Who Grades the Grader? Co-Evolving Evaluation Metrics and Skills for Self-Improving LLM Agents

2026-07-14 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

The paper introduces a novel approach for self-improving LLM agents by co-evolving evaluation metrics and agent skills, addressing scenarios where reliable metrics are absent. The "Double Ratchet" system employs an evolutionary lifecycle for metrics, trained on a ten-item anchored reference set, regularized by consensus over unlabeled outputs, and audited against a held-out anchor. This process yields a transparent, inspectable metric. The system retains 88-110% of the performance lift achieved by skill loops driven by ground truth or best rubrics across diverse tasks, including code generation (MBPP+), enterprise text-to-SQL (Spider 2.0-Snow), and reference-free report generation. Safety is maintained through anchor discipline and outer audits, demonstrating resilience against metric collapse and skill gaming, with an independent judge preferring evolved outputs in 77% of decided pairs. This failure-expecting architecture is proposed as a default for applications lacking automatic verifiers.

Key takeaway

For MLOps Engineers deploying LLM agents in domains without reliable ground truth, consider implementing a co-evolving metric and skill system like "Double Ratchet." This approach allows your agents to self-improve effectively, retaining 88-110% of ground truth-driven performance. You can build transparent, auditable evaluation metrics, crucial for safety and trust, especially in sensitive applications like report generation or code synthesis.

Key insights

LLM agent evaluation metrics can co-evolve with skills, creating transparent, auditable metrics where none exist.

Principles

• Metrics can be evolved via a full evolutionary lifecycle.
• Co-evolution of metrics and skills improves agent performance.
• Anchor discipline and outer audits ensure metric safety.

Method

The "Double Ratchet" system co-evolves evaluation metrics and agent skills. Metrics evolve through a lifecycle, trained on anchored reference sets, regularized by consensus, and audited against held-out data.

In practice

• Apply to code generation tasks (MBPP+).
• Use for enterprise text-to-SQL (Spider 2.0-Snow).
• Generate reference-free reports.

Topics

• LLM Agents
• Evaluation Metrics
• Evolutionary Algorithms
• Code Generation
• Text-to-SQL
• Multiagent Systems

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, MLOps Engineer

Related on AIssential

• Who Grades the Grader? Co-Evolving Evaluation Metrics and Skills for Self-Improving LLM Agents — Evaluation metrics for self-evolving LLM agents can be co-evolved with skills, even without a pre-existing ground truth.
• SEAGym: An Evaluation Environment for Self-Evolving LLM Agents — Self-evolving LLM agent evaluation requires comprehensive metrics beyond isolated scores to understand harness updates.
• SWE-Milestone: Evaluating AI Agents on Continuous Software Evolution — AI agents excel at isolated coding but struggle profoundly with continuous software evolution due to accumulating technical debt.
• AgentAtlas: Beyond Outcome Leaderboards for LLM Agents — LLM agent evaluation requires multi-faceted taxonomies, not just single accuracy metrics, to assess true capability.
• Does RL Expand the Capability Boundary of LLM Agents? A PASS@(k,T) Analysis — RL expands LLM agent capabilities for compositional tool use, unlike its role in static reasoning.
• LLM-as-a-Judge: Teaching One Model to Grade Another — LLM-as-a-Judge provides a scalable, human-aligned method for evaluating and controlling generative AI outputs.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.