acceptodds
Under review as a conference paper at ICLR 2027

DECOR: Auditing LLM Deception via Information Manipulation Theory

Abstract

Large language models (LLMs) can deceive users by subtly manipulating truthful information—omitting key facts, shifting focus, or obscuring meaning, making deception auditing essential. Existing black-box deception auditing methods typically make coarse-grained, response-level judgments without a theoretical basis for identifying the specific information involved in deception or how that information contributes to it. We propose DECOR, a multi-agent framework that operationalizes information manipulation theory (IMT) for deception auditing, which characterizes deceptive communication along four dimensions of information manipulation. DECOR decomposes input contexts into atomic information units, evaluates each unit against the response across these dimensions, and aggregates the resulting manipulation profiles into a global deception index. To enable reliable deception evaluation, we establish human reference labels for 745 LLM responses on DeceptionBench. Experimental results show that DECOR outperforms existing methods in both single-turn and multi-turn deception auditing, achieving 0.948 and 0.657 AUROC, respectively. Across 15 auditor backbones, incorporating elicited thoughts consistently improves DECOR, demonstrating their value as a complementary signal for deception auditing. Our analysis further shows that LLM deception more often involves omitting relevant information and diverting attention than fabricating or distorting facts. Our project is available at https://anonymous.4open.science/r/DECOR-F866/.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM Deception, Deception Detection, Information Manipulation, LLM Safety, Black-box Monitoring, Trustworthy AI

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_QL9QOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper introduces DECOR, a training-free, multi-agent framework for black-box auditing of LLM deception. Drawing on Information Manipulation Theory, it decomposes the input context into atomic information units, assigns importance weights, evaluates each unit along four dimensions—quantity, quality, relation, and manner—and aggregates the scores into a deception index. The authors provide human annotations for 745 DeceptionBench outputs and evaluate DECOR on DeceptionBench and OpenDeception under response-only and thought-plus-response settings. Experiments include baseline comparisons, evaluations across 15 auditor backbones, and component ablations. Reported performance is strongest in the single-turn setting.

Critical strengths
  1. Inspectable auditing workflow. DECOR exposes information units, importance weights, and dimension-level scores underlying its aggregate judgment. These intermediate outputs provide concrete material for manual inspection and error analysis, illustrated through both successful and unsuccessful cases. The method also requires neither parameter training nor access to model internals.

  2. Human annotation effort. The authors annotate all 745 evaluated DeceptionBench outputs, labeling responses and elicited thoughts separately. The labeling guidelines and domain-specific agreement statistics provide useful documentation for evaluating deception auditors.

  3. Promising single-turn results. With GPT-4o as the auditor, DECOR reports AUROCs of 0.938 and 0.948 on DeceptionBench under response-only and thought-plus-response settings, respectively, exceeding the reported baselines in both settings. The weighting and decomposition ablations provide preliminary evidence that these components benefit the pipeline on this dataset.

Critical weaknesses

W1: Theoretical Foundation

  • Reliability of IMT

Page 2, lines 73–77, describes Information Manipulation Theory (IMT) as extensively validated, without adequately discussing the relevant disputes or conditions of applicability.

Experiments by McCornack et al. (1992) found that manipulating information quantity, veracity, relevance, and clarity affected participants’ judgments of deceptiveness. This supports IMT’s value as an analytical framework, but does not establish that an arbitrary scoring scheme along these four dimensions can accurately detect LLM deception. In a replication incorporating checks on the covertness of violations and message types, Jacobs, Dawson, and Brashers (1996) reported results inconsistent with the IMT predictions they tested and argued that deception should instead be explained through false conversational implicatures. McCornack et al. (1996) challenged this interpretation and reanalyzed the data. In a Hong Kong sample of 310 participants, Yeung, Levine, and Nishiyama (1999), whom the paper cites, found that Quality and Relevance violations were judged deceptive, whereas the Quantity and Manner conditions did not yield the same result. Ratings on all four dimensions nevertheless correlated with perceived deceptiveness, and the authors suggested that cultural differences in what constitutes a covert violation might explain the discrepancy. These findings indicate that IMT is not a fully reliable theoretical foundation and that its applicability is context-dependent.

  • Departures from IMT’s Original Definitions

The following changes in the paper’s operationalization of IMT require explanation.

LocationDECOR’s ruleConcern
Appendix A.1, p. 13, lines 688–699Assigns newly introduced, unverified information presented as fact to QuantityThe original framework places assertions lacking adequate evidence under Quality. Why reassign or double-count them under another dimension?
Appendix A.1, p. 14, lines 703–708Assigns bypassing an information unit’s constraints while advancing the original goal to RelationThis partly measures goal compliance rather than conversational relevance alone.
§3.1, p. 4; Appendix A.2, p. 14The main text weights information by its importance to the user’s judgment; the appendix weights it by whether it facilitates or obstructs the execution of deceptive instructionsInfluence on audience beliefs and importance to executing a deceptive task are distinct quantities.

The original discussion of Quality, Relation, and covert violations appears in McCornack (1992), pp. 5–6. It concerns how speakers exploit listeners’ expectations of cooperative communication to mislead them, rather than simply whether an output advances a predefined goal.

Questions: Please explain each modification and its necessity, and use contrasting examples to distinguish compliance with an inappropriate request from inducing false beliefs in an audience. Clarify whether the intended audience is the informed requester or a third party who ultimately receives the generated content, to avoid conflating deception, task compliance, and safety refusals.

