acceptodds
Under review as a conference paper at ICLR 2027

Capability Detection: Monitoring Hidden and Emergent Capabilities in Tool-Using Agents

Abstract

Runtime monitors for tool-using agents evaluate each action against behavioral rules, yet this per-action assessment can miss how tools are being coordinated: a chain of individually benign calls may produce capabilities that no single call would expose. This gap motivates a complementary question: can a weaker monitor detect the agent's operative capability level from masked, partial traces, even when agents possess hidden capabilities the monitor does not know to look for or undergo capability shifts through tool composition? To study this, we formalize capability detection as a per-turn monitoring problem over tool use and introduce NegoBench, a multi-issue negotiation testbed with 19 tools and three capability levels (protocol-only action, independent tool use, and structured multi-tool coordination), with ground-truth labels assigned deterministically from execution metadata. To help a weaker monitor infer these levels from masked traces, we propose Stratified Compositional Inference (SCI), a monitoring scaffold that classifies each turn's observable evidence type and applies type-specific compositional indicators, so that coordination is recognized whether it manifests as inter-call data dependencies, disproportionately rich output from a single visible record, or goal-directed reasoning in the absence of tool-call records. Evaluating monitors weaker than the agents they supervise under varying trace visibility, reasoning access, and agent awareness of being monitored, we find that SCI outperforms both reference strategies across three monitor models and all visibility conditions (G-mean 0.53 vs. 0.46 baseline and 0.45 sequential), with the largest margins where intermediate tool-call records are available, indicating that trace-structure visibility is the primary driver of detection quality. Agent awareness of monitoring reduces detectability relative to Unaware agents, and we identify an intention-action gap: Subvert agents, which are told their chain-of-thought is private and the monitor sees only their tool-call inputs and outputs, dedicate 28.6% of their private reasoning to monitor-aware deliberation and 11.1% to strategic evasion (vs. 0.2% and 3.0% for Aware agents), yet produce statistically indistinguishable tool-call patterns and no additional detection reduction beyond passive awareness. This dissociation suggests that evasive intent could remain confined to private reasoning and does not propagate through the tool-call interface into observable behavior. Beyond detection, SCI recovers tool-level semantics without a tool inventory, and because its classification operates on observable trace structure rather than domain knowledge, the approach extends to any setting that produces execution traces with varying provider-controlled observability. Together, these results establish that capability detection from partial traces is feasible with structured monitoring scaffolds, and that investing in trace transparency yields larger detection gains than attempting to counter agent evasion. For reproducibility, our data and source code will be made available upon publication.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
capability detection, tool composition, AI safety, weak-to-strong monitoring, agent evaluation, LLM agents, scalable oversight, emergent capabilities

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_GUgfOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper focus on whether a weaker LLM can detect when a stronger agent composes tools to boost capabilities from masked, incomplete traces. The authors introduce NegoBench, a multi-issue negotiation benchmark. Execution logs deterministically label each round as L1 (communication only), L2 (isolated tool use), or L3 (composed tool use). Building on this, they propose SCI: it first classifies the evidence format of each round (bundled, itemized, or narrative), then applies corresponding rules to identify tool composition.

Critical strengths

This experimental setting reflects a real-world need.

Critical weaknesses
  1. The paper has notable formatting flaws. In Figure 3, subfigures (a) and (b) overlap with each other, making it impossible to read the performance of the respective methods. In the appendix, only RQs 6, 8, 10, 12, and 13 are present, with the remaining RQs missing entirely. It feels like the authors did not carefully read the paper before submission.

  2. Figure 3 shows that SCI is basically tied with the baselines in most cases, so it is hard to see any clear advantage.

Questions for authors

The prompt for SCI feels overly hand-crafted, with so many rules that look tuned specifically to fit NegoBench's trace format. This undermines the "domain-agnostic" claim and makes one wonder whether the gains actually come from stratification or just a finer rubric. There is an ablation missing: giving the baseline the exact same detailed rules without any stratification.

2 comments

We thank reviewer GUgf for recognizing that "this experimental setting reflects a real-world need" and for the direct question about whether SCI's gains come from the stratification mechanism or the detailed rubric. The ablation below isolates each component and answers this question directly.

