acceptodds
Under review as a conference paper at ICLR 2027

Measuring Say-Do Consistency in Tool-Using Agents: A Claim-Execution Traceability Framework

Abstract

Agent evaluation today follows three paths—state-based, text-based, and goal-plan-action alignment—yet none directly compares what the agent claims to have done against what the tools actually recorded. Deployment risk arises precisely from this divergence: user decisions rest on agent statements. We introduce a criterion-construction procedure for measuring this say-do gap: a scenario-agnostic core (traceability) plus a domain-specific layer deriving claim-execution matching rules from the tool registry, business processes, temporal alignment, and exemptions. Instantiated as a paired text-tool instrument, the procedure runs seven models across three insurance write-operation scenarios (1,780 trajectories), with external replication on 1,980 -bench trajectories. Reading at three refinement levels, it confirms 182 over-promises (40.3% of candidates), every one a fabricated capability no tool supports, a pattern state-based evaluation cannot effectively detect. Transferability is supported by an independent team's timed derivation on an unseen domain and by replication on published benchmark trajectories. The contribution is the procedure, not any specific reading. We release the protocol, codebook, and all adjudication data.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
agent evaluation, say-do consistency, evaluation methodology for agents, benchmark blind spots, AI safety, tool-calling agents, measurement validity

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_zqNqOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper aims to measure the gap between a model's plan of action and actual tool calls it invokes. Outcome based benchmarks including tau-bench miss this since they solely focus on the final state. The authors propose a method of measuring this say-do inconsistency by checking what the model said it did against the tool call history.

The paper contributes an empirical study with 7 models in 3 scenarios producing 1799 conversations. The paper reports that 40.3% of tool calls in 10.1% of all conversations are claimed by the model as having completed the action without invoking the tool itself. The paper's findings are well validated since they use human-annotators in addition to their llm-as-a-judge pipeline and the replication on tau-bench conversations is a good way of showing that this problem exists in the real world.

Critical strengths

The problem is widely observed and needs a good metric, this paper does a good job in scouting related works and positioning itself. Outcome based benchmarks miss this failure mode, this blind spot could be exploited as a reward-hacking mechanism when benchmarks that measure similar metrics are used as RL environments. Related work is thoroughly presented and Table 8 positions this paper's contribution clearly against all other works.

The findings are technically correct and novel. Replication on tau-bench is a really good addition to this paper, it tests model outputs generated by a third party and find that over-promises exist in a popular benchmark as well.

Critical weaknesses

The paper is too dense to read, tons of numbers and details are overwhelming, especially in the introduction's first paragraph where space should be given to motivate ("why") and speak about broader applications ("so what") of the work. The paragraph on "Codebook freezing" is unnecessary in the main text, the space could be much better used to provide examples of the "four-step derivation procedure"

The biggest weakness for me is the structure (or lack thereof) in rigorously establishing which questions to nail down as proxies for the say-do consistency and then answering them methodically. The paper shifts in which questions it aims to answer quite a lot throughout the paper. Specifically I observe this in the results sections where the three observations use a lot of bespoke diction that doesn't connect to the original motivation of the paper. Examples of bespoke diction: "three refinement levels," "anchored escalation," "fabrication flag," "strata," "claim substrate," "rapport decoration," "candidate-adjacent records."

Authors do a good job in reporting where their methods and findings have limitations. However, there's a lack of clarity and scientific rigor throughout the paper which is not what is expected of ICLR papers. Stemming from my first weakness, the problem is widely known but this paper doesn't do a great job in explaining how their experiments answer their question.

Questions for authors

For the inter-rater six-category agreement to be 53.8%–62.5% is fairly low. The authors' defense is that the disagreements mostly fall between "anchored escalation" and "unverifiable" not across the over-promise line, but I would like to see an elaborated explanation as to why is this agreement as low as it is.

Many scenarios were built so the user demands contact with a supervisor or salesperson but the tools offer no way to do it. To what extent is the reported over-promise rate induced by this design, and would it hold if such tools are provided?

Reviewer_JHN4OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper tackles the problem of say/do consistency: an agent can promise to conduct a specific action but actually never execute that action at any point. If these two actions are not in sync this clearly makes agents less trustworthy. The paper proposes a way to measure this discrepancy and conducts some measurements across a few models on the tau-bench benchmark. The results show that agents often over-promise and under-deliver wrt tool calls.

