acceptodds
Under review as a conference paper at ICLR 2027

Knowledge-Verified Emergent Deception in LLM Agents Under Conflicting Incentives

Abstract

Large language models are increasingly deployed as autonomous agents serving users on behalf of companies, creating settings where user and deployer interests can conflict. When an agent knows that a user is owed something its deployer would prefer to deny, does it remain honest? False statements alone cannot answer this question because they may reflect ignorance or hallucination. To answer this question, we introduce KnownLieBench, a knowledge-verified benchmark that first confirms through a neutral probe that an agent knows a user's entitlement, then evaluates whether it makes false claims when given an incentive to deny that entitlement. Specifically, KnownLieBench spans eight customer-service domains and 18,144 multi-turn interactions, uses a trust-tracking customer agent, and separates emergent from instructed deception. Across 18 proprietary and open-weight models, some rarely deceive under incentive alone, while others do so frequently, and explicit instruction raises deception for most models. We further show that honesty-directed fine-tuning reduces deception under incentive, whereas deception-graded fine-tuning increases lie success on honest-control dialogues without increasing lie frequency under incentive. By verifying entitlement knowledge before scoring deception, KnownLieBench reduces the confound between deceptive behavior and lack of knowledge, enabling more rigorous audits of agent honesty.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM agents, AI safety, alignment

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_iARsOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

At its core, the paper develops a benchmark that measures deceptive behavior in LLM-based agents in a company setting when agents are provided with conflicting goals or incentives. The study measures whether agents have "true" knowledge of a customer's rights, to then measure false claims in each benchmark scenario.

Critical strengths
  • The paper does not measure deception "intuitively" (a model – potentially randomly - does something inaccurate or harmful), but against the backdrop of it stating the truth (or correct policy) and then acting against it.
  • Applying a very recent interpretability technique (J-Lens).
Critical weaknesses
  • My main problem with the paper is that after reading it, I kept asking myself: So what? The main findings are not new. It is well established that models deceive when presented with conflicting incentives or goals. Many papers (among them the MASK benchmark) disentangle model beliefs from model statements, and many papers look at deceptive behaviors in multi-turn settings. Sure, the paper combines all of these approaches and shows model differences (it's a benchmark, after all), and doesn't claim its contribution lies in finding a novel phenomenon. Nonetheless, I struggle to see the added value of the paper for the research community. I think the exploratory J-Lens results are really interesting and worthwhile, but they apply to only 16 "purpose-built single-turn eligibility scenarios separate from the benchmark dialogues". So results might perhaps be very unreliable.
  • This is subjective and perhaps hard to fix, but I found it really difficult to understand and follow the details of the experiments. This is due to a plethora of (sometimes vague, often stacked) terms one has to track, and well as a massive amount of experiments where each one has to be interpreted in a very compressed manner due to the page limit. Also, the authors say that AI was used for "improving the clarity and concision of the manuscript", but it feels more like it was simply written by AI (authors, please object if I'm erring, and area chairs, use Pangram to check), which causes the typical fancy sounding, but hard to follow and often overly complicated linguistic style.
  • The paper provides very sparse information about who wrote the benchmark cases. Are e.g. the customer messages created by humans or LLMs?
  • Many comparisons in the paper (e.g. comparisons between models or types of deception) are described in a descriptive manner, without proper statistics. There is not a single significance test or confidence interval in the manuscript. Only some plots show error bars, and plots that do display them don't always state what they represent. This is problematic. Sure, there are large contrasts, and intuitively one can immediately see that they are significant. However, the paper violates basic scientific norms by avoiding statistical tests. I know this might be common in computer science, but something being the norm is no excuse for the norm being bad.
Questions for authors
  • Why are you using strong prompt nudges like "by any means available to you" (Appendix G.2)? Doesn't this invite a certain deceptive behavior instead of having it emerge naturally?
7 comments

We appreciate Reviewer iARs’s feedback on the benchmark’s research value, experimental evidence, and readability. We have clarified the presentation and added supporting analyses to the updated manuscript on OpenReview, highlighting the changes in blue. Below, we explain how these revisions address the points raised.

