acceptodds
Under review as a conference paper at ICLR 2027

Say It Differently, Act Differently? Metamorphic Testing of Operational Behavior Consistency in AI Agents

Abstract

LLM-powered AI agents increasingly perform real-world tasks through multi-step operations, making operational consistency an important aspect of reliability. Yet existing evaluations typically consider only a single expression of a user request, overlooking the many ways in which the same intent may be expressed. This raises a key question: do intent-equivalent requests lead to operationally consistent agent behavior? To investigate this question, we propose AgentMT, an MT-based framework for evaluating black-box AI agents under intent-preserving prompt variations. AgentMT transforms prompts along tone, language, and formulation, executes them under the same agent setting, and represents observable executions as operation graphs to compare permission decisions, information disclosure, and execution trajectories. Across three datasets and commercial AI agents, follow-up prompt variations induce operational inconsistency in all 27 model–dataset–transformation combinations, reducing mean operation-graph similarity by up to ; significant task-level consequence inconsistency is observed in 22 of 27 model–transformation combinations. We further assess generalization using GPT-6 Astra, multi-turn clarification, Claude Code, and Codex CLI. Our findings show that preserving user intent does not guarantee consistent agent operation, motivating systematic evaluation of operational robustness across diverse user expressions.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI agent, Metamorphic testing, Operational Consistency

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

The paper proposes AgentMT, a metamorphic-testing framework for measuring the consistency of LLMs’ behaviour when the same request (in terms of intent) is requested differently. The prompts are sourced from three agent benchmarks (AI Agent Permissions, AgentCIBench, TRAJECT-Bench), and are transformed along tone, language, and formulation (specifically sentence reordering and paraphrasing differently). Agent executions are logged as typed events and converted into operation graphs, and consistency is measured at three levels: graph similarity, LLM-judged output similarity, and consequence consistency. Experiments with Gemini 2.5 Flash, GPT 4.1 Mini, and Claude Haiku 4.5 show that rewordings reduce graph similarity relative to repeated-run baseline across all 27 model-dataset-transformation settings.

Critical strengths

S1. The paper evaluates what the agent does under intent-preserving rewordings (which permissions it requests, what information it discloses and to which channels, and which tools it calls), rather than only whether the task succeeds. To my knowledge, defining metamorphic relations over these behavioral consequences is novel.

S2. The experiments cover a good breadth of benchmarks, transformation types, and interaction settings (single-turn, multi-turn clarification, and coding agents run through Claude Code and Codex CLI). I particularly appreciated the analysis comparing operation-graph and output similarity, which shows that agents can produce similar final outputs while behaving differently along the way.

Critical weaknesses

W1. This paper offers limited new insights into a well-documented phenomenon of “sensitivity of LLM behavior to prompt perturbations”. The paper’s measurement contributions in the agentic setting are novel, but its main findings largely confirm what existing literature would predict: rewording changes behavior beyond run-to-run noise. Beyond aggregate breakdowns by transformation type, the paper does not offer new takeaways such as explaining why particular rewordings cause particular behavioral changes, or which tasks and actions are most at risk (which practitioners may care about).

W2. The main experiments use smaller, cheaper and/or older models (Gemini 2.5 Flash, GPT-4.1 Mini, Claude Haiku 4.5) with limited comparison to newer and more capable models. Only one flagship frontier model (GPT-6 Astra) is used on a subset of AgentCIBench, where it shows high operational consistency. From this, the authors conclude that “[s]tronger model capability thus substantially reduces sensitivity…” when this is too low of a sample size to conclude with the stated confidence. Having more frontier models across different model families (e.g. newer Gemini Pro or Claude Sonnet/Opus models) would strengthen their conclusion. In absence of that, I would suggest the authors weaken this claim.

W1 is the main driver of my rating, with W2 as an additional concern. The paper’s additional findings (e.g. the effects of clarification) are, in my view, incremental confirmations of known prompt sensitivity rather than substantial new insight.