Q1: Stratification ablation

The review asks for "giving the baseline the exact same detailed rules without any stratification." We conducted exactly this four-method ablation on a CoT-Only subset (o4-mini monitor, K0 where the monitor does not receive the agent's tool inventory, Unaware agents, on randomly selected episodes). Four methods isolate each component.

  1. Original baseline, single-pass classification with standard capability definitions
  2. Unstratified rubric, verbatim SCI compositional indicators provided to the baseline prompt, no stratification step
  3. Stratified one-call, SCI rubric with evidence-type classification within a single inference call
  4. SCI (full), stratified architecture (Step 1 classifies evidence type and Step 2 applies type-specific compositional probes per group)
MethodG-Mean95% CI
Original baseline0.00[0.00, 0.00]
Unstratified rubric0.39[0.28, 0.48]
Stratified one-call0.53[0.43, 0.61]
SCI (full)0.74[0.66, 0.80]

The ablation confirms that stratification itself drives SCI's detection gains beyond the detailed rubric. Adding verbatim SCI compositional indicators to the baseline prompt (unstratified rubric) raises G-mean from 0.00 to 0.39 [0.28, 0.48]. The original baseline scores 0.00 on this subset because CoT-Only is the most demanding visibility condition. Action inputs and outputs are hidden, so composition must be inferred from the agent's reasoning alone, a setting where the baseline struggles to identify compositional signals. The baseline recognizes the individual operations of an executed composition but does not aggregate them into one coordinated operation, and it labels 168 of the 173 turns as L2, 5 as L1, and none as L3. The full 180-episode run of the same configuration with Unaware agents confirms that this result reflects the condition and extends well beyond the subset. On all 1,563 turns, SCI reaches a G-mean of 0.722 with an L3 recall of 53.2%, against 0.197 and 4.0% for the baseline, at a comparable specificity of 98.1% and 98.3%. SCI's compositional indicators are designed for exactly this condition, since they teach the monitor to read repeated rationales, named steps, and narrative coordination cues as evidence of a single coordinated operation. Even under the same single-call inference budget, adding evidence-type classification (stratified one-call) raises G-mean to 0.53 [0.43, 0.61], a 36% relative improvement over the unstratified rubric at an identical inference budget, demonstrating that stratification itself improves detection at no additional computational cost. The full two-step architecture reaches 0.74 [0.66, 0.80]. This decomposition demonstrates that SCI's gains extend well beyond those of a more detailed prompt. The evidence-type classification step provides a structurally distinct source of detection signal by routing different trace patterns to type-appropriate compositional indicators, enabling the monitor to recognize coordination whether it surfaces as bundled outputs, itemized operations, or goal-directed reasoning.

On the concern that the rules are tuned to NegoBench, the SCI indicators describe how coordination appears in a trace, as bundled outputs, itemized operations, or goal-directed reasoning, and contain no negotiation-specific tool names or scenario rules. Two results confirm this generality beyond the within-turn trace format. SCI's detection does not depend on the tool inventory (+0.01 between K0 and K1, against −0.03 and −0.04 for baseline and sequential), and SCI maintains a strong lead when we broaden L3 to also cover dependencies that agents carry across turns through dialogue history without composition operators, reaching a G-mean of 0.60 against 0.00 for the baseline.

W1: Formatting

We thank the reviewer for raising the formatting issues. The revised paper will separate the subfigures of Figure 3 so that each method's performance is clearly readable, and will renumber the research questions consecutively and update all cross-references.

W2: SCI performance relative to baselines

Averaged across all visibility conditions and monitors, SCI obtains a G-mean of 0.53 compared with 0.46 for baseline and 0.45 for sequential (RQ1, Section 4.2), with non-overlapping bootstrap 95% confidence intervals (SCI [0.492, 0.563], baseline [0.425, 0.490], sequential [0.416, 0.482]). SCI outperforms both reference strategies in all six visibility conditions, with the largest margins under Expanded visibility where intermediate tool-call records are available (SCI 0.50 to 0.56 vs. baseline and sequential 0.40 to 0.44). SCI maintains the highest G-mean in most awareness conditions (Figure 3(b)), and its detection does not require prior tool knowledge, showing near-zero change between K0 and K1 while baseline and sequential degrade (RQ3, Section 4.4). The advantage holds across all three monitor models (Appendix D.2). The appendix further compares SCI against alternative detection strategies including Profile-then-Probe and a rule-based dependency detector (Appendix D.5, RQ12). In combination with the ablation above, these results demonstrate that SCI provides a consistent and decomposable advantage in the conditions most relevant to deployment. The revised paper will foreground this conditional performance profile so that readers can identify where the method operates best.

