acceptodds
Under review as a conference paper at ICLR 2027

Investigate on Demand: Evidence-Qualified Safety Auditing of Agent Trajectories

Abstract

Trajectory auditing is essential for assessing the safety of LLM agents, as actions judged harmless or suspicious in isolation during execution may have different safety implications in the full trajectory context. One-pass auditing can miss cross-step violations in long trajectories due to context loss, while indiscriminate deep investigation increases computational cost and can introduce false alerts. We introduce IoD-Auditor, a hierarchical on-demand framework for offline trajectory safety auditing. Across trajectories, a lightweight escalation policy uses trajectory length and direct-audit confidence to select between one-pass auditing and multi-agent investigation. Within each investigated trajectory, an evidence-driven multi-agent procedure adaptively examines unresolved evidence gaps. A shared behavioral-evidence record connects authorization, behavioral stage, counterevidence, and observable consequences, guiding further inspection and evidence-qualified adjudication. Across four heterogeneous benchmarks, IoD-Auditor achieves a macro-average F1 of 82.4, outperforming evaluted SOTA one-pass, sequential, and multi-agent investigation baselines. It reduces average model calls by 24% relative to investigating every trajectory, while budget-matched comparisons and controlled ablations demonstrate the benefits of selective escalation and evidence qualification.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM agents, trajectory safety auditing, on-demand investigation, evidence-qualified

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_vvQoOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces IoD-Auditor, a framework for offline safety auditing of LLM-agent trajectories. It combines selective escalation, based on trajectory length and direct-audit confidence, with a TRACE-based multi-stage investigation pipeline. The investigation maintains a structured behavioral-evidence record covering authorization, action stages, supporting evidence, counterevidence, and defenses to guide further inspection and final adjudication. The paper evaluates the framework on four trajectory-safety benchmarks and reports improved classification performance over the evaluated baselines, together with fewer model calls than investigating every trajectory.

Critical strengths
  1. Relevant problem and sensible motivation. Trajectory-level safety auditing is an important problem for tool-using agents. The paper identifies a useful tension: one-pass judgments may overlook distributed evidence, while additional investigation can amplify unsupported suspicions and introduce false positives.

  2. Explicit attention to evidence and counterevidence. Separating authorization, attempted versus realized behavior, and the timing and effectiveness of defenses is a practically meaningful design choice. Requiring investigators to consider counterevidence is particularly appropriate for avoiding the conflation of suspicious behavior with established violations.

  3. Evaluation includes several useful controls. Beyond comparisons across four benchmarks, the paper includes random allocation at matched call budgets, routing and protocol ablations, and adjudicator-model comparisons. These experiments address relevant questions about investigation allocation and the sources of performance differences, although the concerns noted above limit the conclusions that can currently be drawn.

Critical weaknesses
  1. Limited technical contribution and insufficient novelty. The core investigation pipeline is explicitly inherited from TRACE’s Triage–Inspect–Judge framework; adaptive inspection and cross-step evidence aggregation are therefore not new contributions of this paper. The additions primarily consist of threshold-based routing using trajectory length and self-reported confidence, along with prompt instructions to record authorization, behavioral stages, counterevidence, and defenses. Although practically reasonable, these additions appear closer to heuristic scheduling and auditing-prompt refinements than a substantially new algorithm or reasoning mechanism. The budget-constrained objective is not accompanied by an optimization procedure, and the evidence-qualification predicate is primarily implemented through natural-language instructions. The paper therefore does not establish a sufficiently substantive methodological contribution beyond the existing framework.

  2. Missing simple but essential baselines leave the necessity of the investigation framework unproven. The paper does not compare against repeated independent audits with majority voting or single-agent iterative review using the same evidence checklist. The Ensemble-OR/AND baselines in Table 6 combine the direct and investigation branches and do not substitute for these comparisons. Moreover, IoD-Auditor uses more calls than TRACE on several datasets, while Appendix C.3 shows that changing the adjudicating model alone can alter F1 by approximately 11–14 points. Comparisons with matched models and actual inference budgets are needed to determine whether the gains come from the proposed evidence protocol and coordination structure, rather than additional computation, more detailed prompts, or adjudicator selection.

  3. The efficiency evaluation is incomplete and does not adequately support the central cost claims. Cost is measured primarily through model-call counts, despite differences in model choice and context length across roles. These calls are not equivalent units of actual expenditure. The paper does not report input/output token consumption, monetary cost, end-to-end latency, or throughput. In particular, controller–inspector interactions introduce sequential dependencies, whereas independent audits can potentially run in parallel; equal call counts therefore do not imply comparable response times. The paper also lacks performance–cost–latency comparisons with specialized safety judges or guard systems. Although their task settings may differ from offline trajectory auditing, appropriately adapted comparisons under clearly specified information-access conditions are needed to establish practical value.

