Reproduce “Symbolic Learning Enables Self-Evolving Agents” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Wangchunshu Zhou, Yixin Ou, Shengwei Ding, Long Li, Jialong Wu, Tiannan Wang, Jiamin Chen, Shuai Wang and 4 more

The method lets an LLM evaluate an agent’s run and use textual feedback to revise its prompts, tools, and pipeline structure. It works backward through recorded steps, generating language-based guidance rather than numeric gradients. On MATH, GPT-3.5 with the method scores 38.8 accuracy versus 22.5 for Agents w/ AutoPE; this is a benchmark result, not evidence of deployed autonomy.

Key insights

  • Holistic, gradient-based optimization of all symbolic components (prompts, tools, pipeline) is more effective for language agents than optimizing isolated components, analogous to end-to-end training in neural networks.
  • Natural language can effectively represent and compute loss, gradients, and component updates within symbolic agent systems, allowing LLMs to serve as meta-optimizers.
  • Agents can achieve continuous self-improvement and adaptation in dynamic environments by learning from experience and optimizing without requiring explicit ground-truth labels for their language loss.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.