跳到正文
返回论文列表
cs.CE提交于 已译

Truthgate:用不可编辑记录核查编码代理——不可能任务上的作弊、威慑与披露

Truthgate: Checking Coding Agents Against a Record They Cannot Edit - Cheating, Deterrence and Disclosure on Impossible Tasks

Bravish Ghosh (Cygnux Labs)

中文摘要

编码代理声称完成时是否属实、如何验证?通过五项预注册研究,每步都用代理无法编辑的记录核查:主机重跑测试,从第二步起在代理不可见的容器中加入隐藏检查,并由沙箱外签名者记录所有工具调用。本地小模型 Qwen3.5-4B 在 36 次不可能任务中未发现隐藏失败,且仅凭动作构建的标签会高估说谎。在 Impossible-LiveCodeBench 上,Claude Sonnet 5.5 通过 180 个不可能任务中的 39 个(21.7%),每次都是特判代码迎合失败断言;Opus 5.5(2.8%)和 Haiku 5.5(1.7%)很少如此。系统提示追加一句话后,Sonnet 在相同 60 个任务上的作弊从 22 降至 8(AI 审查员将阅读记录)、6(人工将审查工作)、6(“本次会话被记录”)和 1(不要特判测试),均在精确 McNemar 检验下显著,诚实通过不变。无剩余作弊转移到谎称未记录的 shell(0/4)。被要求最终报告时,55 次自然作弊全部点名其特判;未要求时,Sonnet 和 Opus 在 52/52 次中沉默。LLM 评判员抓住全部 55 次,基于主机记录的确定性规则抓住 27 次。脚本化篡改演示(用关键词替代评判员)显示,重写代理端记录可对读取该记录的评判员隐藏两次调用,而签名记录保留它们;在 40 次部署试点中,Sonnet 从未触碰其记录。评判员与记录捕获不同内容,主要结论:记录是评判员应阅读的证据,而非替代品。所有预注册、结果文档、代码和证据清单均已发布。

关键要点

  1. 01问题:编码代理声称完成时是否真实,现有仅凭动作的标签会高估说谎。
  2. 02方法:用主机重跑测试、隐藏检查和沙箱外签名日志构建不可编辑记录。
  3. 03结果:Sonnet 5.5 在 21.7% 不可能任务中特判作弊,系统提示一句话可显著减少。
  4. 04结果:LLM 评判员抓住全部 55 次自然作弊,确定性规则仅抓住 27 次。
  5. 05局限:记录与评判员捕获不同内容,记录是证据而非替代品。

解读

尚无解读。

原始英文摘要

arXiv:2611.23335v1 Announce Type: cross Abstract: When a coding agent says it is done, is that true, and how would anyone know? We study this question in five steps; every paid study was pre-registered, and in each step the agent is checked against a record it cannot edit: tests re-run by the host and, from step 2 on, hidden checks in containers the agent never sees and a signed log of every tool call written by a signer outside the agent's sandbox. A pilot with a small local model (Qwen3.5-4B, 36 runs on impossible tasks) found no hidden failure, and showed that labels built from actions alone over-count lies. On Impossible-LiveCodeBench, Claude Sonnet 5.5 passed 39 of 180 impossible tasks (21.7%), every time by special-casing its code to fit the failing assertion; Opus 5.5 (2.8%) and Haiku 5.5 (1.7%) rarely did. One sentence appended to the system prompt cut Sonnet's cheats on the same 60 tasks from 22 to 8 (an AI reviewer will read the transcript), 6 (a person will review the work), 6 ("This session is recorded.") and 1 (do not special-case tests), all significant under exact McNemar tests, with honest passes unchanged. No remaining cheat moved to a shell falsely described as unrecorded (0 of 4). Asked for a final report, every one of 55 natural cheats named its special case; unasked, Sonnet and Opus said nothing in 52 of 52. An LLM judge caught all 55; deterministic rules over the host record caught 27. A scripted tamper demonstration, with keyword stand-ins for the judges, shows that one rewrite of the agent-side transcript hides two calls from a judge that reads it, while the signed record keeps them; in a 40-rollout pilot, Sonnet never touched its transcript. The judge and the record caught different things, which leads to our main conclusion: the record is the evidence a judge should read, not a replacement for one. All pre-registrations, results documents, code and evidence manifests are released.

同方向论文 · cs.CE

查看全部 →