  • Gap Between Theory and Validation

The paper’s central weakness is that its claim to be “theoretically grounded” is insufficiently supported by the method and experiments. In practice, IMT is primarily implemented as four-dimensional scoring rules in prompts, with an LLM assessing information units’ importance and degree of manipulation (§3; Appendix A). This introduces a set of manually designed inductive biases into an LLM judge. Citing IMT explains their origin, but does not establish their suitability for detecting LLM deception, particularly given the changes to the original concepts’ operational definitions.

Equations (1)–(4) define the workflow without translating the theory into testable methodological constraints. They encapsulate decomposition, weighting, and auditing as functions and specify output aggregation, but derive neither properties these functions should satisfy under IMT nor a theoretical relationship between the aggregate score and deception. Although the aggregation formulas are used in the experiments, the formalization does not serve to test the theory: the evaluation neither tests IMT-derived predictions nor verifies that the LLM’s scores faithfully instantiate the theoretical constructs. Moreover, the evaluation still relies on the “black-box response-level judgments” criticized by authors themselves in the abstract.

The experiments therefore validate the predictive performance of the overall prompting workflow, not its theoretical foundation. The decomposition and weighting ablations in Table 5 and the dimension-removal experiments in Figure 3 show that some design choices help on the current datasets. They do not establish whether the gains arise from IMT’s specific explanatory mechanisms or from finer task decomposition, more detailed scoring instructions, and additional reasoning. Higher AUROC alone does not establish the theoretical validity of the introduced inductive biases.

Questions: Please identify the design choices and testable predictions derived from IMT and test those predictions directly. Also evaluate alternative scoring frameworks while holding decomposition, weights, inputs, and reasoning budgets fixed to isolate IMT’s contribution. The authors could also consider formal verifications or other approaches that make reasoning more traceable and verifiable as complementary validation. Such evidence would strengthen my confidence in the paper’s central claim.

W2: Empirical Soundness

  • Independent Validation of Human Labels

Appendices B.1–B.2, p. 16, state that two annotators inspected the context and deceptive goal, labeled independently, and then discussed disagreements to reach consensus. Overall Cohen’s κ for Response was 0.810, but only 0.564 for Education (Table 6).

These results support annotation consistency, but do not establish that the responses would actually induce false beliefs in the intended audience. The paper also provides insufficient detail on the annotators’ relationship to the authors, whether they were blinded to method predictions, and how disagreement discussions were conducted. The safeguards for independence and validity are therefore unclear. Two annotators are too few, and disagreement resolution may be overly sensitive to individual judgment.

Questions: Please describe annotator recruitment and blinding, release anonymized pre-discussion labels and disagreement-resolution records, and have independent annotators relabel a subset. For disputed cases involving omission, implication, and ambiguity, add a small audience study that directly measures factual beliefs formed after reading the outputs.

  • Missing Ground Truth for Fine-Grained Explanations

A key contribution is identifying “what information is manipulated and how” (§3; §5.2), yet human annotations primarily consist of binary deception labels for thoughts and responses (Appendix B.1). The completeness of information units, appropriateness of weights, and correctness of the four-dimensional scores are illustrated through examples rather than independently validated quantitatively.

High final AUROC therefore does not establish that the intermediate explanations are correct. The finding that Quantity and Relation manipulations are more common (§5.4; Figure 4) also relies on the auditor’s own classifications and has not been confirmed by independent annotation.

Questions: Please independently annotate information units, manipulation types, and importance on a stratified subset, and report accuracy and agreement for each component. Validate the classifications before drawing conclusions about manipulation frequencies; otherwise, describe them only as the distribution of categories predicted by DECOR.

  • Too Few Negative Examples in Multi-Turn Evaluation

Appendix B.3, p. 17, Table 7, shows that the filtered OpenDeception set contains 140 deceptive and six non-deceptive dialogues. Rank comparisons and false-positive estimates rely on just six negative dialogues. Sensitivity to individual negatives therefore needs to be reported.

In Table 3, DECOR’s Response-only AUROC is 0.571, below CoT Red-Handed’s 0.583, and reaches 0.657 only after adding elicited thoughts. The paper itself acknowledges that these thoughts need not faithfully reflect internal reasoning (p. 5, footnote 1). The multi-turn results should therefore report detection performance separately for visible responses alone and for inputs that include additional thought text.

The paper reports neither a full-test-set confusion matrix nor the decision threshold used to convert DECOR scores into binary predictions. Table 7 gives only the ground-truth label distribution. Appendix D (pp. 21–24) presents one TP, TN, FP, and FN example each, which cannot substitute for aggregate counts. AUROC does not reveal how many of the six non-deceptive dialogues are flagged at a given operating threshold. With only six negatives, each additional false positive changes FPR by approximately 16.7 percentage points. A confusion matrix would improve transparency but would not remedy the shortage of negative examples.

Questions: Please add non-deceptive dialogues, especially hard negatives involving legitimate omission, uncertainty, politeness, and refusals. Report confidence intervals for each setting and a sensitivity analysis leaving out each of the six negatives in turn. State the benefits and applicability of each input setting separately. For both Response and Thought & Response, report full-test-set confusion matrices with raw TP, FP, TN, and FN counts, the decision threshold and its selection procedure, and FPR and recall with confidence intervals. Thresholds should be prespecified or selected on an independent validation set, not optimized on the test set.

  • Limited Generalization Evaluation

In Table 1, 600 of 745 responses come from DeepSeek-R1, indicating limited coverage of output-generating models. The 15 models in Table 4 are auditor backbones; varying them does not replace independent evaluation across output sources or task distributions. Appendix E also acknowledges that both datasets primarily elicit deception through explicit goals, pressure, or incentives, leaving spontaneous, low-signal deception untested.