Critical strengths
  • The problem setting is clearly timely. Audibility and traceability of LLM agent actions has become especially important due to increasing misalignment and reliability incidents. If agents do not dependably actually do what they say then this is clearly an issue that deserves attention.
Critical weaknesses
  • It is blatantly clear that the paper is completely comprised of AI generated text. Many sentences are overly formulaic, there is plenty of jargon, many terms are used without definitions, and most paragraphs do not have an understandable flow. This leads to substantive clarity issues throughout the whole paper. Despite repeated attempts of trying to make sense of this paper, I had a very hard time following along. I am not fully opposed to using AI to help with polishing. However, the usage of AI writing in this paper clearly goes beyond simple editing and appears to be substantial and fundamental. I find this very irresponsible.

  • The introduction only motivates the problem very vaguely. It is too short and does convey the intuition.

  • Related work is just a collection of (loosely) adjacent work without a clear flow.

  • The method description is again very vague. Also, there appear to be some experimental numbers and results here despite the fact that there is a separate results section later in the paper. Very confusing.

  • The experimental section is poorly organized and is split across multiple sections.

Questions for authors
  • None really that I can think of that would change my mind.
Reviewer_q8cgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper measures whether an agent's claims about completed actions are supported by tool-level evidence. It derives domain criteria from tool registries, business rules, natural-language action aliases, and time windows/exemptions, then pairs textual claims with invocation records and assigns anchored execution, anchored escalation, over-promise, unverifiable, intentive, or UNDETERMINED labels. The contribution is a measurement procedure and domain-adaptation methodology, rather than a new training algorithm.

For example, an agent may actually submit a refund-exception request but also claim to have emailed a supervisor. If the environment has no email capability, the first claim can be anchored while the second lacks support. The final database state may be correct despite the unsupported textual claim. Non-execution, unverifiability, and future intent need to be distinguished rather than uniformly labeled over-promising.

Critical strengths
  1. The paper clearly separates tool-policy correctness from whether textual claims are supported by execution evidence, identifying a practical failure type that final-state scoring can miss. Examples of erroneous completion claims in PASS trajectories on tau-bench show that the issue is not limited to the authors' custom insurance scenario.

  2. The release includes label-migration records, audit corrections, judging prompts, and some trajectory data. The main text discloses data-driven codebook revisions and moderate six-class agreement. Major counts can be recomputed directly, which is more useful than an unverifiable headline rate.

Critical weaknesses

Overall assessment: the current work is closer to a technical report than a sufficiently validated research contribution, and I recommend 2 / Weak rejection. Its organization of the say-do consistency problem, labeling criteria, and trajectory audit has practical value, but the current evidence does not establish a sufficiently clear and generalizable research contribution.

  1. Criterion construction, label correction, and reliability validation are not sufficiently separated. Discovering and revising rules, correcting labels, and reassessing agreement on the same materials does not replace independent annotation on held-out examples after freezing the rules. The current results are closer to an empirical audit under a particular workflow than validation of a reliable, generalizable measurement framework.

  2. A gap remains between the methodological claims and the released evidence. The paper emphasizes atomic checks of actions, entities, parameters, success status, and time windows, but some released evaluation inputs provide only tool names, without complete original arguments, return values, or success-status evidence. The frozen rule files and per-instance raw execution chains are also incompletely released. The authors need to clarify which rule versions and evidence specifications underlie the reported metrics before those results can support claims of an executable, reproducible method.

Questions for authors
  1. Can an independent team use a frozen, fully released codebook on new examples/domains not used for rule revision, and report binary and six-class confusion matrices, precision/recall, and disagreement adjudication?

  2. Why do the 120 released prompts list only tool names and omit R4/R5? Can the actual prompts, argument/return inputs, and final label versions used for the 90.2% recall result be mapped instance by instance? Please provide an independently frozen codebook and the necessary original execution records.

  3. Given the same registries, arguments/returns, examples, and rule descriptions, how do simple rule matching, a codebook-guided judge, and other dual-input judges compare in performance and human effort?

  4. Please distinguish response-level priority labels from genuinely claim-level coverage. Does measurement reliability hold for registered but unexecuted actions, incorrect arguments, failed returns, delayed execution, and non-write operations?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.