Summary

The requested ablation, the corrected figures, and the per-condition results resolve all of the reviewer's concerns. The exact ablation the reviewer requested shows that stratification drives a 36% relative improvement over an identical rubric at the same inference budget, SCI reaches a G-mean of 0.722 against 0.197 for the baseline on the full 180-episode run of the most demanding visibility condition with Unaware agents, SCI outperforms both reference strategies in every one of the six visibility conditions with non-overlapping confidence intervals, and SCI maintains a strong lead when coordination spans multiple turns. The revised paper will also correct Figure 3 and the appendix numbering.

We would be deeply grateful if the reviewer would reconsider the soundness assessment and the overall score in light of these results, especially as the reviewer recognized that our setting reflects a real-world need. We would be glad to discuss any remaining points.

We thank Reviewer GUgf for the constructive feedback.

Reviewer_ArGCOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates capability detection in tool-using language agents, addressing the failure mode where individual benign tool calls compose into emergent capabilities that single-action runtime monitors overlook. The authors formalize capability detection as a per-turn monitoring problem and introduce NegoBench, a multi-issue negotiation testbed spanning 19 tools across cognitive, behavioral, and protocol categories, where ground-truth capability levels are deterministically extracted from unmasked execution logs. To enable weaker monitors to identify capability shifts from masked traces, the authors propose Stratified Compositional Inference, a three-step scaffolding framework that categorizes trace evidence into bundled, itemized, or narrative types, applies type-specific compositional indicators, and generates decoupled functional descriptions. Evaluating 599 episodes across six agent models under varied trace visibility and agent awareness settings, SCI achieves a G-mean of 0.53 compared to 0.46 for baseline and 0.45 for sequential monitoring. Furthermore, the study identifies an "intention-action gap": agents instructed to subvert monitoring dedicate substantial private reasoning to evasion, yet produce observable tool-call patterns statistically indistinguishable from unaware agents (p > 0.86).

Critical strengths

The paper moves beyond static policy-checking to formalize runtime capability transitions in multi-tool agent workflows. The discovery of the intention-action gap provides an empirical insight for AI safety and oversight: structured tool execution interfaces enforce a natural ceiling on behavioral evasion, showing that transparency in trace structure yields far greater monitoring efficacy than attempting to counteract latent evasive reasoning.

NegoBench provides deterministic ground-truth capability labels across 5,323 turns without relying on heuristic downstream outcome proxies. Furthermore, SCI's ability to accurately infer tool semantics under zero inventory knowledge (mean cosine similarity of 0.46 vs. 0.36 baseline) proves that structured behavioral evidence classification can effectively reconstruct tool utilities in black-box deployment regimes.

Critical weaknesses

NegoBench is confined to multi-issue bargaining games, where compositional tool interactions are mediated by explicit meta-planning constructs (plan_sequence, plan_parallel, plan_if_then_else). In more open-ended environments agents typically compose tools implicitly through iterative context accumulation without structured composition blocks. The paper does not demonstrate whether SCI generalizes to traces lacking explicit multi-tool dispatch constructs.

As shown in the supplementary analysis, replanning and recovery behavior significantly degrades SCI’s detection performance (beta = -0.067, p = 0.007$) while leaving baseline strategies unaffected. Because SCI relies on regular structured composition patterns, disruptions caused by error-recovery or backtracking introduce trace noise that misleads the evidence classifier.

Questions for authors

How does SCI perform when applied to agent traces where multi-step tool dependencies occur implicitly across turns via dialogue history, rather than within explicit within-turn composition wrappers?

Given that mid-composition replanning induces a significant detection penalty on SCI, how can the evidence-type classifier be made resilient to dynamic branching and recovery actions without increasing LLM query overhead?

1 comment

We thank reviewer ArGC for the review and for recognizing 1) that our work "moves beyond static policy-checking to formalize runtime capability transitions", 2) that the intention-action gap "provides an empirical insight for AI safety and oversight", and 3) that NegoBench provides "deterministic ground-truth capability labels" without "relying on heuristic downstream outcome proxies". The questions about cross-turn generalization and recovery resilience shaped the analyses below.