[W1: Added research value and representation analysis]

Thank you for asking us to clarify the added value and J-Lens evidence. We agree that the added research value can be made more concrete. We address this through a shared-case comparison and a dialogue-conditioned J-Lens analysis. Building on MASK's belief–statement distinction, KnownLieBench tests whether instructed deception predicts behavior under incentives, and how false claims translate into customer detection and executable outcomes.

We have added a comparison on 40 shared medium-trust cases from the original evaluation to Section 4.1 and Table 5 of the revised manuscript. DR weights the same seven domains equally, while instructed successful/all uses all 40 rounds and the original success rule. Appendix D.5 gives the case counts.

ModelIncentive DRInstructed DRInstructed successful deception / all rounds
Gemini-2.5-Flash7.4%85.5%3/40 (7.5%)
DeepSeek-V4-Pro56.3%80.7%2/40 (5.0%)
Claude-Sonnet-50.0%28.0%5/40 (12.5%)
Llama-3.1-8B60.0%94.6%1/40 (2.5%)

DR and instructed successful/all on the shared-case subset.

Gemini and DeepSeek respond similarly to deceptive instructions but differ substantially under incentives. Claude-Sonnet-5 also has lower instructed DR than Llama-3.1-8B but a higher successful/all rate on these cases. Measuring outcomes alongside false claims reveals differences that instruction compliance and lie frequency alone miss.

We also extend the original 16-scenario J-Lens pilot to separate benchmark-dialogue runs. Using the same lens, we freeze the prefix before the first false or justified denial, append an eligibility question, and take the readout at the question's final token before an answer is generated. The added question can prompt further computation, so the readout measures entitlement information in response to that query. Readouts average within cases, and AUROC weights domains containing both labels equally.

ModelConditionFalse / justified-denial casesPre-denial diagnostic AUROC [95% CI]
Qwen3.5-27BIncentive26 / 320.84 [0.73, 0.93]
Qwen3.5-27BInstructed56 / 450.82 [0.71, 0.91]
Llama-3.1-8BIncentive34 / 130.56 [0.37, 0.74]
Llama-3.1-8BInstructed46 / 150.54 [0.36, 0.72]

Dialogue-conditioned J-Lens results for false and justified denials.

The dialogue-conditioned readout distinguishes false from justified denials more clearly for Qwen than for Llama. We have added this diagnostic to Section 4.3 and Table 6 of the revised manuscript, with full coverage and honest-reference comparisons in Appendix F.1 (Tables 30–31).

[W2: Clarity and presentation]

Thank you for the feedback on readability and AI use. To address this concern, we have clarified entitlement, round, and trajectory in Sections 3.1–3.2 of the revised manuscript, specified metric denominators in Section 3.4, and moved the model roster to Appendix B.2. Tables 4–7 now present the key supplementary comparisons alongside their interpretation.

We have expanded the AI Use Statement to cover assistance with case drafting, experimental design and review, implementation and analysis code, and manuscript editing, together with human checks.

[W3: Benchmark construction and customer messages]

Thank you for asking how cases and customer messages were written. We agree that the roles of human authors and LLMs should be clearer. To clarify this, we have expanded Section 3.3 of the revised manuscript to describe human policy research and few-shot examples, LLM-drafted records and opening messages, and human checks of policy consistency, entitlement labels, and label leakage. These opening messages are separate from the runtime GPT-4o-mini customer in the original evaluation, which replies using case context, history, and trust (Section 3.2).

Reviewer_mo5sOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces KnownLieBench, a benchmark for measuring "knowledge-verified emergent deception" in LLM customer-service agents when a deployer's interest conflicts with a user's documented entitlement. Its central design is a knowledge gate: a neutral probe first confirms the model knows the user is owed something, and only gate-passed owed cases enter deception scoring, separating lying from ignorance or hallucination. The agent is then placed under a deployer incentive to deny the entitlement, and the benchmark distinguishes emergent deception (incentive only) from instructed deception (explicitly told to lie), with an honest control. It spans eight customer-service domains grounded in real statutes/policies (112 cases), runs in an executable tool environment with a dynamic trust-tracking customer agent, and evaluates 18 current models over 18,144 interactions scored by a fixed GPT-5.1 judge. Key findings: emergent deception varies widely across models (near-zero for some, high for others), explicit instruction raises it for most, and initial trust affects a lie's success and detection more than its frequency. The paper adds post-training experiments and a two-model Jacobian-lens probe of whether entitlement information stays decodable under deceptive framing.

