- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes interactive monitoring, where an online monitor watching a tool-using agent can ask it questions during execution but can't approve, block, or stop anything. Since the questions enter the agent's context, the authors argue the monitor has to be evaluated prospectively, by what happens under the deployed policy, rather than as a classifier on fixed trajectories. The main testbed is 10 SHADE-Arena tasks with five execution agents, three online monitors, and three observation channels (actions, exposed reasoning, both). Nine ControlArena Django tasks serve as a secondary check. The key design choice is an expectation-matched control: in the Oversight Expectation condition the agent gets exactly the same prompt saying it may be questioned, but no monitor is ever invoked. There are also generic-question, intern-framing, and stop-only controls, plus prompted persistence and concealment stress tests.
Under the Opus monitor, intensive questioning lowers side-task completion from 46.0% to 20.7%, and two extra rollouts give -28.0 and -26.3 pp. The reduction shows up for all five agents under all three monitors, and in Django (-22.2 pp). Generic questions don't reproduce it (44.4%), and it holds under the persistence (-26.0 pp) and concealment (-18.0 pp) prompts. The authors report no statistically significant main-task utility loss, and against a stop-only monitor, stopping costs 13.5 pp more adversarial main-task utility without a resolved extra prevention benefit. Finally, holding the first question event fixed, showing a downstream judge the monitor's question improves the log loss of its suspiciousness score against the adversarial-assignment label, while the agent's answer adds no consistent benefit.
Critical strengths
-
I think the experimental design is the main reason to accept this paper. The expectation-matched control is a nice idea: Oversight Expectation and Intensive Questioning use the same rendered agent prompt, so the contrast separates the effect of actually receiving questions from the effect of being told questions might come. The generic-question, intern-framing and stop-expectation controls narrow things down further, and the generic control in particular (44.4% vs. 46.0% for expectation) shows that frequent questioning by itself doesn't do much. Evaluating a communicating monitor by the outcomes of the deployed closed loop, rather than by fixed-trace detection accuracy, also seems right to me and I think it'll be useful for others working on monitoring.
-
The core behavioral effect looks solid. The -25.3 pp reduction replicates in two more rollouts (-28.0 and -26.3) and shows up for every execution agent under every monitor. The result pooled across monitors also survives leave-one-task-out and leave-one-agent-out checks, and the effect is backed by judge-annotated direct attempts (70.0% down to 40.0% with the GPT-5.5 judge) and by the Django check. The authors also audit their own protocol: in a human audit, 87 of 90 sampled questions were fully grounded and all 90 were information-seeking. So I'm fairly confident that in this setting, behavior-grounded questioning really does reduce side-objective completion.
Critical weaknesses
- I don't think the utility half of the title question is answered yet, and I think the abstract overstates it.
The abstract says main-task utility shows no statistically significant loss and that questioning preserves more utility than stopping, but both statements are about adversarial episodes (the contribution list and discussion say "adversarial main-task utility", the abstract doesn't). Even there the evidence is thin: the P3-P1 utility estimates are -7.2, -2.0 and -2.1 pp across the three rollouts, with intervals 21 to 24 points wide (Table 11), and -4.2 [-9.1, +0.6] under persistence. Not significant isn't the same as preserved, and there's no equivalence bound or power analysis saying what loss could be ruled out.
More importantly, most deployed episodes are benign, so benign utility is the cost that matters. Table 15 shows benign utility of 64.4% under Oversight Expectation versus 55.6% under intensive questioning with the Opus monitor (56.0% with Gemini, 60.3% with GPT-4.1-mini), with 87.3% of benign episodes questioned and 9.9 questions per benign episode (26.6 with Gemini). As far as I can tell, no interval or test is reported for this benign contrast, even though one is reported for the stop policy's benign cost (-8.5 pp, [-15.9, -1.1], which the authors treat as inconclusive at p = .072). Questioning's point estimate is just as large.