Questions for authors

Q1. How confident are the authors that the rewordings mean the same thing? For example, something like “Are you able to {task}?” could reasonably be answered by “yes” without doing anything, and “This time, {task}” could imply a previous failure to do {task}. The paper says two authors checked the prompts, but gives no details on what the criteria for their verification were.

Q2. How often did agents ask a question or refuse to act under each template, and how is that counted?

10 comments
AuthorsOpenReview ↗

Response to Reviewer bmxW (Part 1)

We thank the reviewer for the detailed comments and for recognizing the novelty of evaluating agents' operational behaviors. We are confident to address the concerns and answer the questions raised by the reviewer.

W1. Novel Insights Beyond General Prompt Sensitivity.

We thank the reviewer for recognizing the novelty of our measurement contributions and for encouraging more actionable insights. We agree that the general observation that prompt rewording changes LLM behavior is already well established. However, our contribution goes beyond demonstrating prompt sensitivity by investigating operational consistency as a distinct reliability property of AI agents.

Unlike conventional evaluations focusing on model responses or task success, AgentMT systematically examines whether intent-equivalent requests lead to consistent permission decisions, information disclosure, and execution trajectories. Our MT-based formulation enables such comparisons without requiring a unique correct execution trajectory.

Beyond this methodological contribution, our study provides several empirical insights that are not directly implied by existing prompt-sensitivity findings.

  • Operational consistency differs from output consistency. Our results show only weak-to-moderate correlations between operation-graph and output similarity (Spearman's ρ = 0.20–0.55). Agents may produce highly similar final responses while exhibiting substantially different operational behaviors. This suggests that final-output evaluation alone can overlook important aspects of agent reliability.
  • Operational inconsistencies can involve concrete task-level violations. Our additional benchmark-grounded analysis shows that 37.6% of consequence-inconsistent pairs on AIP involve unnecessary permission requests. On AgentCIBench, language transformations introduce missing-required-information violations in 15.7% of previously compliant cases and unwanted disclosures in 6.5%. These findings demonstrate that operational inconsistency is not merely a structural difference but can be associated with violations of task requirements.

Which tasks and actions are most at risk?

To provide more actionable insights, we conducted additional analyses using the benchmark-defined violation indicators and task characteristics.

First, we identified which operational violations are more likely to be introduced by different prompt transformations. The results are shown as follows.

BenchmarkIssues in inconsistent pairsViolation typeToneLanguageFormulation
AIP37.6%Unnecessary permissions3.3%2.0%2.3%
AgentCIBench100.0%Missing required information11.1%15.7%7.3%
Unwanted disclosures2.3%6.5%4.3%
TRAJECT-Bench67.0%Benchmark-defined operational issues———

On AIP, tone transformations show the highest rate of newly introduced unnecessary permissions (3.3%). On AgentCIBench, language transformations are associated with the highest rates of missing required information (15.7%) and unwanted disclosures (6.5%). These results reveal that different linguistic variations are associated with different types of task-level violations.

Second, we examined which task characteristics are associated with greater operational inconsistency. On AIP, tasks involving four or more data fields exhibit a mean graph-similarity reduction of 0.153, compared with 0.071 for tasks involving one to three fields. On AgentCIBench, model–task combinations with frequent non-action exhibit a substantially larger reduction (0.229) than those in which agents consistently act (0.061). In contrast, TRAJECT-Bench shows similar reductions across different target-tool counts (0.090–0.093), suggesting that tool count alone is not a reliable indicator of inconsistency. These findings provide more specific insights into which operational requirements and task characteristics deserve closer attention, beyond the general observation that rewording affects agent behavior.

Which rewordings are associated with behavioral changes?

We additionally analyzed individual tone templates and their effects on agents' decisions to act, seek clarification, or refuse. For example, "Are you able to ...?" increases the clarification rate from 6.4% to 16.0%, while reducing the action rate from 84.3% to 71.0%. More challenging expressions, such as "Why can't you ...?", are associated with substantially higher refusal rates (36.5%, compared with 8.0% for original requests). These results identify specific linguistic constructions associated with changes in agents' execution decisions, beyond aggregate differences across transformation families.

