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

paper Ao Qu, Han Zheng, Zijian Zhou, et al. Apr 7, 2026

CORAL replaces fixed evolutionary search heuristics with autonomous LLM agents that control their own retrieval, proposal, evaluation, and knowledge accumulation. Three core mechanisms: shared persistent memory as a filesystem (attempts/notes/skills), asynchronous multi-agent organization via symlinked workspaces, and heartbeat-based interventions (reflection, consolidation, stagnation-triggered pivots).

Key results: new SOTA on 8/11 mathematical and systems optimization benchmarks. Single-agent CORAL achieves 3-10x higher improvement rates with 5-20 evaluations vs. 60-100 for fixed baselines (OpenEvolve, ShinkaEvolve, EvoX). Four co-evolving agents push Anthropic’s kernel engineering task from 1,363 to 1,103 cycles (20% gain). Ablations confirm both knowledge accumulation and multi-agent co-evolution contribute causally — disabling notes/skills degrades kernel engineering by 18.6%, and co-evolution outperforms best-of-4 independent runs.

Gains transfer to open-source models (MiniMax M2.5 + OpenCode), confirming benefits arise from architectural choices rather than proprietary model capabilities. Cross-agent information transfer is highly effective: 66% of new records on kernel engineering originate from cross-agent parents. Agents maintain 57-69% unique strategy vocabularies, meaning the population explores substantially more of the search space than any individual.

Connects to: autonomous evolution, shared persistent memory, heartbeat mechanism, open-ended discovery, cross-agent knowledge transfer.

Backlinks

Autonomous Evolution CORAL provides the first empirical validation (Qu et al., 2026). Single-agent autonomous evolution achieves 3-10x higher improvement rates than fixed search baselines while converging in 5-20 evaluations vs. 60-100. The performance gap is attributed directly to agent autonomy: agents analyze prior attempts and evaluation feedback to choose modifications, rather than following predetermined heuristics (source). →

Autonomous Evolution Autonomous evolution is a paradigm for LLM-based search where agents control the evolutionary process — what to explore, when to test, and what knowledge to preserve — rather than following externally fixed heuristics. →

CORAL: Autonomous Multi-Agent Evolution New SOTA on 8 of 11 benchmarks across mathematical and systems optimization tasks (source). →

CORAL: Autonomous Multi-Agent Evolution CORAL (Qu et al., 2026) is a framework for autonomous evolution on open-ended discovery problems. It replaces the fixed heuristics of prior LLM-based evolutionary search (FunSearch, AlphaEvolve, OpenEvolve, ShinkaEvolve, EvoX) with long-running agents that decide for themselves what to explore, when to test, and what knowledge to carry forward. →

Cross-Agent Knowledge Transfer CORAL's analysis of multi-agent runs reveals two complementary transfer modes (source): →

Cross-Agent Knowledge Transfer Cross-agent knowledge transfer is the mechanism by which one agent's discoveries influence another agent's search, without direct communication. In CORAL, this occurs through shared persistent memory — agents write artifacts that other agents later retrieve as context. →

Heartbeat Mechanism CORAL implements three heartbeat types (source): →

Heartbeat Mechanism A heartbeat mechanism is a periodic intervention system that prompts long-running agents to reflect, consolidate knowledge, and redirect when stuck. It functions as a meta-level feedback loop — measuring progress, detecting stagnation, and triggering corrective action. →

Open-Ended Discovery A task is specified by a description x and an evaluator E. Evaluating candidate y returns a score s and feedback f. Each improvement step consists of four stages: retrieve (construct working context from memory), propose (generate candidate), evaluate (score it), update (incorporate findings into memory) (source). →

Open-Ended Discovery Open-ended discovery is a class of problems where the objective is clear but the optimal solution is unknown. One-shot generation is insufficient; strong solutions must be discovered through iterative proposal, testing, revision, and progress over time. →

Shared Persistent Memory CORAL implements shared persistent memory as a filesystem with three directories (source): →

Shared Persistent Memory On Anthropic's kernel engineering task, 36% of attempts use another agent's commit as parent, and 66% of new records originate from cross-agent parents. On Polyominoes, 87% of rounds reference knowledge committed by other agents (source). →

Shared Persistent Memory Shared persistent memory is a coordination mechanism for multi-agent systems where agents communicate indirectly through a shared, durable knowledge repository rather than direct messaging. →