ResearchResearch paperTraining & Scaling · Reinforcement Learning · Reasoning & Planning1 source · Oct 21, 2026

VeriEvolve: Evaluator-First Evolution for Tool-Using Agents

We introduce VeriEvolve, an evaluator-first evolution framework that constructs this missing evaluation layer before optimizing the agent.

A central obstacle to self-improving agents is the absence of a reliable fitness signal for realistic stateful tasks.
From task instruction, VeriEvolve synthesizes instruction-grounded challenge states and an executable evaluator that checks them.
Evaluator-guided evolution increases externally verified success from 6/30 to 18/30 for GLM-5.2 and from 27/30 to 30/30 for Claude; refinement removes all three false positives that the weak executor exposes.
Cross-environment studies on PostgreSQL, Playwright, Notion, and GitHub further expose complementary limits from observability, state representation, and search/execution capacity.

[1]VeriEvolve: Evaluator-First Evolution for Tool-Using Agents
arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 21, 12:13 AM
We introduce VeriEvolve, an evaluator-first evolution framework that constructs this missing evaluation layer before optimizing the agent.
A central obstacle to self-improving agents is the absence of a reliable fitness signal for realistic stateful tasks.

Extractive summary: sentences quoted from the sources.

Before this