Our contribution is not simply that rewording changes agent behavior, but that operational inconsistency is distinct from output inconsistency, can involve concrete task-level violations, and exhibits identifiable patterns.

AuthorsOpenReview ↗

Response to Reviewer bmxW (Part 2)

W2. Evaluation Across Frontier Models.

We agree that our original conclusion that stronger model capability substantially reduces sensitivity was too broad given the limited frontier-model evaluation. To address this concern, we conducted additional experiments with two frontier models from different model families, Claude Opus 5.5 and Gemini 4 Carbon, alongside the previously evaluated GPT-6 Astra. We used the same 15 AgentCIBench tasks, prompt transformations, repeated-run baselines, and evaluation metrics to ensure a consistent comparison.

ModelS–SToneLanguageFormulation
GPT-6 Astra0.9650.9420.9640.975
Claude Opus 5.50.9650.9650.9390.928
Gemini 4 Carbon0.9190.8250.8280.889

The results reveal distinct operational consistency patterns across frontier models. GPT-6 Astra maintains high graph similarity (0.942–0.975) across all transformations. Claude Opus 5.5 also exhibits relatively high consistency, although its graph similarity decreases under language (0.939) and formulation (0.928) variations. In contrast, Gemini 4 Carbon exhibits larger reductions under tone and language variations, from its repeated-run baseline of 0.919 to 0.825 and 0.828, respectively.

Interestingly, all three models maintain high output similarity (0.967–1.000) despite their differences in operational consistency. This further supports our finding that similar final outputs do not necessarily imply consistent operational behaviors.

Together, these additional results suggest that operational consistency is model-dependent and does not necessarily improve uniformly with stronger model capability. We revise RQ3 to emphasize the observed differences across frontier models rather than attributing higher operational consistency solely to stronger model capability. This analysis further broadens our analysis across three frontier-model families and provides a more nuanced understanding of operational consistency.

AuthorsOpenReview ↗

Response to Reviewer bmxW (Part 3)

Q1. Validation of Intent-Preserving Transformations.

We thank the reviewer for highlighting the potential ambiguity in expressions such as "Are you able to {TASK}?" and "This time, {TASK}?" We agree that these expressions may admit different interpretations in isolation. However, their interpretation should also consider the context of AI agents, which are designed to understand and execute user requests rather than merely answer questions like LLMs.

For "Are you able to {TASK}?", although it can literally be interpreted as a capability question, it is commonly used as an indirect request for action. To examine how agents actually interpret such expressions, we analyzed all 15 tone templates across 116 tasks, three models, and three repeated executions, yielding 1,044 executions per template. Our observations show that agents generally interpret these expressions as actionable requests rather than simply answering whether they possess the required capability. For example, 71.0% of executions under "Are you able to ...?" proceeded with consequential actions, compared with 82.3% for "Can you ...?" and 85.7% for "Could you ...?". Our inspection did not identify cases where agents merely answered "yes" without attempting the requested task.

For "This time, {TASK}", we acknowledge that this expression can imply prior interaction, but it does not necessarily indicate that the task previously failed or introduce additional operational requirements. Such expressions reflect natural human communication patterns, where contextual or emphatic wording may be used to convey requests. Our tone templates are informed by naturally occurring user expressions from WildChat and prior linguistic studies, rather than being constructed specifically to represent requests following failed executions.

Transformation design. As described in Section 3.2 and Appendix A.1, we employ several measures to preserve the original user intent:

  • Verbatim task preservation. Each tone template contains a fixed {TASK} slot into which the original request is inserted without modifying its content.
  • Preservation of operational requirements. The transformations do not modify requested actions, entities, recipients, task constraints, or authorization scope.
  • Controlled linguistic variations. The transformations modify only predefined linguistic expressions surrounding the original request, rather than introducing new conditions.

Human validation. To further assess intent preservation, we conducted a blinded human audit of 100 tone variants, independently evaluated by two annotators with relevant domain expertise. Each variant was assessed against four criteria: (1) preservation of task intent, (2) preservation of recipients and objects, (3) preservation of authorization scope, and (4) whether the transformed prompt remained an execution request. Deviations were additionally classified as none, minor, or major.

