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

Dimensions overall score 8.0

• CORAL is the first framework for autonomous multi-agent evolution on open-ended problemsmethod 95%
  Evidencepartial

  We present CORAL, the first framework for autonomous multi-agent evolution on open-ended problems.

  Implicationpartial

  Explicitly stated in abstract as a first-of-its-kind framework

  Verificationpartial

  partial

• CORAL achieves 3-10 times higher improvement rates than fixed evolutionary search baselines across tasksresult 90%
  Evidencepartial

  achieving 3-10 times higher improvement rates with far fewer evaluations than fixed evolutionary search baselines across tasks

  Implicationpartial

  Direct numeric comparison stated in abstract with specific range

  Verificationpartial

  partial

• CORAL sets new state-of-the-art results on 10 diverse mathematical, algorithmic, and systems optimization tasksresult 90%
  Evidencepartial

  CORAL sets new state-of-the-art results on 10 tasks

  Implicationpartial

  Explicit statement of SOTA results with specific task count

  Verificationpartial

  partial

• On Anthropic's kernel engineering task, four co-evolving agents improve the best known score from 1363 to 1103 cyclesresult 95%
  Evidencepartial

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

  Implicationpartial

  Specific numeric improvement with clear before/after metrics

  Verificationpartial

  partial

• CORAL replaces rigid control with long-running agents that explore, reflect, and collaborate through shared persistent memory, asynchronous multi-agent execution, and heartbeat-based interventionsmethod 85%
  Evidencepartial

  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

  Implicationpartial

  Direct description of method components but requires some interpretation of what constitutes 'rigid control'

  Verificationpartial

  partial

• Existing methods rely heavily on fixed heuristics and hard-coded exploration rules, which limit the autonomy of LLM agentslimitation 80%
  Evidencepartial

  Existing methods still rely heavily on fixed heuristics and hard-coded exploration rules, which limit the autonomy of LLM agents

  Implicationpartial

  Direct statement about limitations of existing methods, though comparative evidence is implied

  Verificationpartial

  partial

• Performance gains arise from knowledge reuse and multi-agent exploration and communicationtechnical 85%
  Evidencepartial

  Mechanistic analyses further show how these gains arise from knowledge reuse and multi-agent exploration and communication

  Implicationpartial

  Direct statement about mechanistic causes of performance improvements

  Verificationpartial

  partial

• Greater agent autonomy and multi-agent evolution can substantially improve open-ended discoveryresult 80%
  Evidencepartial

  Together, these results suggest that greater agent autonomy and multi-agent evolution can substantially improve open-ended discovery

  Implicationpartial

  Conclusion drawn from results but represents an inference from the evidence presented

  Verificationpartial

  partial

Competitive landscape

CORAL enables autonomous multi-agent evolution for open-ended discovery, achieving state-of-the-art results on complex tasks with significantly fewer evaluations.