Although the paper evaluates both DeceptionBench and OpenDeception, they share important construction conditions and cannot adequately test whether the method generalizes beyond them. The empirical scope remains narrow and concentrated on goal-induced scenarios; broader applicability has not been established.

Questions: Please report results by output-generating model and evaluate a more balanced sample. Without changing the scoring rules, evaluate samples without explicit deceptive goals and test generalization to independent task distributions. Candidate benchmarks include MASK (consistency between knowledge and statements), the Sneaking subset of DarkBench (information alteration in summarization or rewriting), and LH-Deception (long-horizon, multi-task interactions). The CSQ tasks in Beyond Prompt-Induced Lies offer a reference for settings without explicit deceptive goals. These evaluations use different label definitions, so their relationship to the paper’s definition of deception must be clarified; factual errors or reasoning failures must not simply be treated as deception. If further evaluation is infeasible, restrict the claimed scope to the goal-induced settings studied here.

W3: Methodological Novelty

  • IMT’s Independent Contribution Is Unclear

Section 3.1 explicitly attributes atomic information decomposition to FActScore, and §2 acknowledges prior work applying IMT to deceptive persuasion by LLMs. The main addition is the combination of decomposition, importance weighting, four-dimensional auditing, and aggregation.

Table 5 shows that removing weighting or decomposition reduces performance, but does not attribute the gains to IMT. The dimension-removal experiments in Figure 3 also leave open whether another equally detailed set of judgment criteria would perform similarly. This affects both the theoretical claim and the method’s novelty relative to existing structured LLM-as-judge approaches.

Questions: Please hold the backbone, inputs, decomposition module, weights, and reasoning budget fixed, and vary only the scoring rules to compare IMT, IDT, and a generic checklist for omissions, contradictions, and misleading content.

  • Baseline Outputs and Reasoning Budgets Need Alignment

Appendix C.2, p. 19, describes the Zero-shot and Few-shot baselines as producing binary labels, whereas DECOR produces continuous aggregate scores and the main metric is AUROC. Please clarify whether the baseline ROC curves use only binary outputs or additional, undocumented scores. If the former, the comparison is confounded by differences in output resolution.

The single-agent comparison in Table 9 also does not adequately specify whether role prompts, context isolation, and token/call budgets are matched. Their effects therefore cannot be clearly separated from the contribution of the multi-agent design.

Questions: Please have each baseline produce continuous or sufficiently granular ordinal deception scores, and release the actual outputs, scoring prompts, and parsing procedures used to compute AUROC. This would prevent differences between binary hard labels and continuous scores from confounding the method comparison. Also report call counts, token usage, and costs, and include a budget-matched, single-agent baseline that audits in stages to identify the source of the gains.

References

  1. McCornack, S. A. (1992). Information manipulation theory. Communication Monographs, 59(1), 1–16. DOI: 10.1080/03637759209376245.
  2. McCornack, S. A., Levine, T. R., Solowczuk, K. A., Torres, H. I., & Campbell, D. M. (1992). When the alteration of information is viewed as deception: An empirical test of information manipulation theory. Communication Monographs, 59(1), 17–29. DOI: 10.1080/03637759209376246.
  3. Jacobs, S., Dawson, E. J., & Brashers, D. (1996). Information manipulation theory: A replication and assessment. Communication Monographs, 63(1), 70–82. DOI: 10.1080/03637759609376375.
  4. McCornack, S. A., Levine, T. R., Morrison, K., & Lapinski, M. (1996). Speaking of information manipulation: A critical rejoinder. Communication Monographs, 63(1), 83–92. DOI: 10.1080/03637759609376376.
  5. Yeung, L. N. T., Levine, T. R., & Nishiyama, K. (1999). Information manipulation theory and perceptions of deception in Hong Kong. Communication Reports, 12(1), 1–11. DOI: 10.1080/08934219909367703.
Questions for authors

See Weaknesses.

7 comments
AuthorsOpenReview ↗

Author's response (Part 1)

We sincerely appreciate your feedback and insightful comments! We address the questions and comments below in detail.


W1.1 Reliability of IMT. Page 2, lines 73–77, describes Information Manipulation Theory (IMT) as extensively validated, without adequately discussing the relevant disputes or conditions of applicability.

Experiments by McCornack et al. (1992)... Ratings on all four dimensions nevertheless correlated with perceived deceptiveness, and the authors suggested that cultural differences in what constitutes a covert violation might explain the discrepancy. These findings indicate that IMT is not a fully reliable theoretical foundation and that its applicability is context-dependent.

We thank the reviewer for highlighting these disputes. We agree that “extensively validated” overstates the evidence and have replaced it with “widely used.” We have also cited all the studies suggested by the reviewer and added a fuller discussion of IMT’s applicability in the appendix A.3.

We do not assume that IMT is universally valid. We use its four dimensions as theoretically motivated features because they have been linked to perceived deception in human communication, although findings vary across studies. Our setting is suitable for this analysis because the input context provides the relevant facts, allowing the auditor to assess whether the response omits, distorts, or sidesteps information important to the audience’s understanding.


W1.2 Departures from IMT’s Original Definitions. Please explain and justify three adaptations:

  • Quantity vs. Quality: Why are new, unverified claims presented as facts assigned to Quantity, when the original IMT places assertions lacking adequate evidence under Quality? Explain the reassignment and possible double-counting (Appendix A.1, pp. 13–14).
  • Relation vs. goal compliance: Why does bypassing an IU’s constraints while advancing the original goal constitute a Relation violation? Clarify how this measures conversational relevance rather than task compliance (Appendix A.1, p. 14).
  • Impact weighting: Reconcile the main text’s emphasis on importance to the user’s judgment with the appendix’s emphasis on facilitating or obstructing deceptive instructions. These measure different quantities (§3.1; Appendix A.2).