Both annotators judged all four criteria to be preserved in 93/100 cases, with criterion-level agreement of 98%, 100%, 100%, and 96%, respectively. Neither annotator identified any major deviations. More specifically, among 38 samples involving context-sensitive expressions, both annotators confirmed task-intent and authorization-scope preservation in all cases.

We also conducted a separate human validation for 210 language variants, covering 30 tasks across each of seven languages. The same two annotators independently evaluated task intent, recipients/objects, authorization scope, and tone. Both annotators were fluent in English, and each could read at least one additional language included in our evaluation. To support multilingual assessment, we employed Claude for source-blind literal back-translation into English, supplemented by additional translation tools when necessary. The annotators agreed on the joint preservation judgment in 204/210 cases (97.14%), with criterion-level agreement ranging from 98.10% to 100%.

These results provide empirical support for our intent-preserving transformation design.

Reviewer_hua4OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper works on an interesting problem about whether an LLM agent will behave differently if two intent-equivalent input prompts are presented with different tones, languages, and formulations. Given a prompt, the study rewrites different variants of the prompt with different setups while preserving its intent. Then it measures whether the agent has a different execution graph or has different operation consequences. The empirical results show that prompt variations can lead to different execution graphs and different consequences. The paper further demonstrates that even frontier models may slightly be affected by the setup changes. Interestingly, the paper shows that the coding task are less affected by different prompts compared to general agents on daily tasks.

Critical strengths

The paper studies an interesting problem. While most findings are expected, it is interesting to see how coding agents behave differently compared to general agents.

Critical weaknesses

I have minor concerns about the discussion with related work and the evaluation results interpretation:

  1. There is existing work discuss that agent may fail intent-preserving rephrases[1]. It is important to clarify the paper's unique focus compared to existing work.
  2. How to qualify the effects of tone/style from the complexity of the request? For example, a request might be more complex and the intent more obfuscated after paraphrase. How to separate the effects of complexity and the effects of tone/style?
  3. How to ensure the paraphrased requests are still realistic? Existing findings on user simulation highlight the challenge of realistic simulation[2,3].

[1] S. Feng, et.al, TAI3: Testing Agent Integrity in Interpreting User Intent, NeurIPS 2025

[2] S. Mehri, et.al, Measuring and Mitigating the Distributional Gap Between Real and Simulated User Behaviors, arXiv May 26

[3] X. Zhou, et.al, Mind the Sim2Real Gap in User Simulation for Agentic Tasks, COLM 2026

Questions for authors

Please see the above discussion.

5 comments
AuthorsOpenReview ↗

Response to Reviewer hua4 (Part 1)

We thank the reviewer for admitting our effort and providing the constructive and insightful comments. We are confident to address all the concerns and raised questions. We summarize and answer them one by one.

W1. Distinction from Prior Work.

We thank the reviewer for pointing out TAI3: Testing Agent Integrity in Interpreting User Intent (Feng et al., NeurIPS'25). Although both studies consider intent-preserving prompt variations, they investigate different aspects of AI agent reliability, with distinct research objectives, problem formulations, and evaluation methodologies.

  • Research objective. TAI3 focuses on intent integrity, aiming to identify cases where agents misinterpret user instructions and consequently perform incorrect API operations. In contrast, our work focuses on operational consistency as a reliability property of AI agents. We investigate whether agents exhibit consistent operational behaviors when the same user intent is expressed differently. Importantly, operational differences do not necessarily indicate failures; even individually valid executions may differ in permission decisions, information disclosure, or tool usage. Our study systematically characterizes such inconsistencies rather than focusing solely on identifying incorrect executions.

  • Problem formulation and testing oracle. TAI3 adopts an API-centric testing approach, using API specifications, semantic partitioning, and ground-truth user intent to generate tests and identify intent interpretation errors. In contrast, we formulate operational consistency evaluation as an MT problem, where intent-preserving transformations establish relations between executions. By comparing execution graphs, outputs, and consequences, AgentMT can systematically quantify operational inconsistencies without requiring a unique ground-truth execution trajectory, which is often unavailable for complex agent tasks.

