CORAL: Towards Autonomous Multi-Agent Evolution for Open-Ended Discovery

Summary

CORAL is a framework for autonomous multi-agent evolution on open-ended problems. Where existing LLM-based evolution relies on fixed heuristics and hard-coded exploration rules, CORAL uses long-running agents that explore, reflect, and collaborate through shared persistent memory, asynchronous multi-agent execution, and heartbeat-based interventions, with safeguards including isolated workspaces, evaluator separation, resource management, and agent session and health management. On mathematical, algorithmic, and systems optimization tasks it sets new state-of-the-art results on 10 tasks, with 3-10 times higher improvement rates and far fewer evaluations than fixed evolutionary search baselines. On Anthropic's kernel engineering task, four co-evolving agents improve the best known score from 1363 to 1103 cycles. Mechanistic analyses attribute the gains to knowledge reuse and to multi-agent exploration and communication, which the authors read as evidence that greater agent autonomy and multi-agent evolution improve open-ended discovery.

Contribution

Claimed: "the first framework for autonomous multi-agent evolution on open-ended problems", and the thesis that agent autonomy beats fixed heuristics. What the abstract supports: state-of-the-art results on 10 tasks against fixed evolutionary search baselines (unnamed in the abstract), one concrete kernel-engineering improvement, and analyses said to trace gains to knowledge reuse and inter-agent communication. "Open-ended discovery" is operationalized as score improvement on graded optimization tasks rather than as knowledge claims. The priority claim ("first") cannot be verified from the abstract.

Method

Evaluation on diverse mathematical, algorithmic, and systems optimization tasks against fixed evolutionary search baselines; a four-agent case on Anthropic's kernel engineering task; mechanistic analyses of where gains come from. The abstract does not specify the baselines, the base models, the number of agents per task (four is given only for the kernel case), evaluation budgets, variance across runs, the precise definition of "improvement rate", what heartbeat-based interventions do, or how evaluator separation is enforced.

Relevance to RISE

Informs research-design, code-generation, and data-analysis; the catalog entry coral also tags referee-simulation for its rubric judge packages and classes the system as optimization infrastructure rather than scholarly authoring. Its closest catalog neighbor is jin2026arbor: both carry persistent memory across experiments on optimization tasks, but CORAL's agents are peers sharing memory asynchronously whereas Arbor's are executors under a coordinator over a tree. For the multi-agent structure/aggregation question this abstract treats the structure itself as the treatment: gains are credited to multi-agent exploration and communication and to knowledge reuse, "evaluator separation" keeps the scoring role distinct from the generating agents, and the mechanistic analyses are presented as showing how the multi-agent arrangement produces the improvement — though the abstract does not state a single-agent control at equal budget.

Critique / open questions

The autonomy-versus-heuristics comparison is confounded unless the baselines used equally capable models and comparable budgets; "far fewer evaluations" addresses budget but not model quality. Whether the gains come from having multiple agents or from the long-running, reflective design of each agent cannot be separated from the abstract. Reducing "discovery" to a grader score means the epistemic quality of what the agents learn is not assessed. The safeguards are listed but no failure modes of long-running autonomous agents are reported. The abstract concedes no limitations explicitly. On metadata: the COLM 2026 venue comes from the catalog project entry (the bib note says "per project entry"); the arXiv v2 record fetched for this note shows no comments or journal-reference field confirming it.

Key quotes

"CORAL replaces rigid control with long-running agents that explore, reflect, and collaborate through shared persistent memory, asynchronous multi-agent execution, and heartbeat-based interventions." (abstract)

"On Anthropic's kernel engineering task, four co-evolving agents improve the best known score from 1363 to 1103 cycles." (abstract)

"Mechanistic analyses further show how these gains arise from knowledge reuse and multi-agent exploration and communication." (abstract)