acceptodds
Under review as a conference paper at ICLR 2027

Auditing Agent Actions through Query-Conditioned Attribution

Abstract

LLM agents increasingly take consequential actions through interactions with users, policies, and external tools. Auditing these agents requires automated attribution of realized actions to their historical basis. However, existing attribution formulations do not provide question-specific traces for diverse auditing objectives. Additionally, when access to the acting model is limited (e.g., in API-only deployments), applicable methods commonly rely on costly input perturbations or external LLM analysis of complete trajectories. We therefore formulate , a new task that takes a natural-language auditing query as input and recovers the source and ordered intermediate evidence for the query-specified aspect of an action. We instantiate this task with , a benchmark comprising 1,396 auditing queries across policy basis, parameter provenance, failure propagation, and unsafe-behavior tracing. To enable efficient, query-specific attribution, we use small open-weight models as attribution proposers that combine query-conditioned gradient saliency with query-semantic relevance to rank history units. Our proposer consistently achieves stronger source and evidence rankings at lower inference cost than open-weight baselines, improving source MRR by up to 40.9% and evidence MAP by 42.1% with only two forward passes and one backward pass. Controlled evaluations confirm that our proposer improves attribution specificity by adapting its rankings to fine-grained changes in the auditing query. Building on a proposer ensemble, our end-to-end system surpasses the strongest frontier-model baseline in source accuracy (64.5% vs. 60.4%) while reducing empirical deployment latency by 29.9% relative to the fastest frontier API baseline. Code and data are available at https://anonymous.4open.science/r/Agent-Action-Attribution-0E7E.

Primary area:
datasets and benchmarks
Keywords:
Agent Action Attribution, Auditable Agent System

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_TbZkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies query-conditioned agent "action attribution", where the goal is to identify historical sources relevant to a specific auditing question about an agent's action. The authors construct A3 BENCH, which comprises 1000+ queries across four categories: policy basis, parameter provenance, failure propagation, and unsafe-behavior tracing.

They propose a gradient-based attribution method that uses small open-weight proxy models to compute token-level saliency scores for the observed action. These scores are combined with semantic relevance and aggregated into step-level rankings. The final methods uses an ensemble of proposers to recover the source (and intermediate) evidence.

The authors report improvements in attribution ranking quality and inference efficiency over existing methods, along with competitive results against frontier LLM baselines.

Critical strengths
  1. On query-conditioned attribution: I think explicitly conditioning attribution on an auditing question is a useful extension of existing work. Different questions about the same action can require different historical evidence, and the benchmark attempts to capture this through query-specific source and intermediate-evidence annotations. The human validation of these annotations is also useful though I'm still not sure if such a ground truth exists for complex agentic actions.

  2. Attribution methodology: The use of query-conditioned gradients from smaller proxy models is an interesting approach, esp when the acting model is only accessible through an API. The method is also cheaper than perturbation-based attribution, and the ablations suggest that query-conditioned gradients contribute meaningfully to performance. The experiments also show that the rankings change when the auditing query changes (ie. it is sensitive to the query).

Critical weaknesses
  1. On the task formulation and whether attribution is well-defined: I am not entirely convinced that the benchmark captures the complexity of agent action attribution. The authors assume that an action can be traced to an identifiable historical source and a sequence of intermediate evidence. However, agent behavior can arise from multiple interacting influences, diffuse information across a trajectory, or the model's learned behavior rather than a particular historical step. For example, how would this framework handle emergent misalignment, or agent errors that are not directly attributable to an earlier interaction?

    Although the general formulation allows multiple sources, every benchmark instance has exactly one annotated source. This seems like a fairly restrictive assumption, especially for the kinds of complex behaviors that motivate agent auditing. This makes me question how much the task actually tells us about why an agent took a particular action.

  2. On the proxy-gradient attribution: The method computes gradients with respect to a separate proxy model rather than the actual agent that performed the action. The gradients measure how historical tokens affect the proxy model's preference between the observed action and a constructed alternative. It is unclear to me why this should faithfully identify the historical information that influenced the acting model, rather than simply identifying information the proxy finds relevant to predicting the action.

    The experiments show that this is a useful ranking heuristic, but do not show that the recovered evidence actually influenced the original agent. I also think the baselines could be stronger, since existing attribution methods are evaluated without comparable query conditioning. This makes it difficult to separate the contribution of the proposed gradient methodology from the benefit of explicitly providing an auditing query.

Generalization across attribution types also seems uneven. The method performs well on policy-based attribution but underperforms Claude Sonnet 4.6 on parameter provenance. The authors also find that many of these errors come from selecting downstream mentions rather than the actual information source. This seems like a limitation of using gradient sensitivity.

Questions for authors
  1. On task definition: How does the proposed framework handle actions with multiple contributing sources or diffuse historical influences? What happens when an action arises from an agent error without an identifiable source in the visible trajectory? Why does the benchmark restrict every instance to exactly one source despite the general formulation allowing multiple sources?

  2. On proxy attribution: Why should gradients computed through a separate proxy model recover the historical factors that influenced the acting agent? Can the authors validate this with interventions on the open-weight acting model, or investigate how sensitive the attributions are to the choice of proxy model?

  3. On baselines: Could the authors compare against query-conditioned versions of existing attribution methods or simpler query-aware retrieval baselines?

  4. On generalization: Why does the method perform substantially worse on parameter provenance compared with policy-basis attribution?

3 comments
AuthorsOpenReview ↗

Key Takeaways from Our Rebuttal and Additional Experiments

We would like to briefly summarize four key findings from our rebuttal and revised manuscript:

  1. Task formulation (W1/Q1): We clarify that query-conditioned historical attribution addresses a practical auditing need grounded in established research. While acknowledging the single-source limitation, we emphasize that even this setting lacks systematic evaluation of query-conditioned source localization and multi-step evidence recovery. A3Bench represents a meaningful step toward more realistic attribution, with multi-source evaluation as a natural extension.
  2. Intervention study (Q2): On our open-weight subset, acting-model interventions show that proxy-selected evidence has stronger effects on action likelihood than token-budget-matched random controls. For example, the Granite proxy achieves greater excess ΔNLL than acting-model self-attribution (0.234 vs. 0.082).
  3. Query-conditioned baselines (W2/Q3): With identical query-derived contrasts, our proposer outperforms matched perturbation baselines by 43.1–62.1% in source MRR and 18.6–33.5% in evidence MAP.
  4. Parameter provenance (W3/Q4): We identify downstream-mention confusion as the main error pattern and show that applying our existing resolver to the Top-5 candidate union raises parameter-source Hit@1 from 36.5% to 51.5%. Overall, our method still outperforms Claude Sonnet 4.6 (64.5% vs. 60.4%) at lower latency.

Thank you again for your feedback. We hope these additional results help address your concerns.

AuthorsOpenReview ↗

Follow-up on Rebuttal and Additional Results

Thank you for your time and effort in reviewing our paper. As the author-reviewer discussion period is coming to an end, we would like to kindly follow up and ask whether our rebuttal and additional results have adequately addressed your concerns.

For convenience, we briefly highlight the main areas addressed in our rebuttal:

  1. Task formulation (W1/Q1)
  2. Intervention study (Q2)
  3. Query-conditioned baselines (W2/Q3)
  4. Parameter provenance (W3/Q4)

Thank you again for your time and consideration!

We would like to kindly follow up once more as the discussion period approaches its end. We would greatly appreciate it if you could share whether our rebuttal and additional experiments have addressed your main concerns, and whether they might warrant reconsideration of your initial assessment.

Thank you again for your time and consideration.

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

This paper addresses two challenges in agent action attribution: supporting different auditing objectives and maintaining efficiency under limited model access. It introduces a benchmark for evaluating query-conditioned attribution and a proposer that constructs alternative actions to rank historical units using two forward passes and one backward pass. Experiments show that the proposer achieves effective attribution rankings with small open-weight models while reducing inference cost.

Critical strengths
  1. Important problem and useful task formulation. Explicitly conditioning attribution on an auditing query captures the practical need to explain different aspects of the same agent action. The accompanying benchmark provides a useful evaluation resource.

  2. Clear presentation. The paper motivates its contributions through attribution specificity and efficiency under limited model access. The examples and task definitions make the distinction between sources and intermediate evidence understandable.

  3. Simple and efficient method. The approach requires neither proposer training nor access to the acting model’s internals. Results across multiple backbones, component ablations, and error analyses provide substantial evidence for its effectiveness.

Critical weaknesses
  1. The proposer comparison uses different input information. The proposed method explicitly uses the auditing query, while the baselines in Table 1 use only the history and target action. It is therefore unclear how much of the improvement comes from the query information versus the attribution method itself. Evaluating the baselines with the same query-derived objective would make this distinction clearer.

  2. The construction of alternative actions needs clearer explanation and validation. Figure 8 provides templates, but the rules for converting queries into concrete alternative actions are not fully specified. Since attribution scores depend on the contrast between the observed and alternative actions, different reasonable alternatives may produce different rankings. Detailed construction examples and sensitivity tests would help establish whether the results are robust to these design choices.

Questions for authors
  1. Could the baselines in Table 1 be evaluated using the same query-derived contrastive objective as the proposed method? This would help distinguish the benefit of query information from that of the attribution method itself.

  2. Could the authors provide concrete examples and rules for converting queries into alternative actions? How much do the attribution rankings change when different reasonable alternatives are used for the same auditing question?

  3. Section 5.3 reports improved agent accuracy after aggregating action-level scores by agent. How exactly are these scores combined, and does the same aggregation rule also improve performance on the hand-crafted subset of Who&When?

2 comments
AuthorsOpenReview ↗

Key Takeaways from Our Rebuttal and Additional Experiments

We would like to briefly summarize three key findings from our rebuttal and revised manuscript:

  1. Query-conditioned baselines (W1/Q1): With identical query-derived contrasts, our proposer outperforms matched perturbation baselines by 43.1–62.1% in source MRR and 18.6–33.5% in evidence MAP.
  2. Contrast construction clarification and robustness (W2/Q2): We clarify our deterministic query-to-contrast routing with concrete examples and evaluate our design robustness through query paraphrases, contrast paraphrases, and three alternative action choices. Performance remains stable under wording variations, with reasonable robustness to different contrast choices.
  3. Agent-level aggregation in Who&When (Q3): We sum the attribution scores in our aggregation across an agent's historical actions. This improves algorithm-generated accuracy from 41.27% to 51.59%, while hand-crafted accuracy remains unchanged at 58.62%.

Thank you again for your feedback. We hope these additional results help address your concerns.

Thanks for your response and most of my concerns have been addressed. I will keep my score.

Reviewer_3BRMOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces A3Bench, a benchmark for attributing language agent actions to interaction history based on natural language auditing queries. The proposed method contrasts a realized action against a query derived alternative action, combining gradient saliency and semantic relevance from small proxy models to rank history units. A proposer ensemble paired with a small resolver identifies the primary source step, while rank fusion selects intermediate evidence. Empirical evaluations assess source ranking, evidence recovery, query sensitivity, and transfer performance on established attribution benchmarks.

Critical strengths
  • Framing action attribution around an auditing query is a practical and meaningful direction. A single agent action can stem from different causes or rely on different context inputs, and the benchmark explicitly evaluates this distinction.

  • The proposed approach operates without internal access to the primary acting model and runs efficiently using small open weight models. It achieves source accuracy comparable to proprietary frontier models while transferring effectively to existing attribution tasks without retraining.

Critical weaknesses
  • The comparison with perturbation baselines does not control for query information. These baselines are evaluated in their standard target conditioned setup, whereas the proposed method receives a query derived action contrast. This makes it difficult to separate the benefit of query information from the intrinsic quality of the attribution technique. Evaluating a leave one out baseline using the same contrast prompt would provide a fairer comparison.

  • Query conditioning relies heavily on manual rules. Constructing alternative actions requires handcrafted templates and keyword matching specific to the four predefined relations, requiring new templates for any unobserved relation type. The current implementation supports predefined relations, and its applicability to other auditing questions remains untested. In addition, the relevance component compares token embeddings without context.

  • The query conditioning experiment changes the auditing focus, but does not test robustness to rephrasing or transfer to unseen relation types. The transfer benchmarks align closely with the failure and safety relations already present in A3Bench, leaving open whether the pipeline can handle genuinely new auditing tasks.

  • The contribution of proxy gradients relative to standard retrieval remains unclear. The full test evaluation does not include a query based retriever or reranker, leaving open how much the gradient component adds over a matched retrieval baseline.

Questions for authors
  1. Are there results comparing the proposed method against a leave one out baseline using the same action contrast, or against a lightweight query based retriever?

  2. How does query parsing handle wording that does not match the relation keywords, or that matches more than one relation?

If the authors can address the identified weaknesses and my questions, I would be willing to reconsider my score.

3 comments
AuthorsOpenReview ↗

Key Takeaways from Our Rebuttal and Additional Experiments

We would like to briefly summarize the key findings from our rebuttal and revised manuscript:

  1. Query-conditioned baselines (W1/Q1): With identical query-derived contrasts, our proposer outperforms matched perturbation baselines by 43.1–62.1% in source MRR and 18.6–33.5% in evidence MAP.
  2. Contextualized retrieval (W2/W4): Our proposer consistently outperforms E5 retriever (12.4–36.6% higher source MRR). Incorporating E5 into our relevance component does not yield consistent improvements in attribution performance, despite requiring additional encoder computation, supporting the effectiveness and efficiency of our lightweight design.
  3. Robustness (W2/W3): Our additional experiments show stable performance under query and contrast paraphrasing, with further tests confirming reasonable robustness to alternative-action choices.
  4. Transferability and extensibility (W2/W3/Q2): Our generic contrast provides a fallback for unmatched or ambiguous queries without relation-specific templates and shows promising transfer on a new user-preference attribution relation. Lightweight relation-specific adaptation further improves source MRR across all three backbones.

Thank you again for your feedback. We hope these results help address your main concerns.