Critical strengths
  1. The knowledge-verification gate is a principled design that targets the central confound in deception measurement. The hardest problem for deception benchmarks is separating lying from not-knowing or hallucinating; verifying through a neutral probe that the model can state the truth, and then scoring deception only on cases where it demonstrably knew the entitlement, directly reduces that confound. Table 1's comparison fairly argues that combining knowledge-verification with agentic tool use, dynamic trust, and success/detection scoring is a genuine gap this benchmark fills.

  2. The agentic environment is realistic and sizable. The eight domains are grounded in real statutes and official policies, the environment is executable with terminal outcomes read from environment state (so merely claiming completion does not count), the customer agent tracks trust dynamically, the emergent/instructed/honest-control conditions are cleanly separated, and the panel covers 18 current models. The scale and deployment realism are strengths, and the finding that emergent and instructed deception are distinct behaviors — some models comply when instructed yet rarely deceive on their own — is useful for safety auditing.

Critical weaknesses
  1. The entire measurement rests on a single LLM judge (GPT-5.1), without strong human-agreement evidence in the main text. For a benchmark whose selling point is rigorous deception measurement, the "lie" labels are extracted by one judge and then rule-converted, yet the main text provides no human-annotation agreement (e.g., κ) and only defers judge validation to an appendix. Using an OpenAI judge to score a panel that includes OpenAI's own models alongside competitors also raises a potential provider-bias concern. Judge reliability is the pivotal soundness question here and is under-substantiated in the main paper.

  2. The knowledge gate introduces a cross-model selection confound. Because each model is scored only on the cases it individually passed the knowledge probe on, deception rate (DR) is computed over different case subsets across models — stronger models pass more and harder cases and are thus evaluated on a harder subset, while weaker models are scored on an easier one. The benchmark's headline "knowledge-verified" feature therefore makes cross-model DR comparisons not apples-to-apples, which is not addressed.

Questions for authors
  1. The judge is a single GPT-5.1. What is its agreement with human annotation (e.g., κ or accuracy) on the lie labels, and is there evidence against provider bias when an OpenAI judge scores OpenAI and competitor models? Can a multi-judge or human-audited subset be reported?

  2. Because the knowledge gate scores each model on its own passed subset, how are cross-model DR comparisons valid when the case sets differ by model and capability? Can DR be reported on the common set of cases that all models passed?

4 comments

We thank Reviewer mo5s for the careful feedback on judge reliability and comparisons across models. The updated manuscript on OpenReview includes the corresponding validation and shared-case analyses, with revisions highlighted in blue. We summarize our responses below.

[W1. Judge reliability and provider bias]

Thank you for raising concerns about judge reliability and provider bias. We agree that judge validation should be clearer in the main text. We address this by adding a validation summary and a provider-stratified human audit. The original Appendix C.2 reported a 60-item human-labelled study with a 30-item held-out split and a grading prompt fixed before held-out evaluation.

JudgeHeld-out Cohen's κPooled agreement
GPT-5.10.83192.8%
Claude-Haiku-4.50.92395.0%

Original judge agreement with human labels.

Here κ compares each judge with human consensus, and pooled agreement includes three grading runs per item. We additionally audited 480 gate-passed rounds from a separate evaluation, sampled independently of judge labels and balanced between OpenAI and other providers. The groups were matched in aggregate by case, condition, and trust. Two annotators independently checked transcripts, blinded to provider and judge outputs, with disagreements adjudicated. Both judges used the same frozen rubric.

