ExpEvolve: Structuring and Reusing Experience for Open-Ended Discovery

arXiv 2026 53 method

TLDR

ExpEvolve adds idea-level extraction to evolutionary LLM agents, improving sample efficiency and final performance on open-ended math and ML tasks.

Reasoning

The paper presents a clear contribution with theoretical conditions and empirical gains across multiple math tasks and LLMs. However, the abstract lacks discussion of limitations or baseline comparisons, and the available evidence is based on benchmark summaries rather than detailed experimental results.

Read-first score

Read-first score 53, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 64.

Recency 8%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Methodology quality 25%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=analysis,experiment

Topical relevance 42%

53.3

Uses existing LLM keyword relevance scores normalized to 0-100. AI scientist,automated scientific discovery,autonomous research agent,automated research,literature review agent,survey generation,automated experimentation,experiment design agent,AI for scientific research,paper writing agent,research automation,scientific discovery agent

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 93.

Keyword Scores

AI scientist

automated scientific discovery

scientific discovery agent

AI for scientific research

research automation

autonomous research agent

automated research

automated experimentation

experiment design agent

literature review agent

survey generation

paper writing agent

Deep Analysis

Innovations

Introduces idea-level extraction and control into evolutionary coding agents via ExpEvolve.
Discovers and leverages high-level factors from programs that explain scoring differences among programs.
Characterizes theoretical conditions under which discovery with idea factors improves sample efficiency.
Builds a practical framework that uses extracted high-level idea factors to guide exploitation from experience and exploration when discovery plateaus.

Methodology

ExpEvolve extends evolutionary coding agents by extracting high-level idea factors from programs that explain scoring differences, then uses these factors to guide exploitation and exploration. It is evaluated on 11 open-ended mathematical tasks, with transfer experiments on 4 other LLMs and 6 machine learning tasks from MLS-Bench. The work includes theoretical analysis of sample efficiency and empirical comparisons measuring final performance and sample efficiency.

Key Results

ExpEvolve significantly improves final performance and sample efficiency by up to 5.0x across 11 open-ended mathematical tasks. Improvements also transfer to 4 other LLMs and to 6 machine learning tasks from MLS-Bench, with analysis showing better exploration and exploitation via discovered idea-level factors.

Resources

Paper