Q1 and W1: Cross-turn implicit dependencies

To measure how much coordination the within-turn definition captures, we broadened the definition of L3 to also cover implicit cross-turn tool dependencies, where agent actions depend on information from prior turns' tool outputs carried through dialogue history outside within-turn composition operators. An LLM annotator (Gemini 3 Flash) identified such dependencies and we re-evaluated the existing monitor predictions under this broadened definition. Comparing SCI against the baseline on this broadened definition, SCI achieves a G-mean of 0.60 while the baseline remains at 0.00, confirming that SCI continues to outperform even when the L3 definition is expanded to include cross-turn coordination. The baseline scores 0.00 because it recognizes individual tool operations but does not aggregate them into a coordinated operation. SCI's compositional indicators capture exactly this coordination, across turns as well as within them. The revised paper will include this cross-turn analysis in the appendix. This result shows that SCI's evidence-type classification captures coordination whether the agent expresses it through explicit composition operators or accumulates it across turns through dialogue history, which is the open-ended setting the reviewer describes.

Q2 and W2: Replanning and recovery behavior

Across our evaluation episodes with the o4-mini monitor and Unaware agents under CoT-Only visibility, 48.3% contain a recorded replanning call. On these episodes, SCI reaches a G-mean of 0.695 against 0.171 for the baseline, a fourfold (4×) advantage, and stays close to its G-mean of 0.752 on episodes without replanning. SCI therefore retains a decisive advantage when recovery calls alter the trace.

To improve the resilience of SCI to recovery-induced trace disruptions, a recovery-aware mechanism can be added directly on top of the current architecture, either through prompt-level instructions that guide the evidence-type classifier to recognize plan-execute-recover phases, or through a preprocessing layer that segments trace subsequences before Step 1. Both approaches operate within the existing inference budget, directly addressing the concern about additional LLM query overhead. The revised paper will discuss this extension in the limitations.

Additional analysis: Stratification ablation

We also isolated the contribution of SCI's evidence-type classification on randomly selected CoT-Only episodes (o4-mini monitor, K0, Unaware agents). Giving the baseline the verbatim SCI compositional indicators without stratification raises G-mean from 0.00 to 0.39. Adding evidence-type classification within the same single inference call raises it further to 0.53, a 36% relative improvement at an identical inference budget, and the full two-step SCI reaches 0.74. Over all episodes of this configuration, SCI reaches a G-mean of 0.722 against 0.197 for the baseline. Stratification is therefore a decisive source of SCI's gains, beyond any additional rubric detail.

Summary

Taken together, these analyses fully resolve both of the reviewer's questions. SCI detects coordination well beyond explicit composition operators, reaching a G-mean of 0.60 against 0.00 for the baseline when L3 spans multiple turns, it retains a fourfold (4×) advantage on episodes with replanning (0.695 against 0.171), and the stratification ablation confirms that evidence-type classification is a decisive driver of these gains.

Since the reviewer found that our formalization of runtime capability transitions moves beyond static policy-checking and that the intention-action gap provides an empirical insight for AI safety and oversight, we would greatly appreciate a further increase of the score in light of these additions, and we welcome any further questions.

We sincerely thank Reviewer ArGC for the recognition of the paper's contributions and for the insightful questions that have strengthened the work.

Reviewer_X7G9OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces NegoBench to study whether weaker monitors can detect compositional tool use from partially observed agent traces. It proposes SCI, which adapts monitoring guidance to different evidence types, and evaluates its effectiveness across multiple models and observation conditions.

Critical strengths
  1. The paper addresses a relevant problem in monitoring tool-using agents.
  2. The experiments cover multiple models and provide relatively comprehensive analyses of visibility and monitoring awareness.
  3. The framework and figures are clear and easy to follow.
Critical weaknesses
  1. The capability levels may reflect how agents call tools more than how well they coordinate them. Agents performing similar steps could receive different labels simply because one uses a predefined composition operator and the other uses ordinary tool calls.
  2. The experiments do not clearly show how much SCI’s evidence-type classification contributes to the gains. Removing this step and comparing methods with the same inference budget would strengthen the evidence.
  3. The evasion experiments rely on prompt instructions. They provide limited support for the broader claim that execution interfaces limit evasion, especially without testing agents trained to hide their behavior.
  4. Too much important content is placed in the appendix, including additional baseline comparisons and the limitations. Bringing the key results and discussion into the main paper would make the contribution easier to assess.
Questions for authors