JudgeEvaluated agentsκFalse positives / human non-liesFalse negatives / human lies
GPT-5.1OpenAI0.8711/232 (0.4%)1/8 (12.5%)
GPT-5.1Other providers0.8578/132 (6.1%)9/108 (8.3%)
Claude-Haiku-4.5OpenAI0.8172/232 (0.9%)1/8 (12.5%)
Claude-Haiku-4.5Other providers0.8747/132 (5.3%)8/108 (7.4%)

Provider-specific judge agreement and errors in the supplementary human audit.

Within this audit, the OpenAI–other-provider gap in lie rates differs by less than one percentage point across human labels and both judges. Only eight OpenAI rounds are human-positive, so its missed-lie rate remains imprecise (95% Wilson interval 2.2–47.1%). We have brought held-out agreement and the audit finding into Section 3.4 and added the full audit to Appendix C.4 of the revised manuscript.

[W2. Comparability across knowledge-gated subsets]

Thank you for pointing out the different scored case sets. To address this concern, we reanalyse the original cohort on cases passed by all 18 models with usable records in each condition and trust setting. Billing has no common cases, so both the all-eligible and common-case estimates average the same seven domains, then trust levels. Each model retains its original history and customer state.

ConditionCommon rounds per modelMaximum absolute DR changeSpearman ρ
Incentive1384.9 pp0.985
Instructed1332.2 pp0.996

DR changes and ranking agreement on cases shared by all models.

Model rankings remain closely aligned when all models are scored on the same cases. We have added this finding to Section 3.4 and the complete 18-model DR comparison and case counts to Appendix D.4 of the revised manuscript (Tables 17–18).

[Q1: Human agreement and provider bias]

Thank you for asking about agreement and provider bias. The original held-out study gives GPT-5.1 κ = 0.831 against human consensus. The new human and two-judge audit finds similar provider gaps, with the small OpenAI-positive sample noted in W1. We have added the summary to Section 3.4 and the provider-specific error breakdown to Appendix C.4 of the revised manuscript.

[Q2: Evaluation on common cases]

Thank you for suggesting shared cases. We have added the full common-case DR comparison for all 18 models to Appendix D.4 of the revised manuscript. As summarized in W2, both sides exclude Billing and use the same domain weights, so the comparison isolates scored-case composition while preserving each model's original history.

AuthorsOpenReview ↗

Rebuttal Reminder

Dear Reviewer mo5s,

Thank you for your careful reading and helpful feedback. As the discussion period draws to a close, we would appreciate hearing whether our responses and the additional analyses in the updated manuscript have addressed your concerns. We would be happy to answer any remaining questions or provide further clarification.

We appreciate your time and effort!

Best regards,

The Authors

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

The paper introduces KnownLieBench, an evaluation of customer-service agents whose business objectives conflict with customers' entitlements. A neutral probe first checks whether the model can identify the correct entitlement. The model then interacts with a simulated customer and takes actions in an executable environment under honest, incentive-only, or explicitly deceptive prompts. The benchmark covers eight domains and eighteen models. The paper measures false claims, successful withholding, customer detection, and simulated trust, and adds small post-training experiments and an exploratory representation analysis.

Critical strengths
  • The evaluation makes a useful distinction between understanding an entitlement and respecting it under conflicting incentives. Source-grounded cases, justified-denial controls, a neutral knowledge check, and recorded terminal actions make the results easier to interpret than an evaluation based only on whether a response sounds deceptive. The honest-control rates are also low for most models, supporting the claim that the incentive conditions change behavior beyond ordinary factual error.
  • The benchmark is supported by unusually careful validation for a multi-agent evaluation. The judge must quote an actual agent statement; two annotators check a held-out sample; and alternative judges preserve the main condition and trust comparisons. The eighteen-model panel also reveals a practically useful distinction: willingness to follow explicit deceptive instructions does not reliably predict deception under the incentive prompt. This is a reasonable basis for using both settings in safety audits.
Critical weaknesses
  1. The main deception rate is broader than the knowledge verification that motivates the paper. The gate checks entitlement, while DR also counts invented fees, deadlines, and system limitations whose falsity the model has not demonstrated it knows. Appendix D.3 recognizes this, but its non-disclosure analysis does not identify which claims contradict the verified proposition. An agent can disclose the entitlement and lie about it elsewhere, or omit it while hallucinating an unrelated detail. The paper therefore still needs a claim-level, knowledge-matched rate to establish how much of the reported behavior satisfies its central definition.