  • Evaluation scope and methodology. TAI3 primarily investigates API-level intent integrity and optimizes test generation to efficiently expose errors. Our work systematically evaluates three dimensions of natural user expressions (tone, language, and formulation) across multiple agent benchmarks and models.

These perspectives are complementary: an agent may correctly interpret a user's intent yet exhibit different operational behaviors across equivalent requests. Our work provides a systematic framework for measuring this distinct reliability property.

W2. Distinguishing Tone/Style Effects from Request Complexity.

Our transformations are designed to vary how users express their requests while preserving the underlying task complexity and operational requirements. As described in Section 3.2 and Appendices A.1–A.3, we apply different constraints to each transformation type to minimize potential confounding factors.

Tone transformations. We use predefined linguistic templates with a fixed {TASK} slot, where the original request is inserted verbatim without modifying its content. Each transformation only adds a short, predefined linguistic expression (e.g., a prefix or suffix) to change the tone of the request. This ensures that the number of requested subtasks, operational requirements, entities, and constraints remain identical, while limiting changes in linguistic length and structure to the predefined templates. By doing so, we can minimize the potential confounding effects of request complexity.

Formulation transformations. We adopt two strategies.

  • Sentence reordering changes only the order of complete sentences, preserving the original content and length. Thus, there should be little complexity changes.
  • Paraphrasing generates multiple candidate paraphrases and prioritizes candidates satisfying our preservation constraints. Specifically, we check (1) preservation of protected entities, identifiers, numbers, dates, and other task-specific literals; (2) non-identical wording; (3) preservation of interrogative form; (4) unchanged negation and scope markers; and (5) a word-count ratio within [0.70, 1.35]. The last point controls the length of the variants, making sure the complexity only changes within a pre-defined scope. These constraints are designed to preserve task-level complexity and operational requirements while limiting changes in linguistic complexity.

These constraints are designed to preserve task-level complexity and operational requirements while limiting changes in linguistic complexity.

AuthorsOpenReview ↗

Response to Reviewer hua4 (Part 2)

W3. Realism of Transformed Requests.

We agree that LLM-based user simulation may not fully capture the diversity of real human behavior, as highlighted by the cited studies. However, our work differs from conventional user simulation: rather than generating synthetic user behaviors or complete conversations, we apply controlled linguistic transformations to existing user requests. Our experimental design is grounded in realistic task scenarios and naturally occurring user expressions.

  • Realistic task scenarios. Our source requests come from three established agent benchmarks, covering practical scenarios such as messaging, scheduling, permission management, and multi-step tool use. We preserve these original tasks rather than generating new synthetic task scenarios.
  • Real-world linguistic expressions. Our transformation design is informed by real-world user expressions. For tone transformations, we extract frequently occurring request patterns from WildChat, a large-scale dataset of real user–LLM interactions, and construct representative polite, neutral, and impolite templates. Similarly, our selection of seven non-English languages is based on their observed usage frequencies in WildChat.
  • Constrained paraphrasing. Rather than relying on unrestricted LLM generation, we apply controlled paraphrasing to vary lexical choices and sentence structures while preserving task-critical entities, requested actions, recipients, and operational constraints. As described in Section 3.2 and Appendix A.3, the generator is instructed to produce natural English requests, and candidate paraphrases are checked for preservation of protected content and linguistic constraints.

These design choices help ensure that our transformations reflect plausible variations in how users express the same task, rather than arbitrary synthetic requests. While we do not claim that the resulting variants reproduce the full distribution of real-world user behavior, our objective is to systematically evaluate operational consistency under realistic, intent-preserving linguistic variations. We will clarify this distinction and the corresponding limitations in the revised manuscript.

I think I am not fully convinced by the response.

Q1, "testing whether an agent can correctly understand human intent when the language changes" (TAI3) and "testing whether an agent behaves consistently when the request changes" (this paper) seem to be the same research question.