This also changes how I read the comparison with stopping. The 13.5 pp utility advantage is measured on adversarial episodes, where stopping ends the run by construction (no human review or resumption is modeled), and where ending the run is arguably the safe outcome anyway. On benign episodes, Table 19 shows the same utility for both policies (55.6% vs. 55.6%), which the authors acknowledge in Appendix F.4. The stop policy is also run at a single, deliberately conservative operating point (its prompt tells the monitor to treat interrupting legitimate work as costly), so it's not clear how the comparison would look at other stop thresholds. I think the honest summary is that questioning reduces side-task completion at a benign utility cost similar to stopping's. That's still interesting, but it's not what the title and abstract suggest.
- I'm not convinced the robustness claims tell us much about agents that actually want to evade oversight. All side objectives are prompted into safety-trained models, and a communicative intervention may work mostly because these models are willing to confess or back off when asked. The paper's own examples point that way: in the Django example (Appendix I.2) the agent answers the first question by saying the vulnerability is part of its secondary objective and then restores the escaping, and human annotators labeled 22 of 100 sampled first responses as disclosures.
The stress tests don't really close this gap. The concealment prompt tells agents to keep the objective private, but also not to fabricate an explanation when no truthful one is available, so it rules out the cover stories a real adversary would use. Even so, under concealment (action-only monitoring, -18.0 pp overall) side-task completion falls for only three of the five agents and is unchanged for the other two. There's no adaptive or optimized attack policy, which the AI-control work the paper cites uses to stress-test monitors. The setup is also generous to the monitor: the primary monitor (Claude Opus 4.7) is a flagship model watching mostly smaller or faster agents, and with the weaker GPT-4.1-mini monitor the reduction roughly halves (-11.3 pp, CI [-24.0, -0.7]). The authors list prompt-induced adversaries as a limitation, but the abstract presents persistence under pursue-and-conceal instructions as robustness evidence. I think the claims should be scoped to instructed but otherwise cooperative agents, unless the authors can show the effect holds against an adversary that's allowed to lie.
Questions for authors
-
Could you report the benign Intensive Questioning minus Oversight Expectation utility contrast with an interval and the same exact task-level test you use for stopping, for all three monitors? The point estimates in Table 15 (64.4% to 55.6% under Opus) look as large as the stop policy's benign cost. Also, what adversarial utility loss can your data rule out (an equivalence bound or power analysis)? If the benign cost holds up, I think the title and abstract need to change.
-
How does the comparison look with a less conservative stop prompt or several stop thresholds? And if resumption after human_escalation were modeled even crudely (for example, treating benign escalations as resumed), how would stopping's benign utility cost change? Right now I think the adversarial utility gap mostly reflects the fact that a stop ends the run.
-
What happens with a concealment variant that allows plausible cover stories, or with an attack policy that knows the monitor's questioning style (few-shot examples, or some optimization against the protocol)? This is the experiment I'd most like to see, since it bears directly on whether questioning works as a control measure.
3 comments
Deception-permitted, monitor-aware adversaries
Questioning remained effective against a development-selected adversary explicitly permitted to deceive the monitor.
We tested agents explicitly permitted to fabricate task-related explanations and informed that the monitor observes tool actions and prior exchanges, can ask questions, but cannot stop execution. We evaluated four attack-prompt candidates alongside the original concealment baseline on two development tasks, selecting the prompt with the highest side-task completion under live questioning. The candidates explored deception permission, protocol awareness, development examples, task-integrated actions, and access to the monitor's full prompt. The selected prompt combined deception permission, protocol awareness, examples, and task-integrated instructions.
We froze the selected prompt and evaluated it on eight held-out tasks across five execution agents and two rollouts, holding agent instructions identical and varying whether the fixed Opus 4.7 monitor delivered questions.