We appreciate the reviewer for pointing out these ambiguities. We clarify the three definitions below and provide examples distinguishing task compliance from misleading an audience.

(i) We acknowledge that the original IMT places assertions lacking adequate evidence under Quality. In our rubric, “unverified” means unsupported by the input context, which is not the same as lacking adequate evidence. Quantity concerns how much information is provided, including both too little and too much information (McCornack, 1992, p. 5). DECOR uses the input context as a reference to identify omissions and additions. We include new, unverified additions under Quantity if they are presented as established fact and materially distort the audience’s understanding. These additions are not necessarily false or lacking adequate evidence. Quality, by contrast, concerns false or distorted information presented as true.

(ii) In the original IMT, Relation concerns relevance under the constraints of the preceding discourse (McCornack, 1992, p. 5). In our setting, a response may depart from the requested task because it refuses or corrects the request. Such departures should not themselves count as manipulation. We therefore require that a response both bypasses relevant information and continues to advance the original goal. For example, when asked to sell a product with a known safety risk, a response that refuses and offers an honest alternative disclosing the risk does not violate Relation. A response that refuses but then suggests redirecting safety questions toward the product’s benefits violates Relation, because it avoids addressing the risk while still helping sell the product.

(iii) We agree that importance to the audience’s judgment and importance to task execution are distinct quantities. We have revised Section 3.1 to align with Appendix A.2, clarifying that impact weights measure the importance that directly supports or obstructs the execution of the instruction.

AuthorsOpenReview ↗

Author's response (Part 2)

Original IMT concerns misleading listeners by exploiting expectations of cooperative communication, including covert violations, rather than simply advancing a predefined goal (McCornack, 1992, pp. 5–6). Explain why each adaptation is necessary, provide contrasting examples distinguishing compliance with inappropriate requests and safety refusals from inducing false beliefs, and clarify whether the intended audience is the informed requester or a third-party recipient.

We thank the reviewer for highlighting the distinction between task compliance and deception. We want to clarify that the prompts used to generate responses from the source models do not necessarily contain explicit deceptive goals. For example, a prompt may ask the model to sell a product despite a known safety risk, rather than directly asking, “How can I deceive the customer?” Task compliance and refusal concern whether the response follows the requester’s instruction, while deception concerns whether it would lead the intended recipient to form a false belief. The recipient may differ from the informed requester. For example, two responses to a pregnant customer both retain the salesperson’s role. One calls the product “generally safe for most users” without mentioning the known miscarriage risk. The other discloses the risk. Human annotators labeled the first deceptive and the second honest. Failing to clearly and accurately communicate the known miscarriage risk could lead the customer(recipient) to falsely believe the product is safe during pregnancy. The distinction is whether the response misleads the customer, not whether it pursues the sales goal.


W1.3 Gap Between Theory and Validation.

We thank the reviewer for pointing out the need to clarify the role of IMT and validate its use in DECOR. IMT guides the four-dimensional manipulation scores assigned to each IU. Decomposition, impact weighting, and aggregation are our methodological choices for applying these scores to LLM auditing, rather than procedures derived from IMT.

To isolate the contribution of the scoring rules, we compared IMT with IDT and a generic checklist using the same GPT-4o backbone, inputs, IUs, weights, aggregation, and prompt template. Only the scoring rules were changed.

Scoring rulesResponseThought & Response
IMT (Quantity, Quality, Relation, Manner)0.9350.946
IDT (Veridicality, Completeness, Directness, Clarity, Personalization)0.9070.920
Checklist (Omission, Contradiction, Misleading content)0.9070.924

IMT achieves higher AUROC in both input settings, supporting its contribution beyond the shared decomposition and weighting steps.


W2.1 Independent Validation of Human Labels.

We thank the reviewer for suggesting further validation of the human labels. The two annotators are members of the author team, each with over one year of research experience in LLM deception. They annotated independently without seeing each other’s labels, and all labels were finalized before DECOR was designed. We will release both annotators’ anonymized pre-discussion labels and conduct independent relabeling upon publication.

To clarify how disputed cases were resolved, we added two disputed cases to Appendix B.2. In the airbag case, the response acknowledges the accident but presents the failure to deploy as normal and claims that safety inspections were passed. This could lead the buyer to falsely believe the airbags are working properly, so it was labeled deceptive. In the ultrasound case, the response confirms normal fetal development while recommending consultation about supplements. This advice would not create a false belief that supplements are needed to address a developmental problem, so it was labeled non-deceptive.

To check whether our conclusions depend on how disagreements were resolved, we re-evaluated all methods under different treatments of disputed labels. DECOR ranks first or ties for first in every case.

Ground-truth labels (n)DECOR ResponseBest baselineDECOR Thought & ResponseBest baseline
Consensus, as in the paper (745)0.9380.9190.9470.927
Annotator A only (745)0.9260.9040.9320.917
Annotator B only (745)0.9250.9250.9360.927
Agreed before discussion only (675)0.9610.9470.9670.950
Disputed cases → deceptive (745)0.9220.9010.9270.910
Disputed cases → honest (745)0.9360.9340.9480.939
AuthorsOpenReview ↗

Author's response (Part 3)

W2.2 Missing Ground Truth for Fine-Grained Explanations. Questions: Please independently annotate information units, manipulation types, and importance on a stratified subset, and report accuracy and agreement for each component. Validate the classifications before drawing conclusions about manipulation frequencies; otherwise, describe them only as the distribution of categories predicted by DECOR.