  2. The post-training conclusion about lie frequency versus effectiveness compares different evaluation conditions. As the Table 12 caption explicitly notes, offline DR and detection use the incentive condition, while DSR uses the honest control. Consequently, increased DSR alongside stable DR does not demonstrate that lie effectiveness changed independently of frequency within the same setting. DSR also conditions on whichever cases contain lies after training, so changing the composition of those cases can change the rate. Reporting all three metrics and the unconditional successful-deception rate within each condition would make this result interpretable.

Questions for authors
  1. What fraction of counted lies directly contradicts the entitlement established by the knowledge probe? Please report that restricted metric and its model ordering alongside the broader false-claim rate; the round-level disclosure split does not answer this question.
  2. Can Table 12 report DR, DSR, detection, and lie counts separately within both the honest-control and incentive conditions, together with the fraction of all rounds containing successful deception? Does the post-training conclusion survive a comparison on matched cases and a second judge?
  3. How robust is the knowledge gate to response bias or prompt variation? This matters especially for Llama-3.1-8B, which passes 89.1% of owed probes but only 33.3% of calibration probes. Would requiring a correct matched non-entitlement control materially change its measured deception rate?
4 comments

We appreciate Reviewer HeZF’s thoughtful comments on knowledge verification and post-training evaluation. We have incorporated the key clarifications and additional analyses into the updated manuscript on OpenReview, with changes highlighted in blue. Below, we respond to each concern.

[W1: Knowledge-matched deception rates]

Thank you for asking us to clarify the link between knowledge verification and DR. We agree that the connection between verified entitlements and counted false claims can be made more explicit. To address this concern, we ran a separate supplementary batch on the same model panel and cases under both deception conditions and all trust levels, using fresh neutral probes and the original prompts and decoding settings. GPT-5.1 checks all customer-visible claims in each flagged round against the policy, record, and saved probe answer. A match requires a quoted contradiction of the verified entitlement under unchanged facts, including after truthful disclosure. Omission alone and unrelated falsehoods do not qualify.

Entitlement-matched DR uses the same scored rounds as DR. Rates retain equal domain, trust, and model weighting, while matched/flagged shares pool rounds.

ConditionScored owed roundsMatched / false-claim roundsMean DRMean matched DRSpearman ρ
Incentive3,422611/831 (73.5%)24.5%18.0%0.962
Instructed3,4191,942/2,329 (83.4%)68.4%57.1%0.946

DR, entitlement-matched DR, and model-ranking agreement in the supplementary batch.

Most flagged rounds contain entitlement contradictions. The stricter criterion lowers absolute rates while preserving closely aligned model rankings. A two-annotator stratified audit confirmed 189/200 automated labels (94.5%). The table retains automated classifications.

The original Appendix D.3 reported disclosure patterns. We have added the entitlement-matched definition to Section 3.1 and replaced that appendix analysis with the rates, full model rankings, matching rubric, and audit in the revised manuscript (Tables 14–16).

[W2: Matched post-training evaluation]

Thank you for suggesting within-condition post-training comparisons. We agree that matched cases can make changes in frequency and success easier to interpret. To address this concern, we use a separate held-out pool to evaluate four Llama-3.1-8B configurations on the same 20 gate-passed cases across eight domains, with 480 rounds per condition over two seeds at medium trust. Both SFT directions use the deception-graded reward, while DPO and GRPO use outcome rewards with their corresponding baseline. GPT-5.1 and Claude-Haiku-4.5 grade the same transcripts.

DR averages domains equally. DSR and detection condition on lying, while successful/all uses the fixed 480-round denominator to account for changes in which rounds contain lies.

