现了“运行内(Within a run)”不改动权重的动态自适应与“跨运行(Across runs)”的搜索轨迹蒸馏。

在性能方面,A-SR 在多个科学基准测试和真实世界科学发现任务中展现出压倒性的优势,显著提升了开源大模型在复杂数学表达式发现方面的准确率与泛化能力。


A-SR: Self-Evolving Agentic LLMs for Symbolic Regression via Hierarchical Coordination

arXiv: 2608.04872 [cs.CL]
Authors: Wenxiao Zhao, Dong Liu, Kaiyi Xu, Feng Liu, Zhen Zhao, Fei Ben, Shu Wang, Wenhao Li, Ying Nian Wu, Fenghua Ling, Haobo Li, Lei Bai
Submitted: August 5, 2026 (v1); Last revised: August 6, 2026 (v2)


📌 Summary

符号回归是从数据中发现闭式数学方程的任务。传统由大语言模型(LLM)引导的方法通常采用僵化的提议循环,将异构的搜索失败压缩为基本的标量分数和单一的提示词。

为了克服这些局限,作者引入了 A-SR,这是一个自进化的智能体框架,它将主要的控制单元从简单的表达式编辑转变为基于角色的证据视图

该框架的核心亮点包括: * 分层协同(Hierarchical Coordination): 通过在专用协同协议、在线评估器-奖励角色策略和状态路由的进程内存之间动态路由任务,来协同公式的发现。 * 细粒度评估器反馈(Granular Evaluator Feedback): 刻画可靠性与生产力,更新角色级效用,并将优秀的模体(motifs)、失败轨迹和有效性诊断路由至不同的专用智能体。 * 双时间尺度自进化(Two-Timescale Self-Evolution): * 运行内(Within a run): 在不修改 LLM 权重的情况下动态适应搜索过程。 * 跨运行(Across runs): 将记录的搜索轨迹蒸馏到开源 LLM 中,作为基于角色的提议先验。

Symbolic regression is the task of discovering closed-form mathematical equations from data. Traditional Large Language Model (LLM)-guided methods often utilize a rigid proposal loop that compresses heterogeneous search failures into a basic scalar score and a single prompt.

To overcome these limitations, the authors introduce A-SR, a self-evolving agentic framework that shifts the primary control unit from simple expression edits to role-conditioned evidence views.

Key highlights of the framework include: * Hierarchical Coordination: Coordinates formula discovery by dynamically routing tasks among specialized coordination protocols, an online evaluator-reward role policy, and state-routed process memory. * Granular Evaluator Feedback: Characterizes reliability and productivity, updates role-level utilities, and routes elite motifs, failure traces, and validity diagnostics to distinct specialized agents. * Two-Timescale Self-Evolution: * Within a run: Adapts the search process dynamically without modifying LLM weights. * Across runs: Distills recorded search trajectories into open-source LLMs to serve as role-conditioned proposal priors.


📊 Performance Highlights

  • LLM-SRBench (LSR-Synth Domains): 在四个科学领域取平均,使用 Llama-3.1-8B 时,A-SR 将 Acc@0.0125.79% 提升至 48.30%。同时,轻量级的 A-SR-LoRA 变体将 Qwen3-4B 的性能从 24.58% 提升至 38.29%
  • 真实世界科学发现(Real-World Scientific Discovery): 在四个复杂的真实世界科学任务中,A-SR 在汇报的 8 个指标中的 7 个 上实现了最佳的分布内或分布外归一化均方误差。
  • LLM-SRBench (LSR-Synth Domains): Averaged across four scientific domains, A-SR improves Acc@0.01 from 25.79% to 48.30% using Llama-3.1-8B. Meanwhile, the lightweight A-SR-LoRA variant improves Qwen3-4B performance from 24.58% to 38.29%.
  • Real-World Scientific Discovery: On four complex real-world scientific tasks, A-SR achieves the best in-distribution or out-of-distribution normalized mean squared error across 7 out of 8 reported metrics.


(Note: Original license icon preserved below per formatting requirements)
license icon view license