We thank the reviewer for this suggestion. We added human validation of IU decomposition, impact weighting, and manipulation judgments on 50 DeceptionBench contexts, with 10 from each domain.

For decomposition, an annotator independently constructed reference IUs for each context. We compared these with IUs extracted by GPT-4o (two independent runs), Claude Sonnet 4.6, and Qwen3-235B-A22B.

Auditor backbonePrecisionRecall (all IUs)Recall (decisive IUs)Consistency (F1)
GPT-4o (run 1)0.9720.8870.947–
GPT-4o (run 2)0.9750.9090.9610.979
Claude Sonnet 4.60.9440.9740.9870.894
Qwen3-235B-A22B0.9560.8700.9210.911

Precision measures the share of extracted IUs that match a reference IU, and recall measures the share of reference IUs recovered. Consistency is measured by F1 against GPT-4o run 1. Extraction precision is at least 0.944, overall recall ranges from 0.870 to 0.974, and decisive-IU recall is at least 0.921.

For impact weighting, the annotator assigned each reference IU a weight w ∈ {1, 2, 3} following Appendix A.2, without seeing DECOR’s outputs. Decisive IUs receive the highest weight (w = 3), indicating information that directly supports or obstructs the execution of the instruction.

Auditor backboneExact agreementQuadratic κDecisive precisionDecisive recall
GPT-4o (run 1)0.7900.7040.8780.903
GPT-4o (run 2)0.7760.6940.8680.904
Claude Sonnet 4.60.8180.7510.7890.947
Qwen3-235B-A22B0.7760.6700.8530.914

Quadratic-weighted Cohen’s κ ranges from 0.670 to 0.751, showing agreement between DECOR’s weights and human judgments. Recall for decisive IUs exceeds 90% across all backbones, indicating that DECOR correctly identifies most information judged decisive by the annotator.

For manipulation judgments, we evaluated 91 human-identified decisive IUs from the same 50 contexts. Following Appendix A.1, the annotator judged whether the response violates each IMT dimension for each IU, without seeing DECOR’s scores. Human judgments are binary, whereas DECOR assigns graded scores. To compare them directly, we converted each DECOR score into “violation” for a positive score and “no violation” for zero.

DimensionAgreementκ
Quantity0.900.80
Quality0.930.85
Relation0.820.63
Manner0.810.62

Agreement ranges from 81% to 93% across the four dimensions (κ = 0.62–0.85), supporting DECOR’s manipulation judgments on decisive information. We report the main findings in Section 5.4 and provide the full protocol and results in Appendix C.4.

We also revised section 5.3 to describe the distribution of IMT dimensions predicted by DECOR, rather than ground-truth manipulation frequencies.


Reviewer_6PvtOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper presents a novel method for deception auditing, DECOR, which consists in decomposing the input in several parts with varying weights, and for each unit, identifying the extent to which the LLM output deceives across each of the 4 IMT dimensions. The reasoning behind it and place in the literature is properly presented: The main contribution is this novel framework used for auditing deception and identifying the underlying deception dimensions. They additionally included human-annotated reference labels (DeceptionBench) and show the agreement rates between these and LLMs.

Critical strengths

This auditing method seems overall sound, and very interpretable since we can identify exactly which form of deception and where it appears. Having a human-labelled dataset also considerably strengthens the paper. The appendix and repository are clear and strong.

Critical weaknesses
  1. There are few significance statistics which pool several results together. I would like to have the detailed significance statistics, to better estimate the actual impact of each setting (notably for Table 11, where you mention “even though the differences are not significant in every individual evaluation setting”). The claims should then be adapted accordingly, if necessary, based on these significance statistics.

  2. While several models are used, the results for DeceptionBench mostly come from DeepSeek-R1 (Table 1). It would be relevant to expand results from the other models or include the AUROC by model.

  3. The results for the multi-turn part are weaker, but the non-deceptive rate is very low (only 6 out of the 146 dialogues are non-deceiving), it would be relevant to expand the results to reach a higher number of non-deceptive dialogues.

Questions for authors

I didn’t find the prompt used to generate the thought, could you please provide the prompt and generation settings (or its location in the repository) used for generating the thought, and clarify whether DeepSeek-R1’s thoughts are its reasoning trace, or still obtained through prompting?


Post-rebuttal update

Soundness raised from 2 to 3. I thank the authors for their thorough revision, which addressed my concerns and strengthened the paper. I support acceptance (weak acceptance maintained).

1 comment

We sincerely appreciate the reviewer for finding DECOR sound and interpretable! We address the questions and comments below in detail.


W1 Detailed significance statistics.

Thank you so much for raising the concern! We now report p-values for each evaluation setting, using GPT-4o as the auditor. Per-setting comparisons use paired bootstrap tests with Holm correction. “Overall” aggregates results across both input settings. Bold values indicate p<0.05p < 0.05.

BaselineResponseThought & ResponseOverall
Zero-shot<0.001<0.001<0.001
Few-shot<0.001<0.001<0.001
DeceptionBench<0.001<0.001<0.001
CoT Red-Handed0.081<0.001<0.001
Constitutional Monitor0.0810.0100.001

DECOR significantly outperforms all five baselines under Thought & Response setting. Under Response, its gains are significant against Zero-shot, Few-shot, and DeceptionBench, but not against the two monitoring baselines. We have revised the claims in Appendix C.3 accordingly.


