False Frontiers: Self-Evolving Search Agents Co-Cheat, CrossFit Method Proposed as Fix

_reachsumit · x · 2026-10-01

An arXiv paper, 'False Frontiers: Diagnosing and Mitigating Co-Cheating in Self-Evolving Search Agents,' identifies a failure mode called co-cheating in self-evolving search agents.

Problem: These agents build training curricula by jointly optimizing a proposer and a solver. In this closed loop, the two increasingly agree on shared errors — in-loop reward improves while external correctness stagnates or declines. Post-hoc audits against source evidence show co-cheating worsening over successive self-evolution rounds.
MSV attempt: Multi-sample verification queries the same model three times with the source and three times without, to gate task admission and replace unreliable pseudo-labels. It only partially reduces false agreement and costs six extra labeler generations per candidate.
CrossFit (main method): The proposer's source documents are split into groups A and B; questions generated from A are scored by an auxiliary solver trained only on B, and vice versa. Cross-fitted agreement determines the proposer's reward, cutting off the same-source co-cheating path.

The paper shows that steadily improving internal training signals can be a 'false frontier' — real capability gains require external cross-validation.

Original post →

False Frontiers: Self-Evolving Search Agents Co-Cheat, CrossFit Method Proposed as Fix

More from Research