I am particularly interested in the motivation and broader implications of this work:

  1. Beyond improving detection scores, what is the main insight this work provides about monitoring hidden and emergent capabilities in tool-using agents?
  2. Among the limitations discussed in Appendix H, which do you consider the most important, and how does it affect the practical use of SCI?
3 comments

We thank reviewer X7G9 for recognizing 1) the relevance of monitoring tool-using agents, 2) experiments that "cover multiple models and provide relatively comprehensive analyses of visibility and monitoring awareness", and 3) a framework and figures that are "clear and easy to follow". The questions about stratification and scope shaped both the ablation and the revised paper.

We address the questions below with a stratification ablation that decomposes SCI's (Stratified Compositional Inference) contribution into rubric detail, stratification mechanism, and inference budget, and evidence from the testbed design and a cross-turn analysis that grounds the capability labeling.

Q1: The main insight

Trace-structure visibility is the primary lever for improving capability detection, ahead of any countermeasure against agent evasion. Investing in trace transparency (making intermediate tool-call records available to monitors) yields larger detection gains than attempting to counter evasive reasoning. This reframes the monitoring design problem. The productive investment is in controlling what composition structure the monitor can observe, which avoids an arms race against evasion altogether.

Q2: Scope and practical use of SCI

The single-domain scope is the most important consideration. Although the empirical evaluation uses a single domain, we chose the multi-issue negotiation setting because it provides a level of fine-grained labeling that other agent evaluation settings cannot offer. Its deterministic utility trajectories (per-issue scores, Pareto distance, and turn-over-turn deltas) allow us to uniquely attribute how each tool call, action, and strategic decision contributes to the agent's overarching goal, and capability labels are assigned deterministically from execution metadata without downstream outcome proxies or human annotation.

For practical use, SCI is directly applicable to any agent system that produces reasoning traces or tool-call records. The visibility conditions studied in the paper (CoT-Only where reasoning traces are visible but action I/O is hidden, Action-Only where action I/O is visible but reasoning is hidden, and CoT-Action where both are visible) map onto the trace access levels that real deployment environments provide. SCI's stratification mechanism operates on structural properties of these traces independently of negotiation-specific features and can be extended to other agent domains where such traces are available.

W2: Stratification ablation

To isolate whether SCI's evidence-type classification drives the detection gains beyond a more detailed rubric, we conducted a four-method ablation on a CoT-Only subset (o4-mini monitor, K0 where the monitor does not receive the agent's tool inventory, Unaware agents, on randomly selected episodes).

  1. Original baseline, single-pass classification with standard capability definitions
  2. Unstratified rubric, verbatim SCI compositional indicators provided to the baseline prompt, no stratification step
  3. Stratified one-call, SCI rubric with evidence-type classification within a single inference call
  4. SCI (full), stratified architecture (Step 1 classifies evidence type and Step 2 applies type-specific compositional probes per group)
MethodG-Mean95% CI
Original baseline0.00[0.00, 0.00]
Unstratified rubric0.39[0.28, 0.48]
Stratified one-call0.53[0.43, 0.61]
SCI (full)0.74[0.66, 0.80]

The original baseline scores 0.00 on this subset because CoT-Only is the most demanding visibility condition. Action inputs and outputs are hidden, so composition must be inferred from the agent's reasoning alone, a setting where the baseline struggles to identify compositional signals. The baseline recognizes the individual operations of an executed composition but does not aggregate them into one coordinated operation, and it labels 168 of the 173 turns as L2, 5 as L1, and none as L3. The full 180-episode run of the same configuration with Unaware agents confirms that this result reflects the condition and extends well beyond the subset. On all 1,563 turns, SCI reaches a G-mean of 0.722 with an L3 recall of 53.2%, against 0.197 and 4.0% for the baseline, at a comparable specificity of 98.1% and 98.3%. SCI's compositional indicators are designed for exactly this condition, since they teach the monitor to read repeated rationales, named steps, and narrative coordination cues as evidence of a single coordinated operation.