  4. Important numerical inconsistencies undermine confidence in the evaluation. Table 4 reports 450 unsafe and 90 safe TraceSafe trajectories, whereas Table 2 reports 100% precision, 34.4% recall, and 67.2% accuracy for Full-GLM. Under the stated unsafe-class evaluation, the reported precision and recall imply approximately 45.3% accuracy on that class distribution, not 67.2%. The full-system F1 scores also differ substantially between Tables 2 and 3: 79.3 versus 73.1 on TraceSafe and 67.0 versus 71.7 on MCPHunt. The authors need to specify the evaluation samples, aggregation procedures, and run configurations for each experiment and provide confusion matrices. These discrepancies cannot be explained by ordinary rounding, and the principal results cannot currently be reconciled.

  5. The length-stratified results conflict with the stated routing policy. According to Section 4.2 and the implementation details, every trajectory with at least four steps is routed directly to investigation. Under matched configurations, IoD-Auditor and All-Investigate should therefore execute the same investigation branch on these trajectories. Nevertheless, Figure 4 shows substantial differences in call counts for length groups entirely above this threshold, including approximately 11 versus 19 calls in the six-to-ten-step group. Independent stochastic runs may introduce variation, but these differences cannot simply be attributed to selective routing when the routing decisions are identical. The authors need to clarify whether different internal budgets, implementations, or experimental configurations were used; otherwise, these results do not support the proposed explanation for the efficiency gains.

  6. The evaluation does not directly validate the claimed reasoning benefits of the evidence protocol. In Table 8, evidence qualification reduces false-positive rates but also lowers recall. On ATBench500, for example, recall decreases from 96.34% to 92.68% relative to “No qualification,” while F1 remains unchanged at 89.94%. This result is also consistent with a more conservative decision policy and does not, by itself, establish more accurate interpretation of behavioral evidence. Comparisons at matched recall are needed, together with independent evaluation of evidence-citation correctness, behavioral-stage identification, and defense-effectiveness assessment. The current trajectory-level classification metrics do not directly validate these capabilities, despite their central role in the claimed contribution.

Questions for authors

see weaknesses.

Reviewer_o56bOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

IoD-Auditor combines selective escalation with evidence-guided investigation for offline trajectory auditing. It extends the TRACE workflow by deciding which trajectories need deeper inspection and requiring stronger evidence before raising safety alerts.

Critical strengths
  1. The motivation is practical: deeper investigation can increase both cost and false positives.
  2. Explicitly considering authorization, counterevidence, and execution outcomes is useful.
  3. The evaluation includes four benchmarks, allocation baselines, and paired confidence intervals.
Critical weaknesses
  1. Several reported results need reconciliation. Tables 2 and 3 give different full-router results, and the TraceSafe metrics are difficult to reconcile with the class distribution in Table 4.
  2. The routing analysis is unclear. With the stated threshold, trajectories of at least four steps should enter the same investigation branch, yet Figure 4 shows substantial differences between IoD-Auditor and All-Investigate in these groups.
  3. The related work omits Wen et al., Learning Efficient Guardrails for Compliance-ICML 2026. Its focus on policy-compliance detection differs from general safety auditing, but PolicyGuard-4B and PolicyGuardBench are relevant comparisons for the overlapping setting.
Questions for authors
  1. Could you reconcile the results and evaluation populations across Tables 2–4?
  2. Why do IoD-Auditor and All-Investigate differ on trajectories that must enter investigation?
  3. How were routing thresholds selected, and do the savings hold in tokens and latency?
Reviewer_onb3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes IoD-Auditor, a hierarchical framework for offline safety auditing of LLM agent trajectories. It uses trajectory length and direct-audit confidence to selectively trigger deeper investigation, where a structured behavioral-evidence record guides adaptive inspection and final adjudication. Experiments across multiple benchmarks demonstrate improvements in auditing accuracy and computational efficiency compared to existing approaches.

Critical strengths

The paper addresses the important problem of balancing accuracy and computational cost in agent trajectory safety auditing. The authors evaluate their approach on four benchmarks, with budget-matched comparisons and ablation studies showing promising improvements in auditing performance and efficiency.

Critical weaknesses

The framework largely builds on TRACE, with a simple routing strategy based on trajectory length and LLM-reported confidence. However, the confidence scores are not calibrated, and the routing relies on manually selected thresholds. The effectiveness of this strategy depends on the reliability of LLM-reported confidence and the choice of thresholds, neither of which is sufficiently justified.

The paper shows that evidence qualification reduces false positives, but it is unclear whether this is because the system makes more accurate judgments or simply becomes more conservative. Also, the intermediate evidence assessments rely on LLMs, but the paper does not evaluate their accuracy or reliability, raising concerns about the effectiveness of the proposed evidence qualification.

Questions for authors
  1. Since the intermediate evidence assessments are generated by LLMs, how do the authors ensure their accuracy and reliability? Is there any evaluation of these judgments?

  2. How were the routing thresholds selected, and how reliable is uncalibrated LLM confidence for deciding when deeper investigation is needed?

  3. Table 2 reports 100% precision, 34.4% recall, and 67.2% accuracy for Full-GLM on TraceSafe. However, Table 4 shows 450 unsafe and 90 safe trajectories, which appears inconsistent with these metrics. Could the authors clarify that?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.