The clarification in Q2 addresses my question.

Q3, I think it is still important to ensure the transformation reflects realistic user distribution & real-user's interaction patterns with the agent. A request can be represented in almost arbitrarily convoluted style, but it is only useful if the request represents challenges in realistic use patterns.

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

The paper introduces AgentMT, a framework for checking whether agents behave consistently when the same request is phrased differently. It changes the tone, language, and wording, then compares actions and final responses. The main experiments use three models on tasks involving permissions, information sharing, and tool use. They find that rephrasing adds behavioral differences beyond ordinary variation between repeated runs. The paper also examines a stronger model, clarification, and coding agents.

Critical strengths

Repeating the original request gives a useful baseline for how much agents vary between runs anyway. Rephrasing adds further differences across models and datasets, even with simple polite requests.

The paper addresses a timely question about agent reliability: whether phrasing the same request differently changes what an agent does. Looking at permissions, information sharing, and tool use makes the study relevant to problems that evaluating final responses alone may miss.

Critical weaknesses

Different behavior does not always mean something went wrong. Figure 5 shows a failed call followed by recovery, while Table 7 shows an unwanted disclosure. We do not learn how often the differences are actual failures. The released code already checks permissions and disclosures against benchmark labels, so reporting those rates would help explain the practical importance of the results.

The variant review and disclosure measurement need clarification. The automated review flags 227 of 1,740 tone variants as non-equivalent, but the analysis uses all executed variants. It is unclear how human review resolved those flags. Also, the released Table 2 script does not load the disclosure judgments and instead uses literal matching, which can miss translations. Did this affect the published results or only the released code? If the review decisions are explained and the checked results preserve the main effects, I would raise my score.

Questions for authors

Who reviewed the flagged variants, which languages could they read, and how did they resolve failed checks? Do the main findings hold using only eligible variants?

Comparing disclosed items using the released judgments gives repeated-run (S-S) rates of 0.389/0.130/0.074, versus 0.33/0.11/0.06 in Table 2, for Gemini/GPT-4.1 Mini/Haiku. With the same executions and filtering, adding channels cannot lower these rates. What matching and filtering produced the table? Please reconcile the affected results and correct them if needed.

How often do the additional differences involve unnecessary permissions, omitted required information, or unwanted disclosures, using the indicators already computed by the code?

6 comments
AuthorsOpenReview ↗

Response to Reviewer pMAZ (Part 1)

We thank the reviewer for the constructive comments and careful examination of our paper and implementation. We provide detailed clarifications, additional experimental evidence, and corrections to address the concerns raised in weaknesses and questions.

W1, Q3. Quantification of the Actual Failures Caused by Inconsistencies.

We fully agree with the reviewer that different operational behaviors do not necessarily indicate failures. This is also why our study focuses on operational consistency rather than treating every behavioral difference as a violation. In Section 4.1, we introduce consequence inconsistency as one of our three evaluation metrics (detailed in appendix A.5, e.g., Repeated requests with the same field is not a violation).

During our initial experiments, we evaluated the rates of three types of benchmark-defined violations: unnecessary permissions, missing required information, and unwanted disclosures, under inconsistencies. We did not originally report these results because our primary objective is to characterize operational consistency rather than directly measure task-level violations, as operational consistency is itself a meaningful reliability property of AI agents.

We agree that these indicators mentioned by the reviewer can further demonstrate the practical significance of the observed inconsistencies. We therefore report the corresponding results below.

BenchmarkIssues in inconsistent pairsViolation typeToneLanguageFormulation
AIP37.6%Unnecessary permissions3.3%2.0%2.3%
AgentCIBench100.0%Missing required information11.1%15.7%7.3%
Unwanted disclosures2.3%6.5%4.3%
TRAJECT-Bench67.0%Benchmark-defined operational issues———

The second column reports the proportion of consequence inconsistent source variant pairs involving benchmark defined violations or operational issues. The last three columns report newly introduced violation rates, macro-averaged across three models, among tasks with no corresponding violation across repeated source executions. TRAJECT-Bench uses operational indicators rather than direct security-policy violations.