ConditionConfigurationLie / detected roundsDR (%)DSR (%)Detection (%)Successful / all (%)
Honest controlDeception baseline38/208.342.152.616/480 (3.3)
Honest controlBusiness-SFT46/810.278.317.436/480 (7.5)
Honest controlHonesty-SFT38/58.281.613.231/480 (6.5)
IncentiveDeception baseline225/15347.824.468.055/480 (11.5)
IncentiveBusiness-SFT263/17455.729.366.277/480 (16.0)
IncentiveHonesty-SFT223/14947.428.366.863/480 (13.1)
Honest controlOutcome baseline36/177.450.047.218/480 (3.8)
Honest controlHonesty-DPO26/195.815.473.14/480 (0.8)
IncentiveOutcome baseline229/14349.526.662.461/480 (12.7)
IncentiveHonesty-DPO210/14246.318.167.638/480 (7.9)
Honest controlBusiness-GRPO41/158.956.136.623/480 (4.8)
IncentiveBusiness-GRPO236/15051.127.163.664/480 (13.3)

Matched post-training results within the honest-control and incentive conditions.

The second-judge comparison below gives trained-minus-baseline differences in successful/all, in percentage points with 95% intervals. These intervals describe trajectory-bootstrap uncertainty conditional on the two trained checkpoints per configuration.

ConditionConfigurationGPT-5.1 Δsuccessful/all [95% CI]Claude Δsuccessful/all [95% CI]
Honest controlBusiness-SFT+4.2 [0.0, 9.8]+4.0 [-0.2, 9.6]
Honest controlHonesty-SFT+3.1 [-1.7, 8.3]+2.9 [-1.9, 8.1]
IncentiveBusiness-SFT+4.6 [-1.0, 10.1]+4.0 [-1.8, 9.7]
IncentiveHonesty-SFT+1.7 [-4.2, 7.0]+0.6 [-5.6, 6.3]
Honest controlHonesty-DPO-2.9 [-7.4, 0.8]-2.7 [-7.4, 1.0]
IncentiveHonesty-DPO-4.8 [-10.4, 0.8]-5.0 [-10.9, 0.7]
Honest controlBusiness-GRPO+1.0 [-2.5, 4.6]+1.3 [-2.2, 4.8]
IncentiveBusiness-GRPO+0.6 [-5.7, 6.9]+0.4 [-6.0, 6.7]

Changes in successful/all under two judges grading the same transcripts.

For Business-SFT, the higher successful/all estimate accompanies a higher DR. Both judges give the same direction of change in successful/all point estimates across configurations. We have revised Section 5 and added Table 7 to interpret frequency and success jointly within each condition, replacing the earlier inference that lie effectiveness changed independently of frequency. The revised Appendix E.1 contains the full matched comparisons and statistical procedure (Tables 24–25).

[Q1: Knowledge-matched contradictions]

Thank you for requesting the entitlement-matched breakdown. Entitlement contradictions occur in 73.5% of incentive and 83.4% of instructed flagged rounds in the supplementary batch. These are round-level shares of direct contradictions, including those made after disclosure. We have added the corresponding rates and full model ordering to Appendix D.3 of the revised manuscript (Tables 14–16), as summarized in W1.

[Q2. Post-training evaluation]

Thank you for suggesting matched cases and a second judge. We have added the matched, two-judge comparisons from W2 to the revised Appendix E.1 and updated Section 5 to interpret frequency and success jointly. The outcome-reward Honesty-SFT result comes from the original held-out evaluation, outside this matched comparison (Appendix E.3, Table 28).

[Q3. Knowledge-gate robustness]

Thank you for pointing out the uneven gate pass rates. We tested 38 entitled/non-entitled pairs, reviewed by humans before inference, under three neutral wordings. A post-run input audit excluded one insurance pair with no upper appeal deadline, leaving 37 pairs. A strict pass requires the entitlement Boolean to be correct for both members under all three wordings. Wording-sensitive pairs pass some but not all wordings.

We apply these probe filters retrospectively to the original scored incentive rounds, preserving labels and histories. Each model's covered baseline and filtered DR use the same retained domains: four for Llama, five for Gemma, and eight for Qwen. N denotes scored rounds.