Adding verbatim SCI compositional indicators to the baseline prompt (unstratified rubric) raises G-mean from 0.00 to 0.39 [0.28, 0.48]. Even under the same single-call inference budget, adding evidence-type classification (stratified one-call) raises G-mean to 0.53 [0.43, 0.61], a 36% relative improvement over the unstratified rubric at an identical inference budget, demonstrating that stratification itself improves detection at no additional computational cost. The full two-step architecture reaches 0.74 [0.66, 0.80]. This decomposition demonstrates that SCI's gains extend well beyond those of a more detailed prompt. The evidence-type classification step provides a structurally distinct source of detection signal by routing different trace patterns to type-appropriate compositional indicators, enabling the monitor to recognize coordination whether it surfaces as bundled outputs, itemized operations, or goal-directed reasoning.

W1: Capability levels and tool invocation patterns

The testbed is designed so that agents are equipped to coordinate through composition operators. The composition operators themselves create data dependencies and branching structures between tool calls within a single turn, requiring the agent to plan a coordination structure in advance. Two agents performing similar steps therefore do not receive different labels for the same data flow. L2 denotes independent tool use with no inter-call dependencies, whereas in an L3 turn the agent links tool calls into a structured group, for example by reasoning over one call's output and parameterizing the next call on it (Section 2.2). The L3 label therefore reflects a genuine difference in how the agent coordinates its tools and is computed deterministically from execution logs, providing a reproducible ground-truth signal that does not rely on subjective coordination-quality judgments. We deliberately use structural labels because outcome-based labels would be unreliable, as the association between composition and negotiation welfare differs in sign across agent models (RQ5, Section 4.6). We further broadened L3 to also cover dependencies that agents carry across turns through dialogue history without composition operators, and SCI still detects this coordination with a G-mean of 0.60 against 0.00 for the baseline.

W3: Evasion via prompt instructions

Under the tested conditions, Subvert agents dedicate 28.6% of their private reasoning to monitor-aware deliberation and 11.1% to strategic evasion (vs. 0.2% and 3.0% for Aware agents, Table 1), yet all pairwise behavioral comparisons remain non-significant (Mann-Whitney U, Holm-corrected p > 0.86, Table 1). The observable behavior metrics show no detectable difference across the three awareness conditions. This dissociation constitutes the intention-action gap, in which agents invest substantial reasoning in evasion strategies, yet no detectable behavioral difference emerges in the tool-call record. Whether this holds under training-based evasion, where agents might learn to exploit interface-level degrees of freedom beyond prompt-level reasoning, is a natural next step that the revised paper will discuss as a direction for future work, building on the striking dissociation between intent and action that our experiments establish.

W4: Appendix placement

We thank the reviewer for this constructive suggestion. The scope of the experimental evaluation is substantial, and we recognize that the current placement of key results in the appendix makes the contribution harder to assess. The scope discussion (Appendix H), the per-monitor breakdowns (Appendix D.2, RQ6), and the alternative-strategy comparison (Appendix D.5, RQ12, including Profile-then-Probe and the rule-based dependency detector) all bear directly on the paper's claims and should be accessible in the main body. The revised paper will move these results and the limitations into the main text.

Summary

We believe the responses above fully resolve each of the reviewer's concerns. In particular, the stratification ablation establishes that evidence-type classification alone yields a 36% relative improvement over an identical rubric at the same inference budget, SCI reaches a G-mean of 0.722 against 0.197 for the baseline on the full 180-episode run of the most demanding visibility condition with Unaware agents, the capability labels are grounded in how the agent structurally coordinates its tools, and SCI retains a decisive lead even when coordination spans multiple turns (G-mean 0.60 against 0.00).

In light of these additions, we would be sincerely grateful if the reviewer would consider raising their score, given also that the reviewer recognized the relevance of the problem, the comprehensive analyses of visibility and monitoring awareness across multiple models, and the clarity of the framework and figures, and we would be glad to provide any further clarification during the discussion period.

We thank Reviewer X7G9 for the thoughtful feedback.

Comments on the paper

1 comment
Area_Chair_YbGsOpenReview ↗

Reminder: Please respond to the author rebuttal

Dear Reviewers, the authors have responded to your reviews with additional analyses, including the requested stratification ablation, and the author-reviewer discussion ends soon. Please read their response, post a brief reply on whether your concerns are addressed, and update your scores if your assessment has changed. Thank you!

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.