Source-linked AI summary
Group-Evolving Agents: Open-Ended Self-Improvement via Experience Sharing
Zhaotian Weng, Antonis Antoniades, Deepak Nathani, Zhen Zhang, Xiao Pu, Xin Eric Wang
TL;DR
Existing self-improving agents are constrained by fixed architectures and isolated evolutionary branches, limiting cumulative use of exploratory experience. GEA evolves groups that share and reuse experience across agents, achieving stronger coding-benchmark performance than open-ended self-evolving baselines and comparable or superior results to human-designed frameworks. Its gains transfer across coding models and support more robust framework improvement, while open-ended exploration introduces alignment, computational, and interpretability concerns.
Problem
Fixed architectures and isolated evolutionary branches limit self-improving agents’ ability to transcend initial designs and convert exploratory diversity into cumulative progress.
Method
GEA treats a group of agents as the evolutionary unit, selecting parent groups and sharing aggregated experience to produce and evaluate same-sized offspring groups.
Results
GEA significantly outperforms state-of-the-art open-ended self-evolving methods on SWE-bench Verified and Polyglot, while matching or surpassing human-designed frameworks.
Takeaways & Limitations
Group-level experience sharing more effectively converts exploratory diversity into sustained progress, with improvements transferring across coding models and supporting stronger robustness.
Takeaways & Limitations
Open-ended exploration may introduce directions misaligned with human intent, consume substantial computational resources, and produce increasingly complex systems that are difficult to understand.
Abstract
from arXiv · showhide
Open-ended self-improving agents can autonomously modify their own structural designs to advance their capabilities and overcome the limits of pre-defined architectures, thus reducing reliance on human intervention. We introduce Group-Evolving Agents (GEA), a new paradigm for open-ended self-improvements, which treats a group of agents as the fundamental evolutionary unit, enabling explicit experience sharing and reuse within the group throughout evolution. Unlike existing open-ended self-evolving paradigms that adopt tree-structured evolution, GEA overcomes the limitation of inefficient utilization of exploratory diversity caused by isolated evolutionary branches. We evaluate GEA on challenging coding benchmarks, where it significantly outperforms state-of-the-art self-evolving methods (71.0% vs. 56.7% on SWE-bench Verified, 88.3% vs. 68.3% on Polyglot) and matches or exceeds top human-designed agent frameworks (71.8% and 52.0% on two benchmarks, respectively). Analysis reveals that GEA more effectively converts early-stage exploratory diversity into sustained, long-term progress, achieving stronger performance under the same number of evolved agents. Furthermore, GEA exhibits consistent transferability across different coding models and greater robustness, fixing framework-level bugs in 1.4 iterations on average, versus 5 for self-evolving methods.
1. Introduction
GEA addresses the limited cumulative value of isolated evolutionary branches by making groups, rather than individual agents, the unit of open-ended evolution. It shares and reuses experience within groups, improving coding performance, diversity consolidation, robustness, and transfer across models.
- Existing self-improving agents struggle to exceed pre-defined architectural limits, leaving progress dependent on continued human intervention.
- Tree-structured self-evolution isolates branches, so exploratory diversity often produces short-lived variants rather than cumulative progress.
- GEA treats a group of agents as the evolutionary unit and enables explicit experience sharing and reuse across group members.
- 71.0% on SWE-bench Verified and 88.3% on Polyglot significantly exceed open-ended self-evolving baselines of 56.7% and 68.3%, respectively.
- GEA consolidates exploratory diversity into sustained progress, with improvements transferring across GPT- and Claude-series models and stronger robustness to framework-level perturbations.
- 71.0% versus 71.8% on SWE-bench Verified and 88.3% versus 52.0% on Polyglot show performance comparable to or exceeding human-designed frameworks.
2. Related Work
Prior work mainly refines individual agents along linear or tree-structured evolutionary paths. Although branching supports diverse exploration, these approaches limit open-ended exploration and the production of novel, learnable artifacts.
- Most self-improving systems iteratively refine a given agent toward a specific objective using linear, chain-based evolution.
- Goal-oriented chain-based evolution inherently limits exploration across diverse directions in open-ended solution spaces.
- Open-ended AI systems are characterized by continuously generating artifacts that are novel, interesting, and learnable from a human perspective.
- Lineage-based tree strategies enable branching exploration and can help avoid local optima, but individual parents independently produce offspring.
3. Method
GEA selects diverse, capable parent groups from an archive and evolves same-sized offspring groups through shared evolutionary experience. Reflection converts aggregated traces into directives and framework-level patches, which are evaluated and archived.
- Method overview: GEA archives all discovered agents and evolves groups through parent-group selection followed by open-ended group evolution.
- Parent Group Selection: Parent selection balances downstream performance with novelty, using task-success representations and cosine dissimilarity to form a top-K group.
- Open-Ended Group Evolution: Unlike independent evolution, group sharing lets agents integrate complementary directions and transform transient diversity into long-term useful experience.
- Open-Ended Group Evolution: The parent group produces an offspring group of the same size, with each agent using its own history and experience aggregated from other members.
- Open-Ended Group Evolution: Evolutionary traces include framework patches, predicted task patches, execution workflows, tool histories, and evaluation outcomes exposing failure modes.
- Open-Ended Group Evolution: Shared experience is analyzed by reflection to generate directives for workflows, tools, or prompts; evolution turns them into framework patches, and acting evaluates the updated agents.