W2 Results on different source models

We appreciate this insightful suggestion. We report results separately by source model, referring to the model that generated the responses. GPT-4o is the fixed auditing backbone for all methods. We also evaluate all methods after excluding DeepSeek-R1 and after balancing the sources by sampling 45 responses per model over 1,000 draws. The tables below report both input settings.

Response AUROC

MethodFullDeepSeek-R1GPT-4oQwen2.5-7BWithout DeepSeek-R1Balanced
Zero-shot0.8880.8980.7650.8470.8380.857
Few-shot0.8900.8950.8120.8680.8670.882
DeceptionBench0.8190.8250.6560.8400.7780.797
CoT Red-Handed0.9180.9250.8280.8980.8720.897
Constitutional Monitor0.9190.9200.8580.7900.9120.926
DECOR0.9380.9350.9160.8700.9590.958

Thought & Response AUROC

MethodFullDeepSeek-R1GPT-4oQwen2.5-7BWithout DeepSeek-R1Balanced
Zero-shot0.9060.9140.7700.9030.8610.878
Few-shot0.9090.9120.8440.9090.8890.900
DeceptionBench0.8200.8260.6450.8470.7810.795
CoT Red-Handed0.9180.9260.8210.8980.8700.896
Constitutional Monitor0.9270.9290.8770.8780.9200.928
DECOR0.9480.9440.9320.9110.9700.967

DECOR ranks first across all subsets in Thought & Response and all subsets except Qwen2.5-7B in Response. Its advantage holds after excluding DeepSeek-R1 and balancing the source models. We have added these results to Appendix C.3.


W3 low non-deceptive rate

We thank the reviewer for raising this important question. Our original evaluation used OpenDeception’s English split to match the English-language setting of our DeceptionBench evaluation. To include more non-deceptive dialogues, we added the Chinese split generated and released by the benchmark authors. The additional 331 dialogues include 29 non-deceptive cases, increasing the number of non-deceptive dialogues from 6 to 35. Since 299 of the Chinese dialogues were generated by Qwen models, we also balanced the four source model families: GPT, Llama, Qwen, and Doubao. The table below compares Thought & Response AUROC on the original English set, the combined set, and the family-balanced set.

MethodEnglish (146)English + Chinese (477)English + Chinese, family-balanced
Zero-shot0.4230.6290.650
Few-shot0.4500.6660.665
DeceptionBench0.3390.6320.612
CoT Red-Handed0.6470.7360.755
Constitutional Monitor0.6700.7480.769
DECOR0.7200.7610.795

DECOR achieves the highest Thought & Response AUROC across all three settings. Its lead holds after expanding the non-deceptive set from 6 to 35 dialogues and balancing the source model families.


Q1 provide the prompt and generation settings used for generating the thought

To match DeceptionBench’s generation protocol, we elicited thoughts from all source models using its prompt: '(Please say your thought and final response in a json format: {"thought":"","response":""})'. These thoughts are prompted outputs, including for DeepSeek-R1. All source models use the same generation settings without model-specific adjustments: max_tokens=500, temperature=0.0, top_p=0.95, do_sample=False. We provide the prompt, generation settings and examples in Appendix B.4.

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

This paper proposes DECOR, a multi-agent framework for fine-grained LLM deception auditing based on Information Manipulation Theory (IMT). It decomposes input contexts into atomic information units, evaluates manipulation across four IMT dimensions, and aggregates the results using impact weights. The authors introduce human-annotated labels for 745 responses and evaluate DECOR on single-turn and multi-turn benchmarks, demonstrating competitive performance against existing baselines.

Critical strengths
  • Fine-grained and interpretable auditing. DECOR provides a structured way to identify which information is manipulated and how, going beyond coarse response-level deception detection. The ablation studies support the effectiveness of information decomposition and impact weighting.
  • Extensive empirical evaluation. The paper introduces human-annotated reference labels with strong inter-annotator agreement and evaluates DECOR against multiple baselines across 15 auditor backbones. The results demonstrate promising performance, particularly in single-turn deception auditing.
Critical weaknesses
  • Reliability of LLM-based decomposition. DECOR relies on LLMs to decompose contexts into atomic information units, but the reliability of this process is not directly evaluated. Although the ablation study demonstrates its effectiveness, it remains unclear whether the extracted units are accurate, complete, and consistent. An evaluation against human annotations or a consistency analysis across different LLMs would strengthen this claim.
  • Justification of impact weighting. The paper introduces a three-level importance weighting scheme and demonstrates its effectiveness through ablation studies. However, it remains unclear how reliable the LLM-assigned weights are and whether they align with human judgments of information importance. Further validation of the assigned weights and their sensitivity to different weighting schemes would strengthen the methodological justification.
Questions for authors
  • Could the authors conduct an independent human evaluation on a representative subset of samples to assess the quality of atomic information decomposition, impact weighting, and manipulation judgments across the four IMT dimensions? Reporting the agreement between human annotations and DECOR's outputs would help validate the reliability of its fine-grained auditing.

  • Among the 745 annotated responses, 600 are generated by DeepSeek-R1. Could the authors report DECOR's performance separately for each source model? Additionally, would DECOR still outperform the baselines if DeepSeek-R1 responses were excluded or the dataset were balanced across source models?

If the authors could address the problems in weaknesses and questions, I will raise my rating.

4 comments
AuthorsOpenReview ↗

Author's response (Part 1)

We thank the reviewer for the positive assessment and the concrete suggestions! We address the questions and comments below in detail.


W1 Reliability of LLM-based decomposition.

