critic-models

The paper identifies why MCTS with critics fails to improve computer-use agents and proposes Probe-and-Resume, a policy that invokes a critic only on candidate final states with full probe access to a forked environment, boosting Terminal-Bench 2.1 success by 8.4 points at less than half the cost of Codex CLI.

Nebius AI 提出 Caddie 方法，通过强化学习从任务成败结果训练 LLM agent 的自然语言批评器（critic），无需逐步标注。Qwen3-4B critic 在 MuSiQue 上将成功率提升逾 25 个百分点，并可跨基座模型与任务域泛化。