4. Experiments
GEA is evaluated on repository-level software engineering and multilingual code synthesis using staged benchmark protocols and comparisons with self-evolving and human-designed frameworks. The experimental design isolates coding capability, out-of-domain generalization, and the role of experience sharing.
- Benchmarks: GEA is evaluated on SWE-bench Verified and Polyglot, covering repository-level software engineering and multilingual code synthesis.
- Benchmarks: SWE-bench Verified uses staged screening from 10 sanity-check tasks to a 50-task Verified-Mini evaluation for agents demonstrating basic coding functionality.
- Benchmarks: Polyglot reports pass@1 on multilingual algorithmic code generation and uses a separate unseen 50-task medium set as an out-of-distribution test.
- Experimental setup: GEA uses groups of two with four nearest neighbors, running group evolution for 30 SWE-bench iterations and 20 Polyglot iterations.
- Comparisons: The study compares GEA with a strict single-parent tree baseline that prevents experience reuse across evolutionary branches.
- Comparisons: Final GEA performance is also compared against human-designed frameworks, including OpenHands + GPT-5 on SWE-bench Verified.
5. Results and Analysis
GEA outperforms the self-evolving DGM baseline on both coding benchmarks and matches or exceeds human-designed agents. Its gains reflect consolidation of diverse tool and workflow innovations, transfer across coding models, and faster repair of framework-level bugs.
- Main Results: 71.0% vs. 56.7% on SWE-bench Verified and 88.3% vs. 68.3% on Polyglot, GEA outperforms DGM under the reported comparison.GEA improves from 20.0% to 71.0% on SWE-bench Verified and from 38.2% to 88.3% on Polyglot.
- Main Results: 71.0% vs. 71.8% on SWE-bench Verified and 88.3% vs. 52.0% on Polyglot, GEA matches or exceeds human-designed agents.The paper attributes this result to meta-learning without human intervention.
- Evolution Analysis: GEA reaches 88.3% on Polyglot in 4 iterations and 71.0% on SWE-bench Verified in 8 iterations, with benchmark-specific evolutionary patterns.Polyglot uses larger, concentrated patches, while SWE-bench Verified uses smaller, more distributed patches.
- Evolution Analysis: GEA integrates eight of nine key tool functionalities, while DGM integrates five; four tools explored by DGM remain isolated because of lineage isolation.Five of GEA’s integrated tools originated from different parent agents.
- Evolution Analysis: 17 unique ancestors contribute to GEA’s best agent versus 9 for DGM, and GEA’s worst top-5 agent reaches 58.3% versus DGM’s best agent at 56.7%.The comparison indicates broader integration across GEA’s population.
- Transferability: GEA’s improvements transfer across GPT- and Claude-series models because they primarily target agent workflows and tool usage rather than model-specific prompting.The best-discovered agent consistently outperforms the corresponding initial agent on both benchmarks.
- Robustness: 1.4 iterations versus 5, GEA repairs injected framework-level bugs faster than DGM across five independent trials.The repair evaluation tests whether a bug-free group member’s experience can guide repair of a faulty agent.
6. Conclusion
GEA makes a group of agents, rather than an individual, the unit of evolution and enables explicit experience sharing. The resulting consolidation improves performance, transfers across models, and supports faster recovery from framework-level bugs.
- Conclusion: GEA treats a group of agents as the fundamental evolutionary unit, allowing agents to share and reuse complementary evolutionary experiences.This replaces individual-centric evolution with group-level experience integration.
- Conclusion: GEA converts transient exploratory diversity into long-term useful experience and achieves stronger performance with the same number of evolved agents.The conclusion contrasts this consolidation with individual-centric self-evolving approaches.
- Conclusion: GEA’s workflow and tool-use improvements transfer consistently across GPT-series and Claude-series models rather than overfitting to one coding model.The paper identifies agent-level workflow and tool enhancements as the primary source of improvement.
- Conclusion: Group-level experience reuse lets better-performing agents guide repair of faulty agents, enabling recovery from framework-level bugs in fewer evolution iterations.The conclusion presents this as a robustness advantage over individual-centric self-evolving approaches.
Impact Statement
GEA can autonomously modify agent implementations, but open-ended evolution raises alignment, computational-cost, interpretability, and safety concerns. The experiments therefore use isolated sandbox environments, while broader applications remain potential extensions.
- Impact Statement: Open-ended evolution may produce directions misaligned with human intent, consume substantial computational resources, or create increasingly complex systems that are difficult to understand.The paper frames these as inherent considerations of open-ended exploration.
- Impact Statement: All experiments run in isolated sandbox environments, limiting potential impacts on host systems.The sandbox boundary follows the paper’s stated safety measure.
- Impact Statement: Although this work focuses on coding capabilities, the paradigm may have broader applications such as self-improvement for bias mitigation.The paper presents these applications as potential extensions rather than evaluated results.
A.1 Cost Estimate
The primary cost for both GEA and DGM comes from benchmark evaluation, and their overall costs are similar because they generate the same number of agents.
- Cost Estimate: Approximately USD 13,000 per method on SWE-bench and USD 1,500 on Polyglot, the estimated cost of a full run is dominated by benchmark evaluation.The paper states that GEA and DGM have very similar overall costs under the matched-agent setting.
A.2 Case Study
The case study traces the evolutionary trajectories of GEA’s top-three agents on SWE-bench Verified, showing a shared emphasis on workflow and tool-use improvements.
- Across all three top-performing agents, performance-improving patches primarily enhance workflows and tool usage rather than model-specific prompting.