Newest papers

VeriEvolve: Evaluator-First Evolution for Tool-Using Agents

Yueshen Li, Kamer Ali Yüksel, Gökhan Tür, Dilek Z. Hakkani-Tur, João Pedro Maia · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.29005 · cited by 0 works (OpenAlex) · open-access PDF

Portuguese

Um obstáculo central para os agentes de auto-melhoria é a ausência de um sinal de aptidão confiável para tarefas de estado realistas. Introduzimos \textsc{VeriEvolve}, um framework de evolução do avaliador-primeiro que constrói esta camada de avaliação perdida antes de otimizar o agente. A partir da instrução da tarefa, o VeriEvolve sintetiza estados de desafio baseados em instruções e um avaliador executável que os verifica. Durante cada rodada de evolução do fluxo de trabalho, o avaliador é congelado e serve como o sinal de aptidão. Entre as rodadas, o VeriEvolve revisa o avaliador convertendo os estados finais de execuções fortes e fracas em um par de casos de teste contra-fatal.

Machine translation (Meta M2M100); the English original follows.

Abstract

A central obstacle to self-improving agents is the absence of a reliable fitness signal for realistic stateful tasks. We introduce \textsc{VeriEvolve}, an evaluator-first evolution framework that constructs this missing evaluation layer before optimizing the agent. From task instruction, VeriEvolve synthesizes instruction-grounded challenge states and an executable evaluator that checks them. During each round of workflow evolution, the evaluator is frozen and serves as the fitness signal. Between rounds, VeriEvolve revises the evaluator by converting the end states of strong and weak runs into a pair of counterfactual test cases. Across all 30 MCPMark Filesystem tasks, VeriEvolve constructs 30 evaluators with 141 challenge states, achieving 141/141 agreement on the generated suites and 113/122 agreement on held-out states produced by executor agents. 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. These results show that executable evaluators can be synthesized and used to guide search when task success is observable and an external verifier is available for calibration.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Executor, Executable, Computer science, Task (project management), Counterfactual thinking, Stateful firewall

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange