acceptodds
Under review as a conference paper at ICLR 2027

Who Is Allowed to Act? Benchmarking Decision Rights in Document-Grounded AI Agents

Abstract

An agent’s ability to execute an action does not establish its right to do so. As AI agents begin to act on information spread across policies, approvals, records, budgets, and other documents, they must determine whether the available information actually gives them sufficient grounds and authority to take a proposed action. We define decision-right reasoning as the task of determining whether a proposed action is permitted, not permitted, or cannot be resolved from the available evidence. We organize this decision around six factors, covering the sufficiency of the evidence, the validity of the relevant authority, the reversibility and criticality of the action, its temporal validity, the requirements for recovery, and any applicable cost or budget constraints. Each judgment is linked to the documents that support it, so that the final decision and its underlying basis can be examined separately. We have tested this formulation through independent human annotation across multiple document-based scenarios, followed by a separate review of the supporting decisions and recovery and temporal requirements. The results so far show that people can apply the decision rules consistently and trace their judgments to specific evidence. We use this formulation to construct a benchmark with controlled changes to the underlying documents, cases where the available information does not support a definite decision, and checks for shortcuts based on candidate position or surface wording. The benchmark examines whether language models can distinguish actions that can be carried out from actions for which the available evidence and authority provide sufficient grounds to proceed.

Primary area:
datasets and benchmarks
Keywords:
AI agents, decision-right reasoning, agent evaluation, document-grounded reasoning, large language models, datasets and benchmarks

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_mRYnOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This benchmark attempts to measure how often AI make decisions when the evidence or permission given is not sufficient for making those decisions. However, the benchmark is unmotivated by the literature, very unclear, poorly constructed, and poorly analyzed so I mark clear reject.

Critical strengths

None

Critical weaknesses
  • Ungrounded Task and Scenario Construction: The benchmark relies on the EACTRN axes to establish ground truth, but it's not clear to me why these axes are the ones to focus on. Only A (operational authority) is grounded in the literature. Without literature grounding, ECTRN seem like random axes generated by AI to try to be comprehensive. Further, it's not explained how the tasks in the benchmark are generated or whether they're necessary at all or relate to downstream impacts. This makes me believe this benchmark is a benchmark for creating a benchmark's sake rather than attempting to measure something useful. Human study is very unclear and there is no appendix explaining additional details.
  • Shallow Analysis: This is a benchmark paper and therefore it should provide thorough analysis of model performance on the benchmark. Instead, the results section just basically tells me that models over-authorize. The error analysis is just 2 confusion matrices that don't actually dig into the reasoning traces of these models. Much deeper review is needed. Please also quantify uncertainty when possible.
  • Very Old Models: The paper only evaluates Llama 3.3 70B which released in Dec 2024 and Mistral-Large-Instruct-2411 which released in Nov 2024. Since then, frontier models have gotten much better. I cannot tell if this benchmark would be saturated by current models (if it's saturated, it's not a useful benchmark).
Questions for authors
  • In organizational settings, is liability or other accountability actually defined by EACTRN axes? If not, why focus on EACTRN here?
  • Why did the models over-authorize? I would like a much deeper analysis than currently provided.
Reviewer_i79AOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper builds a benchmark to analyse the authorisation of AI actions, which is assessed using a six-axis framework, including evidence sufficiency, authority validity, reversibility and criticality, temporal validity, recovery requirements, and cost or budget constraints. The combination of these six axes defines three labels, including PERMITTED, NOT PERMITTED, and UNKNOWN. Two LLMs were evaluated on the established benchmark.

Critical strengths

The paper raises an important concern: AI actions must be authorised. It proposes a framework to analyse this concern and builds a benchmark.

Critical weaknesses

I feel that the experiments are insufficient. Only two LLMs were evaluated on the full benchmark. This makes it difficult to assess the difficulty of the benchmark and whether the benchmark can be widely used.

The authors emphasise six-axis decision-right reasoning to produce the action-right classification (ARC), but only the final ARC labels were analysed in the main experiments. This is a mismatch between the claimed reasoning framework and the measured classification performance.

Questions for authors

Can the authors explain why they chose six axes rather than more or fewer? Why are there no experiments analysing the roles of the six axes in the LLMs’ predictions? Can the authors evaluate more LLMs, include a strong baseline, and conduct ablation experiments? Positive answers to these questions could lead me to increase my recommendation score.

Reviewer_9d4BOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the tasks of “decision-right reasoning”, which determine whether an AI agent is entitled to carry out a proposed action, given information spread across multiple documents. The task is formulated as classifying each candidate action as PERMITTED, NOT_PERMITTED, or UNKNOWN. The judgment is made along six axes: (1) evidence sufficiency, (2) operational authority, (3) criticality and reversibility, (4) temporal validity, (5) recovery correctness, and (6) binding resources. The axis labels are aggregated with a failure-dominant rule: if any axis fails, the action is not permitted; if none fails but some remain unresolved, the outcome is unknown. The authors build a benchmark of 360 tasks and 1,080 candidate actions over fictional policies and documents, including 84 controlled pairs in which only a single decision-relevant fact is changed.

Critical strengths
  1. In an organizational policy setting, the motivation that what an agent can do differs from what it is allowed to do is timely. Treating UNKNOWN as a candidate answer, rather than forcing a binary decision, also seems closer to how such decisions arise in practice.
Critical weaknesses
  1. The paper does not provide enough information on dataset construction and experiments to make the work reproducible, and there are inconsistencies in what is reported.

    • The data construction process is barely described. In the main text, the inter-annotator agreement results in Section 4 are effectively the only evidence of data quality, while the lexical diagnostics and duplicate checks in Section 3 are mentioned without any results. The appendix covers construction only for the 60- and 120-task stages and says nothing about how the 240 tasks added in the final expansion (I27_121–360) were built. This is a significant gap in the transparency and documentation of the construction process. In addition, the term “AI” are frequently mentioned throughout the appendix suggest that this is a synthetic dataset, and evidence is needed on how well it reflects real-world [A]. The human annotation in Section 4 only reports agreement among annotators; it does not address whether the scenarios themselves are realistic.
    • The AI Use Statement is inconsistent with the appendix and should be revised. The statement reads: “AI tools provided auxiliary support during manuscript preparation, including language polishing. The authors are responsible for the final content and conclusions.” However, the appendix shows that AI also played a central role in data generation and auditing.
    • The experiments in Section 6 appear intended to show that the benchmark is challenging for LLMs. However, it is unclear whether any prompt optimization was performed and which prompts were actually used. It is also important to know how frontier models perform, beyond Llama-3.3-70B and Mistral-Large-2411.
  2. The six axes proposed in Section 2 are central to this work, yet apart from operational authority, no references are given to support their definitions. Grounding the axes in relevant prior work would help establish their validity.

[A] Bean et al., Measuring What Matters: Construct Validity in Large Language Model Benchmarks, NeurIPS 2025

Questions for authors

Please See Weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.