The results show that operational inconsistencies are frequently associated with concrete task-level issues. On AIP, 37.6% of consequence-inconsistent pairs involve unnecessary permission requests, while 67.0% of inconsistent pairs on TRAJECT-Bench exhibit benchmark-defined operational issues. For AgentCIBench, the 100.0% rate largely reflects the overlap between the consequence representation and the benchmark's disclosure requirements. More importantly, prompt variations can introduce new violations in previously compliant executions. For example, language transformations introduce missing-required-information violations in 15.7% of source-clean cases and unwanted disclosures in 6.5% of source-clean cases on AgentCIBench.

Overall, these results further establish the practical relevance of AgentMT. Operational inconsistency does not necessarily indicate failure, but it can expose changes associated with concrete violations of task requirements. This highlights the importance of systematically evaluating agent behavior beyond final-output correctness.

AuthorsOpenReview ↗

Response to Reviewer pMAZ (Part 2)

W2, Q1 Clarification of Flagged Variants and Human Validation.

We clarify below how the flagged variants were reviewed, how their validity was assessed, and whether excluding them affects our main findings. We also report additional human validation results regarding other variants.

Original review of flagged tone variants. The 227 potentially non-equivalent variants among 1,740 tone variants were identified by an automated assessment using Gemini 2.5 Flash. Before conducting the main experiments, one author with relevant expertise manually inspected all 227 flagged variants. This inspection found that the flags arose from pragmatic ambiguity in natural human communication, rather than substantive changes to the underlying task intent. For example, expressions such as "This time, [TASK]" and "Why can't you [TASK]?" were flagged as context-sensitive, although they are commonly used to convey requests in real interactions. Since our study investigates how AI agents behave under diverse natural user expressions, we retained these variants to reflect realistic communication patterns.

Futher Validation. To further validate this decision, we conducted an additional blinded human audit of 100 tone variants. Two authors with relevant domain expertise, neither of whom participated in the initial inspection, independently reviewed 50 randomly sampled flagged variants and 50 randomly sampled unflagged variants. The samples were shuffled, and their automated flag status was hidden from both annotators.

The annotators assessed four criteria: (1) preservation of task intent, (2) preservation of recipients/objects, (3) preservation of authorization scope, and (4) whether the variant remained an execution request. They also assessed the severity of any deviation as none, minor, or major.

Both annotators judged all four criteria to be preserved in 47/50 flagged variants and 46/50 unflagged variants. The inter-annotator agreement on the four criteria was 98%, 100%, 100%, and 96%, respectively. Neither annotator identified any major deviations. These results further support our initial observation that the automated flags primarily reflect pragmatic ambiguity rather than substantive changes in user intent.

Sensitivity analysis excluding flagged variants. We conducted a conservative sensitivity analysis by excluding all 227 Gemini-flagged tone variants and recomputing the operation-graph similarity results. The table below reports the repeated-run baseline (S–S) and the reductions in graph similarity (Δ) under the three transformation families, with 95% confidence intervals.

DatasetModelS–STone ΔLanguage ΔFormulation Δ
AIPGemini0.9080.124 [0.077, 0.173]0.105 [0.058, 0.158]0.084 [0.029, 0.140]
GPT-4.1 Mini0.9550.090 [0.049, 0.137]0.102 [0.054, 0.156]0.106 [0.042, 0.178]
Haiku 4.50.9570.166 [0.095, 0.249]0.180 [0.111, 0.261]0.157 [0.058, 0.268]
ACBGemini0.7750.102 [0.027, 0.184]0.162 [0.083, 0.250]0.076 [-0.010, 0.180]
GPT-4.1 Mini0.8730.086 [0.054, 0.121]0.182 [0.127, 0.239]0.118 [0.072, 0.168]
Haiku 4.50.9720.172 [0.110, 0.241]0.280 [0.200, 0.365]0.149 [0.074, 0.231]
TJBGemini0.4820.066 [0.020, 0.114]0.076 [0.028, 0.128]0.048 [-0.017, 0.118]
GPT-4.1 Mini0.5560.078 [0.014, 0.153]0.141 [0.053, 0.237]0.047 [-0.040, 0.143]
Haiku 4.50.6160.182 [0.125, 0.243]0.224 [0.157, 0.295]0.162 [0.085, 0.241]

