acceptodds
Withdrawn from ICLR 2027

Probe-Driven Adversarial Co-Evolution for Distinguishing Hallucination and Deception in Large Language Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

While large language models (LLMs) are increasingly used for question answering and decision-making, they can still produce incorrect responses. These errors may arise from (1) hallucination, where the model unintentionally produces incorrect content, or (2) deception, where the model intentionally produces false or misleading information. Existing methods mainly rely on LLM outputs or follow-up interactions to identify errors, but usually do not determine whether an error comes from hallucination or deception. Distinguishing the two is essential for diagnosing model failures and applying targeted mitigation strategies. We argue that effective deception detection should actively uncover diverse deceptive patterns, enabling the detection to generalize across varying deception strategies. In this paper, we propose a black-box method that combines probe interaction with adversarial strategy co-evolution to distinguish hallucination from deception in LLMs. The probe interaction module collects behavioral evidence by observing how the model explains, verifies, and revises its answer under interrogation. The adversarial optimization module optimizes a lie generator and a discriminator through a multi-round competition, allowing the detection strategy to cover broader deceptive patterns and target deception-specific behavioral cues. Experiments across multiple LLMs and datasets show that our method outperforms strong baselines.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
large language model, agent deception, llm hallucination

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_NjMkOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper propose a probe-driven method to classify the hallucination and deception responses. It utilize a adversarial updating paradigm to generate hallucination-deception pairs and update probe set. Experiments show the effectiveness of proposed method on factual question-answering and science multi-choice questions.

Critical strengths
  • It is a new problem to identify hallucination and deception responses.
  • The proposed method is clear to understand.
Critical weaknesses
  • In practical application, we can simply treat both hallucination and deception as false responses and filter them out. Why do we need to identify hallucination and deception responses separately? Further statements about its significance is needed.
  • The proposed method is only tested on limited scopes. Is the proposed method generalized to other domains such as code, math, or open-world questions? More evaluations could be helpful.
  • The additional cost can not be ignored. There are 65 additional probe questions needed to sample with the audting model. Analysis on computation costs is needed.
Questions for authors

Please see the weaknesses.

Reviewer_WzJ6OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes a black-box method to distinguish hallucinated answers from deliberately generated false answers. It pairs natural Qwen3-14B errors with GPT-5.5-generated incorrect responses to the same questions, obtains responses to 65 diagnostic probes, and alternates generator and discriminator/probe updates over eight rounds. The selected strategy combines probes with learned textual feedback. Experiments on TriviaQA and SciQ report gains over monitoring and judging baselines, component ablations, and transfer across models and datasets.

Critical strengths
  1. A concrete adaptive method for a useful diagnosis problem. The motivation for distinguishing failure causes beyond factual correctness is clear (pp. 1–3, Sections 1–3.2). The generator/discriminator feedback and probe revision give a specific, understandable approach to explore that goal (pp. 4–5, Sections 3.3–3.5). This is a potentially useful methodological contribution even if its present interpretation needs narrowing.

  2. Informative evaluation within the constructed task. Table 2 studies individual components; Tables 3–4 report both successful and poor transfer. The split is by question, and Appendix A reports question-cluster bootstrap intervals, paired permutation tests and Holm correction, explicitly retaining the nonsignificant Qwen3.6/TriviaQA outcomes (p. 5; pp. 7–8; pp. 13–14). These details support learning about the adaptive procedure's behavior, although they do not resolve the identification concerns below.

Critical weaknesses
  1. Class and source-model provenance are confounded. Section 4.1.2 (p. 5, lines 264–269) assigns natural hallucinations to Qwen3-14B and the paired deceptive traces to GPT-5.5. Question pairing controls topic, but not source style, instructions, error construction or response structure. High held-out accuracy can therefore also reflect differences between the generation pipelines. Tables 3–4 do not report a control that breaks the source/class association in the underlying traces. This obstructs the paper's central claim about distinguishing error causes, rather than proving that the classifier actually uses any particular shortcut. Resolving it requires source-controlled evidence, not just an explanation of the existing question pairing.

  2. Separate-model replay is not direct observation of the original generator's behavior. Appendix B.2 (p. 15, lines 758–766) inserts a supplied reasoning trace and answer into a fresh conversation for each probe, while B.3 explicitly states that Model B is separate from the original trace generator (lines 789–807). Its replies can be useful text-mediated features, but do not directly observe whether the original generator knew the answer, persisted under its original objective, or revised the answer in its original context. This limits the abstract's behavioral interpretation and Section 4.6's discussion of probe cues. B.1 supplies a reference answer and explicitly requests an incorrect trace, so the synthetic intervention is meaningful; the unresolved issue is the bridge to a broader diagnosis of the original model's error cause. Validating that bridge, or substantiating a narrower contribution, requires more than a wording clarification.

Questions for authors
  1. Evidence for the current central claim: Is there an existing held-out control in which source does not determine class, such as a shared source producing both natural errors and instructed false answers, or a crossed source/class design? For instructed-false samples, how is correct-answer capability established, and how much of the gain remains? Such evidence could materially change my assessment. As supporting clarification, please state the paired populations, counts and invalid-output rules and reconcile Tables 1 and 5: for example, SciQ/DeepSeek ACC is 88.65−75.73=12.92 points in Table 1 but 14.21 in Table 5. Different subsets or aggregation may explain this. Appendix A already reports uncertainty and tests; the question concerns their populations and effect estimates.

  2. Clarification or validation of the measurement claim: Which exact models generate the original traces, probe replies, discriminator decisions and feedback in each setting? Does any reported experiment interrogate the original generator with its original instructions retained, or are all probes independent Model B replays? What evidence supports the claimed original-generator diagnosis in the latter case, and what narrower text-mediated classification claim would the authors retain if that bridge cannot be supported? The answers could change how much contribution the current results establish.

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

The method pairs GPT-generated deceptive CoTs with hallucinated ones, and co-evolves an adversary generator and a probe-based discriminator. Experiments classify deception v.s. hallucination on TriviaQA and SciQ across 4 LLMs against 5 baselines.

Critical strengths
  1. The paper is well-presented and well-motivated.
  2. The research topic of deception detection in LLMs is timely.
  3. The paper's insight on pairing both types of error per question to control topic/format is interesting.
Critical weaknesses
  1. The submission does not provide the AI Use Statement required by this year's Author Guidelines.
  2. The paper assumes that GPT-5.5 can generate CoTs that instantiate real deception, which may not be grounded and it needs further justification.
  3. The paper's presentation is below expectations. While Figure 1 has a very detailed layout of the pipeline flow, the concrete mathematical definitions and notations in the writing of the sections are missing. One example is that how feedback control in Section 3.5 is executed is not formally described in mathematical language. Pure words would not be effective, and they weaken the readability. Additionally, the submission may need to reconsider the visual presentation (e.g., put some boxes to bound the Prompts) for Appendix B. The current presentation of Appendix B is violating the template's right margin.
Questions for authors

N.A.

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

Related papers

Discussion (0)

Sign in to comment.