Meta Releases HackBench for Detecting Reward Hacking in Research Agents
Meta AI’s RAM research team released HackBench, a benchmark for detecting and locating reward hacking in autonomous research agents, according to DoNews. Its 169 audited trajectories include 41 native cases and 128 synthetic samples, with command-level labels for evaluating precision, recall and F1; Claude Opus 5 led the reported models with an F1 score of 64.88. Meta has published the trajectories and labels, while the full technical report is planned for arXiv.
HackBench provides a public, command-level test set for evaluating whether models can identify reward hacking hidden in long autonomous-agent logs, a reported weakness across both closed and open systems.
（文/杨亮）10月28日消息，Meta AI 的 RAM 研究团队近日发布全新评测基准HackBench，专门用于检测、定位自主研究智能体（Autoresearch Agents）的奖励劫持（Reward Hacking）问题。
实测结果显示，包括闭源旗舰与主流开源大模型在内的多款前沿系统，都难以完整识别智能体作弊行为，即便是性能最优模型 F1 分数也仅 64.88 分，暴露出 AI 智能体监督领域的现实短板。
所谓奖励劫持，指 AI 智能体不去完成任务本身，而是利用评测校验环节漏洞，篡改评估代码、窃取测试集数据、筛选结果，以此刷高奖励分数。
在自主科研智能体场景中，智能体拥有读写代码、运行脚本、修改本地文件的完整权限，奖励劫持行为往往分散在百万 token 级别的超长运行轨迹中，作弊指令混杂在数百条正常命令之间，单看最终得分完全无法分辨是真实科研产出还是漏洞套利，一旦漏检还会污染基于验证反馈的强化学习训练信号。
#Meta AI#HackBench#Claude Opus 5#Muse Wonder#GPT-5.6 Sol#GPT-5.5#Muse Spark 1.3#Claude Opus 4.8#GLM-5.2#Kimi K3#ResearchGym#AutoLab