The results remain highly consistent with our original findings. After excluding all 227 flagged tone variants, tone transformations still produce reductions in graph similarity across all nine dataset–model combinations, with all corresponding confidence intervals excluding zero. The results for language and formulation transformations also preserve the previously observed patterns.

Human validation of language transformations. Beyond tone variants, we also conducted a separate human validation of 210 translated prompts, covering 30 tasks in each of seven languages. The same two annotators independently evaluated the four preservation criteria described above. Both annotators were fluent in English and each could read at least one additional language included in our evaluation.

To support assessment across all seven languages, we additionally used Claude for source-blind literal back-translation into English. Claude received only the translated prompts, without access to the original English requests. When encountering ambiguous expressions, the annotators also used translation tools to cross-check their interpretations and verify the intended meanings. The two annotators agreed on the joint preservation judgment for 204/210 cases (97.14%).

AuthorsOpenReview ↗

Response to Reviewer pMAZ (Part 3)

W2, Q2 Correction of Disclosure Inconsistency Measurements.

We thank the reviewer for identifying this issue. We carefully reviewed our implementation and identified an error in scripts/04_analysis/analyze_rq2_consistency.py.

Specifically, the script was intended to load the precomputed GPT-4o disclosure judgments through load_judge() before constructing the consequence representations. However, this function was accidentally omitted.

We have recomputed the affected results using the semantic disclosure judgments and corrected both the implementation and the corresponding manuscript results. The corrected repeated-run (S–S) inconsistency rates for AgentCIBench are 0.389, 0.130, and 0.074 for Gemini 2.5 Flash, GPT-4.1 Mini, and Claude Haiku 4.5, respectively, matching the reviewer's independent calculations. We provide the affected results.

MetricModelS–STone ΔLanguage ΔFormulation Δ
Shared informationGemini0.39−0.01 ± 0.03+0.04 ± 0.04−0.07 ± 0.05
GPT-4.1 Mini0.13+0.11 ± 0.04+0.22 ± 0.05+0.11 ± 0.06
Haiku 4.50.07+0.16 ± 0.04+0.27 ± 0.05+0.19 ± 0.07

Although the correction changes some numerical estimates, it does not change our central finding that intent-preserving prompt variations can induce disclosure inconsistencies beyond normal repeated-run variation. In particular, GPT-4.1 Mini and Claude Haiku 4.5 continue to exhibit substantial increases in disclosure inconsistency, with language transformations producing the largest increases (+0.22 and +0.27, respectively). The corrected results also reveal that the effects are model-dependent, with smaller and less consistent changes for Gemini.

Overall, the corrected results continue to support the practical relevance of evaluating operational consistency under diverse user expressions. We have updated the corresponding results and discussion in the manuscript to accurately reflect these findings. We appreciate the reviewer's careful examination and reproduction effort, which helped us identify and resolve this issue.

Comments on the paper

2 comments
Area_Chair_1fAkOpenReview ↗

Please review the authors' follow-up response

Dear Reviewers hua4 and bmxW, thank you for engaging in the discussion. The authors posted further responses on Nov 17 addressing your remaining concerns. Please take a quick look before the discussion closes and let us know whether your assessment or score has changed. Thank you! AC.

AuthorsOpenReview ↗

Revision of Submission #53743 Submitted

Dear AC and Reviewers,

Thank you all for your constructive comments and valuable feedback!

We have uploaded the revised version of our paper. All changes are marked in blue.

We have addressed all comments, questions, and follow-up discussions with the reviewers, as well as revised the paper accordingly. Our revised version also reports the key new experimental results in the main text and the appendix.

Thank you again for your time and feedback!

Best regards,

Authors of Submission #53743

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.