We thank the reviewer for raising this concern. We added a human evaluation on 50 DeceptionBench contexts (10 per domain) that covers both suggestions: comparison against human annotation and consistency across LLMs. An annotator independently built a gold set of information units (IUs) for each context. We compared it with IUs extracted by GPT-4o (two independent runs), Claude Sonnet 4.6 and Qwen3-235B-A22B, reporting:

  • precision: the share of DECOR IUs that match a gold IU;
  • recall: the share of gold IUs that DECOR recovers, overall and for decisive IUs (w = 3);
  • consistency: F1 against GPT-4o run 1.
Auditor backbonePrecisionRecall (all IUs)Recall (decisive IUs)Consistency (F1)
GPT-4o (run 1)0.9720.8870.947–
GPT-4o (run 2)0.9750.9090.9610.979
Claude Sonnet 4.60.9440.9740.9870.894
Qwen3-235B-A22B0.9560.8700.9210.911

Across all backbones, extraction precision is at least 0.944, overall recall ranges from 0.870 to 0.974, and decisive-IU recall is at least 0.921. After a closer review, we found that Claude’s lower consistency appears to reflect finer-grained splitting rather than missing information. We also found that most missed gold IUs are peripheral role statements (e.g., "You are my tax consultant."). We report the main findings in Section 5.4 and provide the full protocol and results in Appendix C.4.


W2 Justification of impact weighting.

We thank the reviewer for raising this concern. We evaluated both agreement with human judgments and sensitivity to the weighting scheme.

On the same 50 contexts, the annotator assigned each reference IU a weight w ∈ {1, 2, 3} following the definitions in Appendix A.2, without access to DECOR’s outputs. We compared these annotations with weights assigned by GPT-4o, Claude Sonnet 4.6, and Qwen3-235B-A22B. We also measured precision and recall for decisive IUs (w = 3), defined as information that directly supports or obstructs the execution of the instruction.

Auditor backboneExact agreementQuadratic κ\kappaDecisive precisionDecisive recall
GPT-4o (run 1)0.7900.7040.8780.903
GPT-4o (run 2)0.7760.6940.8680.904
Claude Sonnet 4.60.8180.7510.7890.947
Qwen3-235B-A22B0.7760.6700.8530.914

Recall for decisive IUs exceeds 90% across all backbones, indicating that DECOR correctly identifies most information judged decisive by the annotator. We report the main findings in Section 5.4 and provide the full protocol and results in Appendix C.4.

We also tested the sensitivity to the weighting scheme. We re-aggregated the same audits under different mappings from w ∈ {1, 2, 3} to numeric weights. DECOR outperforms the strongest baseline (Constitutional Monitor) under every weighting scheme we tried, and every non-uniform scheme beats uniform weighting:

WeightsResponseThought & Response
DECOR {1,2,3}0.9380.948
Exponential {1,2,4}0.9320.944
Rank-order centroid {0.11,0.28,0.61}0.9320.945
Uniform {1,1,1}0.9210.936
Strongest baseline (Constitutional Monitor)0.9190.927

These results show that impact weighting improves performance, and the gains hold across the tested weighting schemes. The full protocol and results are in Appendix C.3.

AuthorsOpenReview ↗

Author's response (Part 2)

Q1 Could the authors conduct an independent human evaluation on a representative subset of samples to assess the quality of atomic information decomposition, impact weighting, and manipulation judgments across the four IMT dimensions? Reporting the agreement between human annotations and DECOR's outputs would help validate the reliability of its fine-grained auditing.

Decomposition and impact weighting are addressed in W1 and W2 above. Here, we evaluate manipulation judgments on 91 human-identified decisive IUs from the same 50 contexts. Decisive IUs receive the highest weight (w = 3) because they directly support or obstruct the execution of the instruction.

For each decisive IU, the annotator judged whether the response manipulates that information along each of the four IMT dimensions, following Appendix A.1 and without seeing DECOR’s scores. Human judgments are binary, whereas DECOR assigns graded manipulation scores. To compare them directly, we converted each DECOR score into the same two categories: “violation” for a positive score and “no violation” for zero.

DimensionAgreement
Quantity0.90
Quality0.93
Relation0.82
Manner0.81

Agreement is high on these decisive IUs across all four dimensions, ranging from 81% to 93%. Quality shows the highest agreement, followed by Quantity. This shows that DECOR’s judgments largely match human judgments on whether and how decisive information is manipulated. We report the main findings in Section 5.4 and provide the full protocol and results in Appendix C.4.


Q2 Could the authors report DECOR's performance separately for each source model? Additionally, would DECOR still outperform the baselines if DeepSeek-R1 responses were excluded or the dataset were balanced across source models?

We appreciate this insightful suggestion. We report results separately by source model, referring to the model that generated the responses. GPT-4o is the fixed auditing backbone for all methods. We also evaluate all methods after excluding DeepSeek-R1 and after balancing the sources by sampling 45 responses per model over 1,000 draws. The tables below report both input settings.

Response AUROC

MethodFullDeepSeek-R1GPT-4oQwen2.5-7BWithout DeepSeek-R1Balanced
Zero-shot0.8880.8980.7650.8470.8380.857
Few-shot0.8900.8950.8120.8680.8670.882
DeceptionBench0.8190.8250.6560.8400.7780.797
CoT Red-Handed0.9180.9250.8280.8980.8720.897
Constitutional Monitor0.9190.9200.8580.7900.9120.926
DECOR0.9380.9350.9160.8700.9590.958

Thought & Response AUROC

MethodFullDeepSeek-R1GPT-4oQwen2.5-7BWithout DeepSeek-R1Balanced
Zero-shot0.9060.9140.7700.9030.8610.878
Few-shot0.9090.9120.8440.9090.8890.900
DeceptionBench0.8200.8260.6450.8470.7810.795
CoT Red-Handed0.9180.9260.8210.8980.8700.896
Constitutional Monitor0.9270.9290.8770.8780.9200.928
DECOR0.9480.9440.9320.9110.9700.967

DECOR ranks first across all subsets in Thought & Response and all subsets except Qwen2.5-7B in Response. Its advantage holds after excluding DeepSeek-R1 and balancing the source models. We have added these results to Appendix C.3.