AuthorsOpenReview ↗

Follow-up on Rebuttal and Additional Results

Thank you for your time and effort in reviewing our paper. As the author-reviewer discussion period will end in about two days, we would like to kindly follow up and ask whether our rebuttal and additional results have adequately addressed your concerns.

For convenience, we briefly highlight the main areas addressed in our rebuttal:

  1. Query-conditioned baselines (W1/Q1)
  2. Contextualized retrieval (W2/W4)
  3. Robustness (W2/W3)
  4. Transferability and extensibility (W2/W3/Q2)

Thank you again for your time and consideration!

Thank you for the additional experiments. The query-matched baselines and E5 comparisons address my main concerns about comparison fairness and the benefit beyond retrieval. Template dependence and the limited transfer evaluation remain limitations, but my main concerns have been addressed. I have raised my score to 3.

Comments on the paper

2 comments
Area_Chair_JVAoOpenReview ↗

Please engage in the discussion

Reviewers: if you have not responded to the authors' reply, please read it and engage in the discussion. Thanks!

AuthorsOpenReview ↗

Final Author Response and Rebuttal Summary

We sincerely thank the Area Chair and all three reviewers for their time and constructive feedback. As the discussion period concludes, we would like to provide a structured summary of our core contributions, reviewer feedback, and how the main concerns have been addressed during rebuttal.

1. Core contributions

Our work introduces (i) a new query-conditioned agent action attribution task and A3Bench, with 1,396 auditing queries and source/intermediate-evidence annotations; (ii) an efficient proxy-based attribution method that operates without access to the acting model's internals; and (iii) an empirically effective system that surpasses the strongest frontier-model baseline in source accuracy (64.5% vs. 60.4%), while achieving 29.9% lower measured latency than the fastest frontier API baseline.

2. Recognized strengths

  • Well-Motivated Task Formulation (3BRM, xzVn, TbZk): Addressing the practical need for query-conditioned attribution, where different auditing questions about the same agent action require different historical evidence;
  • Useful Benchmark and Evidence Annotations (3BRM, xzVn, TbZk): Providing a valuable evaluation resource with query-specific source and intermediate-evidence annotations, including human validation highlighted by TbZk;
  • Practical and Efficient Attribution Method (3BRM, xzVn, TbZk): Using lightweight open-weight proxy models to perform query-conditioned attribution without accessing the acting model's internals, while avoiding costly perturbation-based computations;
  • Strong Empirical Support (3BRM, xzVn, TbZk): Demonstrating competitive frontier-model performance and transfer without retraining (3BRM), effectiveness across backbones and ablations (xzVn), and meaningful gradient contributions and query-sensitive rankings (TbZk).

3. Rebuttal outcomes and responses to reviewer concerns

During rebuttal, both reviewers 3BRM and xzVn explicitly acknowledged that their main concerns had been addressed following new experiments on fair baseline comparisons, contextualized retrieval, contrast robustness, and generalization to a previously unseen auditing relation. Reviewer 3BRM raised their score to 3, while reviewer xzVn maintained their positive score of 3.

While reviewer TbZk has not yet provided further feedback during author-reviewer discussions, we carefully addressed all four major concerns through targeted experiments and clarifications:

  • Task formulation (W1/Q1): To address whether attribution is well-defined for complex agent behavior, we clarified that query-conditioned historical attribution is a distinct, practically important auditing task supported by prior literature. Single-source annotation defines a controlled evaluation scope, not an assumption that all actions have a unique historical cause; multi-source attribution remains a natural extension.
  • Proxy attribution validity (W2/Q2): To address whether proxy gradients identify evidence influencing the acting model, we conducted the requested interventions on 255 examples. Results show that proxy-selected evidence has stronger effects on the acting model's action likelihood than token-budget-matched random controls, with some proxies outperforming acting-model self-attribution (0.234 vs. 0.082 excess ΔNLL).
  • Baseline fairness (W2/Q3): To test whether our gains stem merely from additional query information, we evaluated baselines with identical query-derived contrasts. Our method maintains substantial improvements (43.1–62.1% higher source MRR). Additional E5 comparisons further establish benefits beyond query-aware retrieval.
  • Parameter provenance (W3/Q4): To explain the weaker performance on parameter provenance, we highlighted our existing error analysis, which identifies downstream-mention confusion as the main failure mode. Our existing resolver-based mitigation substantially improves source identification (Hit@1: 36.5% → 51.5%), demonstrating that this limitation can be partially addressed without retraining.

We believe these additional results and clarifications substantially strengthen the empirical and conceptual support for our work and directly address the principal concerns raised in the reviews. We sincerely appreciate the Area Chair's consideration of the revised manuscript and the full rebuttal evidence in the final assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.