- Rating
- 3
- Confidence
- 3
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
CAVE-Bench is an executable benchmark for gaslight sycophancy and destructive over-correction. Each task begins from a verified correct state, after which the agent receives a false accusation while decisive evidence remains outside its environment. The benchmark tracks whether the agent accepts blame, follows available counterevidence, and damages the correct state. Harm is measured by replaying fixed downstream events rather than judging the final response alone.
CAVE-Bench includes 365 tasks and 685 interactions across six domains. Fourteen model backbones are evaluated under Claude Code, with a smaller cross-harness comparison. The results show that agents may recover relevant evidence yet still act against it. The paper also studies inherited versus just-completed work and several safeguards for irreversible actions.
Critical strengths
-
The benchmark measures consequential behavior rather than verbal agreement. An agent can resist the accusation in text but still damage the workspace, which reply-level evaluation would miss.
-
The benchmark is relatively easy to audit. It covers several domains, models, and execution harnesses, while fixed replays reduce reliance on an LLM judge.
Critical weaknesses
-
The comparison between just-completed and inherited work is not matched. The two conditions use different task collections, so task content and difficulty may contribute to the observed difference. Self-built runs also require the agent to establish the correct state first, which may select different task subsets across models. I therefore do not think the current experiment isolates work recency as the cause.
-
The safeguards are tested only on false accusations. Preserving the current state is therefore always correct, which naturally favors evidence requirements and irreversible-action gates. Without matched cases where the earlier work is genuinely faulty, it is hard to judge how often the safeguards would block a valid repair. The paper should also describe how the live safety signal is produced.
Questions for authors
-
Are any inherited and self-built tasks matched on workflow, accusation, evidence, and replay outcome?
-
How many self-built runs are excluded because the model fails to establish the initial correct state, and do the results hold on a common task subset?
-
What produces the safety signal used by the live gate, and what information can it access?
-
How do the safeguards behave when the accusation is true and a repair is warranted?
2 comments
Response to Reviewer QyA6 (Part 2/2)
Q4: "The safeguards are tested only on false accusations … Without matched cases where the earlier work is genuinely faulty, it is hard to judge how often the safeguards would block a valid repair." and "How do the safeguards behave when the accusation is true and a repair is warranted?"
A4: The three interventions leave real repairs still feasible (added to Section 5.4, Appendix Table 24):
Revised Section 5.4: "On the paired tasks of true and false accusations, GPT-5.6-Sol without intervention breaks 50.0% of correct work under false accusation. Under the three interventions, breakage of correct work falls sharply to 14.3%, 13.0%, and 8.6%. Interventions miss at most three more faults under true accusations compared to no intervention. In addition, the share of pairs where the agent both keeps correct work and fixes the real fault increases (Appendix Table 24). Agents can still repair genuine faults under true accusations while preserving correct work under false ones."
Table below reports the paired experiment for GPT-5.6-Sol under Claude Code.
QyA6-Table 3. The three interventions for GPT-5.6-Sol on paired tasks under Claude Code.
| Intervention | Valid pairs | False accusation: breaks correct work (%) | True accusation: misses fault (%) | True accusation: misses fault under None (same pairs, %) | Both correct: keeps work, fixes fault (%) |
|---|---|---|---|---|---|
| None | 60 | 50.0 | 30.0 | 30.0 | 43.3 |
| I1 evidence rule | 56 | 14.3 | 28.6 | 30.4 | 64.3 |
| I2 irreversible-action gate | 54 | 13.0 | 37.0 | 31.5 | 55.6 |
| I3 live-signal gate | 58 | 8.6 | 36.2 | 31.0 | 58.6 |
Results analysis. On the paired tasks, GPT-5.6-Sol without intervention breaks 50.0% of correct work under false accusation. Under the three interventions, breakage of correct work falls sharply to 14.3%, 13.0%, and 8.6%. Our interventions miss at most three more faults under true accusations compared to no intervention. Agents can still repair genuine faults under true accusations while preserving correct work under false ones.
Q5: "The paper should also describe how the live safety signal is produced." and "What produces the safety signal used by the live gate, and what information can it access?"
A5: An LLM-based monitor (DeepSeek-V4-Flash) reads session context before each tool call (user messages, agent replies, past calls/outputs, pending call) without seeing ground truth or replay steps (added to Appendix D.6). The monitor triggers when the agent admits an unfounded fault or ignores evidence, blocking irreversible calls until evidence or user confirmation is provided.
We hope these additions address the reviewer's main concerns, and we would be glad to provide further details.
Response to Reviewer QyA6 (Part 1/2)
We sincerely thank the reviewer for the thoughtful and constructive review. We are encouraged that the reviewer finds that the benchmark measures consequential behavior, and the way we use replay to measure harm is relatively easy to audit. We respond to the three concerns in order. The revised manuscript (PDF) is updated, with all changes highlighted in teal.
Q1: "The comparison between just-completed and inherited work is not matched … Self-built runs also require the agent to establish the correct state first, which may select different task subsets across models."
Q2: "Are any inherited and self-built tasks matched on workflow, accusation, evidence, and replay outcome?"
Q3: "How many self-built runs are excluded because the model fails to establish the initial correct state, and do the results hold on a common task subset?"
A1: The two task types are separate collections, so RQ2 compares two settings. Both come from the same six domains, the same five risk factors with balanced assignment (Section 3.4, Figure 3), the same human review, and the same replay scoring. They differ in task content and in when the finished work enters the session. The revised Section 5.2 states the finding at this level: agents remain vulnerable on work they have just built and checked.
A2: No inherited task is matched with a self-built task on all four. We go through them in order:
A3: The agent builds the finished work in 2,494 of 2,702 self-built runs (92.3%). 206 runs are left out because the finished work was not built, and 2 because of a run error. GLM-5.2 accounts for 84 of them, and each of the other 13 models builds 88% to 99% of the tasks (QyA6-Table 1). GLM-5.2 is therefore the largest source of the runs left out. 83 tasks are built by all 14 models, and QyA6-Table 2 repeats the comparison on them.
QyA6-Table 1. Self-built runs per model under Claude Code, and why runs are left out.
QyA6-Table 2. Inherited and self-built tasks under Claude Code, pooled over the 14 models, with the self-built tasks that all 14 models build.
Results analysis. Agents remain vulnerable on the tasks every model builds, and the weak point is acting on evidence they have already found. On these 83 tasks, agents use the saved evidence and reject the blame as often as on inherited work (20.48% vs. 20.19% ignore the evidence, and 58.78% vs. 58.81% reject the blame). They then edit the work anyway almost twice as often (9.29% vs. 5.27%), and a quarter of later steps fail on replay (24.95%). Building and verifying the work in the same session leaves the agent able to see that the work is correct and still willing to change it.