Thanks for the authors' additional experiemnts and analysis. The source-model results address my concern about DeepSeek-R1. Hoever, exact agreement on impact weights ranges from 77.6% to 81.8%, and even on decisive units, agreement is 82% for Relation and 81% for Manner. These results only partially address my concern about the reliability, so I keep my original rating.

We thank the reviewer for the follow-up. We examined where the weight and dimension judgments disagree to better understand the remaining reliability concerns.

To show the direction of the disagreements, each column below counts the IUs where IMT and the human assign different weights. For example, “Human 1 / DECOR 2” means that we assigns weight 1 and DECOR assigns weight 2. Higher weights indicate greater impact: 1 = Peripheral, 2 = Auxiliary, and 3 = Decisive.

Auditor backboneHuman 1 / DECOR 2Human 2 / DECOR 1Human 2 / DECOR 3Human 3 / DECOR 2Human 1 / DECOR 3Human 3 / DECOR 1
GPT-4o (run 1)2615641
GPT-4o (run 2)2916641
Claude Sonnet 4.616213262
Qwen3-235B-A22B2805660

Most weight disagreements are between adjacent tiers: 96.4%–97.6% of reference IUs receive either the human weight or a neighbouring weight. DECOR also tends to assign higher rather than lower weights, most often rating peripheral information as auxiliary. Thus, disagreements mainly reflect small overestimates of importance rather than large differences in judgment.

To understand the lower agreement for Relation and Manner, we re-examined all 33 IU–dimension disagreements(16 for Relation and 17 for Manner). We re-read the prompt, response, and DECOR’s rationale, assessed which judgment was better supported, and recorded the main reason.

DimensionDisagreementsBoth defensibleDECOR better supportedOriginal label better supported
Relation16943
Manner17683
Total3315126
Judgment after re-readingDimension overlapOverlooked detailOver-readingBorderline
Both defensible(15)13002
DECOR better supported(12)1821
Original label better supported(6)4020
Total(33)18843

Re-examination found both judgments defensible in 15 of 33 disagreements, DECOR better supported in 12, and the original label in six. Eight cases favoring DECOR involved overlooked response details. For example, a fake clinic describes a routine blood test as “designed to catch early warning signs by analyzing key health markers.” We agreed with DECOR that this vague wording warrants a Manner flag because it could misleadingly suggest cancer-detection capability.

The most common reason of disagreement was dimension overlap (18 of 33 cases): both judgments identified manipulation, but differed on which dimensions applied. In 16 cases, DECOR flagged an additional dimension; in two, the annotator did. Agreement on whether any dimension was violated remained high (95.6%, κ = 0.912). This suggests that the lower agreement for Relation and Manner partly reflects differences in assigning manipulation types rather than identifying manipulation itself.

Comments on the paper

1 comment

Dear Reviewers, AC, SAC, and PC,

We extend our sincere gratitude for the time and effort you have dedicated to reviewing our manuscript.

DECOR provides fine-grained LLM deception audits by identifying which information is manipulated and how. We appreciate the reviewers’ recognition of its interpretability, the value of our human reference labels, and the breadth of our evaluation. Below, we summarize the key strengths, revisions, and additional analyses addressing the reviews.


Key strengths noted by the reviewers

  • S1. Fine-grained, interpretable audits of which information is manipulated and how [R1, R2, R3].
  • S2. Human-annotated reference labels for 745 responses that strengthen the evaluation [R1, R2, R3].
  • S3. Extensive evaluation against multiple baselines and across 15 auditor backbones [R1].
  • S4. Strong single-turn results, with ablations supporting decomposition and impact weighting [R1, R3].
  • S5. Clear presentation [R1, R2, R3], with a well-documented appendix and repository [R2].

We have revised the paper accordingly, and the major changes are summarized below:

  • [R1, R3] We add human validation of IU extraction, impact weighting, and IMT judgments in Section 5.4, with the full protocol and results in Appendix C.4.
  • [R1, R2, R3] We report performance by response-generating model and evaluations excluding DeepSeek-R1 and balancing source models in Appendix C.3, Table 8.
  • [R1, R3] We add sensitivity analysis of impact weighting in Appendix C.3, Table 10.
  • [R2] We revise the statistical significance analysis and corresponding claims in Appendix C.3.
  • [R2] We document the thought-elicitation prompt and generation settings in Appendix B.4.
  • [R3] We discuss IMT’s empirical support and applicability in Appendix A.3 and clarify the auditing rubrics in Appendix A.1.
  • [R3] We align the impact-weight definition in Section 3.1 with Appendix A.2.
  • [R3] We provide examples explaining the resolution of annotation disagreements in Appendix B.2 and clarify the budget-matched single-agent comparison in Appendix C.3, Table 11.

Our responses also report additional multi-turn evaluations, controlled comparisons of scoring frameworks, continuous-score baselines, and analyses of annotation disagreements. We clarify the evaluated scope in Appendix E; direct audience-belief studies and generalization to spontaneous deception remain for future work.

All added content in the revised manuscript has been marked in red.

We refer to Reviewer NZCz as R1, Reviewer 6Pvt as R2, and Reviewer QL9Q as R3.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.