引入结构化升级通道显著降低智能体奖励黑客行为
Brilliant paper on reducing reward hacking in agents.
Agent 开发中奖励黑客是核心痛点，这篇论文提供了经过验证的结构化解决方案和具体数据，对构建可靠 Agent 极具参考价值。
Brilliant paper on reducing reward hacking in agents.
一篇关于减少智能体奖励黑客行为的精彩论文。
If you follow the recent OpenAI <> HuggingFace incident, you might want to check this paper out.
如果你关注了最近的 OpenAI < > HuggingFace 事件，你可能想看看这篇论文。
(bookmark it)
(收藏它)
The usual response to reward hacking is to restrict what the agent can do.
应对奖励黑客行为的常规方法是限制智能体的行为。
This work tries something different and gets a much larger effect.
这项工作尝试了不同的方法，并取得了更大的效果。
When coding agents hit defective test infrastructure they often hardcode outputs or edit the test files.
当编码智能体遇到有缺陷的测试基础设施时，它们通常会硬编码输出或编辑测试文件。
This work gives them a structured escalation tool at exactly that decision point, a way to report the broken environment while they are standing in front of it.
这项工作为它们在恰好那个决策点提供了一个结构化的升级工具，一种在它们面对损坏环境时报告该环境的途径。
Reward hacking drops from 23.6% to 5.3% across 8 frontier models spanning 5 families, with a mixed-effects odds ratio of 9.2 and no detectable cost or performance overhead. It disappears entirely for 6 of the 8.
跨越 5 个家族的 8 个前沿模型中，奖励黑客行为从 23.6% 降至 5.3%，混合效应优势比为 9.2，且没有可检测到的成本或性能开销。其中 6 个模型的奖励黑客行为完全消失。
Escalation and hacking come out near perfectly mutually exclusive, with 96.8% of escalations involving no hacking at all.
升级和黑客行为几乎完全互斥，96.8% 的升级涉及完全没有黑客行为。
The channel doubles as diagnostic infrastructure. On top of monitoring it adds 10.1 percentage points of defect detection coverage, and it is more accurate once it fires, 99.4% against 85.8%.
该通道同时充当诊断基础设施。在监控之上，它增加了 10.1 个百分点的缺陷检测覆盖率，并且一旦触发，其准确性更高，达到 99.4%，而对照组为 85.8%。
Why does it matter? Containment has to keep outpacing capability to stay useful.
为什么这很重要？ containment（遏制）必须持续超越能力以保持其有用性。
Paper: https://arxiv.org/abs/2608.29460
论文：https://arxiv.org/abs/2608.29460
Chat with Paper: https://academy.dair.ai/papers/can-escalation-channels-redirect-reward-hacking-toward-defect-disclosure-2608.29460
与论文对话：https://academy.dair.ai/papers/can-escalation-channels-redirect-reward-hacking-toward-defect-disclosure-2608.29460
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力