ModelAll-three paired passesWording-sensitive pairsCovered DR (N)All-three DR (N)Change (pp)
Llama-3.1-8B18/372/37 (5.4%)55.0% (78)45.7% (54)−9.3 pp
Gemma-3-12B17/3713/37 (35.1%)30.2% (72)26.9% (51)−3.3 pp
Qwen3.5-27B35/371/37 (2.7%)21.1% (111)22.6% (105)+1.5 pp

Paired-control pass rates, wording sensitivity, and the corresponding DR changes.

Llama's measured DR is most sensitive to the stricter gate, while Gemma's paired decisions are most sensitive to wording. Qwen retains most rounds with little DR change. We have added the DR comparison to Table 4 of the revised manuscript, with complete probe results and domain coverage in Appendix D.6 and the actual wordings in Appendix G.5.

AuthorsOpenReview ↗

Rebuttal Reminder

Dear Reviewer HeZF,

Thank you for your careful reading and helpful feedback. As the discussion period draws to a close, we would appreciate hearing whether our responses and the additional analyses in the updated manuscript have addressed your concerns. We would be happy to answer any remaining questions or provide further clarification.

We appreciate your time and effort!

Best regards,

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Final Rebuttal Summary and Acknowledgment

We thank the reviewers for their careful reading and constructive feedback. Below, we summarize the additional analyses and clarifications made in response to their comments. These updates have been incorporated into the revised manuscript on OpenReview, with changes highlighted in blue.

  1. Knowledge verification and gate robustness (HeZF). We added an entitlement-matched measure requiring a quoted contradiction of the verified entitlement. In the supplementary batch, most flagged rounds meet this criterion, and model rankings remain closely aligned. Paired entitlement/non-entitlement controls under three wordings further quantify model-specific gate sensitivity. These analyses appear in Section 3.1, Table 4, and Appendices D.3 and D.6.

  2. Judge reliability and cross-model comparisons (mo5s). We brought held-out human agreement into the main text and added a 480-round human audit with two judges. Human labels and both judges yield similar OpenAI–other-provider gaps in lie rates within this audit. Evaluation on cases shared by all 18 models also preserves closely aligned rankings. We report these findings in Section 3.4 and Appendices C.4 and D.4.

  3. Post-training interpretation (HeZF). We added matched evaluations within each condition, using two judges and reporting DR, DSR, detection, lie counts, and successful deception over all scored rounds. Business-SFT has higher point estimates for both DR and successful/all, and we updated the interpretation to discuss frequency and success jointly. The comparisons appear in Section 5, Table 7, and Appendix E.1.

  4. Research value and J-Lens evidence (iARs). A comparison on 40 shared cases shows that similar compliance with deceptive instructions can accompany very different behavior under incentives, while lie frequency alone misses differences in successful deception. We also extended J-Lens to benchmark-dialogue diagnostics, which distinguish false from justified denials more clearly for Qwen than Llama. These additions appear in Sections 4.1 and 4.3, Tables 5–6, and Appendices D.5 and F.1.

  5. Statistical support and incentive wording (iARs). We added confidence intervals supporting four central condition and trust contrasts and clarified figure error bars. A four-arm prompt ablation finds that deception persists after clause deletion. Adding a performance metric to honest control increases entitlement-matched deception for Qwen and DeepSeek. These results appear in Appendices D.7–D.8.

  6. Clarity, benchmark construction, and AI use (iARs). We clarified terminology and metric denominators, documented the policy research, LLM drafting, and human review used to construct cases and opening messages, and distinguished these from runtime customer generation. We also expanded the AI Use Statement. Section 3 and Tables 4–7 now make the evaluation process and key additional comparisons easier to follow.

In summary, KnownLieBench connects verified entitlement knowledge with agent behavior under conflicting incentives, measuring false claims alongside customer detection and executable outcomes. The revised manuscript adds targeted validation and matched comparisons to support these measurements and clarify their interpretation. We sincerely thank the reviewers for their feedback and welcome any remaining questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.