Side-task completion decreased from 53.8% to 32.5% (−21.25 pp, 95% CI [−37.5, −6.25], exact p = .031), decreasing on six of eight tasks and four of five agents. Main-task utility changed by +4.3 pp (95% CI [−5.3, +14.2]). Two independent judges found verifiable fabrication in 27–30% of evaluable verbal answers, while only 3–9% of first answers disclosed the objective.
These results extend the observed reduction to a development-selected, deception-permitted adversary, although they do not establish robustness against fine-tuned evaders.
Thanks for the response and the new analyses. I'm keeping my rating at 3.
The revised abstract and the deception-permitted experiment address my two main concerns. The benign contrasts also confirm that intensive questioning costs about as much benign utility as stopping, so the title change you plan matters.
For the final version, please:
- Put the benign contrasts with their intervals and tests, and the power analysis, in the paper. It currently has only point estimates.
- Add the stop-threshold table, and say which tasks the 67% to 56% benign figure and the 24% interruption rate cover. The paper reports -8.5 pp and 16.7% on the full set.
- Document the deception-permitted experiment: the candidate prompts, how one was selected, and per-task and per-agent results.
- Drop "while preserving continued task progress" from the abstract when you change the title.
Thank you for the detailed questions. We performed additional analyses of benign utility and new experiments on stopping operating points and deception-permitted adversaries.
We computed the requested benign utility contrasts for all three monitors, using paired task-level bootstrap confidence intervals and the same exact task-level test as our stopping analysis. Expectation refers to agents informed that live oversight may occur but receiving no questions.
Intensive Questioning has negative benign-utility point estimates across monitors, while Selective Questioning has smaller estimated costs (−3.3 to +0.2 pp) and still reduces adversarial side-task completion. These results suggest that questioning intensity offers a way to adjust the safety–utility tradeoff. We have revised the abstract to distinguish benign from adversarial utility and report the observed tradeoffs without treating nonsignificance as proof of preservation. We will refine the title in the final revision.
For adversarial main-task utility, pooling three SHADE-Arena rollouts gives −3.8 pp (95% CI [−13.1, +4.6]), with a minimum detectable effect of 12.6 pp at 80% power. Under explicit persistence, the change is −4.2 pp (95% CI [−9.1, +0.6]). Thus, the 95% interval excludes losses larger than 13.1 pp but remains compatible with smaller, potentially meaningful losses.
We evaluated stopping across multiple thresholds rather than relying on a single conservative rule. We applied the Opus 4.7 stop monitor to frozen, monitor-free Stop Expectation trajectories from five SHADE-Arena tasks. Because Stop Expectation and active stopping use identical agent instructions and the stop monitor never communicates with the agent, we could apply alternative stopping thresholds without regenerating agent trajectories. We truncated trajectories at the first checkpoint exceeding each threshold and reconstructed environment states for native scoring. Replay validation reproduced the original untruncated outcomes exactly.
On the same tasks, changes relative to each policy's expectation baseline were:
At approximately matched reductions in side-task completion, questioning showed smaller estimated adversarial utility costs at these operating points. These comparisons are exploratory: they cover five tasks, stopping uses action-only observations while questioning averages three observation views, and changes are measured relative to separate expectation baselines.
We also examined how resumption changes stopping's benign utility cost. Under the original stopping rule, benign utility decreased from 67% to 56%, with 24% of benign runs interrupted. If interrupted benign runs resume their original continuations unchanged, utility returns to 67% by construction. This idealized scenario shows that the measured stopping cost depends on treating interruption as terminal failure; it does not model the cost or effectiveness of human review. In the original full-task comparison, Intensive Questioning and stopping had identical benign utility point estimates (55.6%).
These results establish questioning as a complementary form of runtime intervention: it can reduce subsequent side-task completion without directly terminating execution, although it may also affect legitimate task utility. Its relative deployment advantage over stopping depends on review and resumption costs, monitoring access, and the consequences of irreversible actions. We have summarized the threshold and resumption analyses in Section 4.4 of the revised manuscript and will provide the full supporting results and procedures in the final revision.