acceptodds
Under review as a conference paper at ICLR 2027

Generate, Verify, Retract: Exploiting the Generation–Verification Gap in Vision-Language Models

Abstract

Vision-language models (VLMs) can generate visually unsupported objects, which then become part of the autoregressive context and may influence subsequent generation. We uncover a substantial **generation–verification gap**: on paired generated object events, the same frozen VLM's explicit object-existence judgments detect hallucinations substantially better than generation-token probabilities (AUROC 0.8584–0.9022 vs. 0.6093–0.6155). This suggests that although a VLM may fail to generate only visually grounded content, it can often recognize an unsupported object once that object has been proposed. Motivated by this observation, we introduce **RETRACT**, a decoding framework that transfers the model's explicit verification capability into its generation process. RETRACT distills object-existence judgments from the frozen VLM into a lightweight verifier operating on candidate-specific internal representations. During decoding, each object candidate is verified after being processed but before being retained in the generation context, allowing unsupported candidates to be locally retracted before they influence subsequent generation. This design avoids repeated explicit verification queries and requires no modification to the VLM parameters. Across multiple VLM backbones and hallucination benchmarks, RETRACT consistently reduces object hallucinations while preserving generation quality, with only modest inference overhead. Code is available at [https://anonymous.4open.science/r/RETRACT/](https://anonymous.4open.science/r/RETRACT/).

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Vision-Language Models, Object Hallucination, Visual Grounding, Hallucination Mitigation, Autoregressive Decoding, Visual Verification

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_CDyVOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
2(−1)
Soundness
2
Read review
Paper summary

This paper studies the gap between generation and explicit object-existence verification in vision-language models. On paired object events from 10,000 COCO images, explicit verification by the same frozen model identifies unsupported objects substantially better than generation-token probabilities. Motivated by this observation, the paper proposes RETRACT. A lightweight linear verifier is trained to predict the model’s own object-existence judgments from candidate-specific, post-commit hidden states. During decoding, eligible object candidates are processed and checked before being retained. Rejected candidates are removed from the output, their KV-cache entries are masked, and replacements are selected using the saved pre-candidate logits. Experiments cover three backbones and CHAIR, AMBER, and OpenCHAIR, with additional analyses of verification signals, representation positions, retraction behavior, and inference cost. The results generally show reduced object hallucination with small measured latency overhead, although some content-coverage loss remains.

Critical strengths

A well-supported motivation based on paired measurements. The generation–verification comparison evaluates both signals on the same object events across three backbones and reports image-bootstrap confidence intervals. Verification achieves AUROC values of approximately 0.8580.858–0.9020.902, compared with 0.6090.609–0.6160.616 for generation-token risk. This provides convincing evidence for the specific gap being exploited, beyond illustrative examples.

A practical approach with reasonably broad empirical validation. Distilling the frozen backbone’s judgments into a linear verifier is computationally attractive. Experiments span three model families and three captioning benchmarks, while reporting coverage and latency alongside hallucination metrics. The post-commit, uncertainty-based, and random-rejection controls provide useful evidence that the learned verification signal contributes to the improvements.

Critical weaknesses

The evidence does not establish the claimed benefit of removing rejected candidates from future context. In Table 4, removing KV masking improves both CHAIR-i and CHAIR-s on all three backbones. For Qwen and Gemma, it also improves recall relative to full RETRACT; for example, Qwen CHAIR-i improves from 6.966.96 to 5.945.94, while recall increases from 64.1964.19 to 64.7664.76. The paper reports this result transparently, but it weakens the explanation that preventing rejected objects from influencing subsequent generation is responsible for the gains. Furthermore, similar output lengths do not establish preserved information content: Table 1 shows Qwen CHAIR recall decreasing from 66.9066.90 to 64.1364.13. The evidence supports useful hallucination reduction, but the mechanism and content-preservation claims should be narrower.

The incremental contribution over closely related verification-guided decoding methods is not sufficiently established. PSRD already distills a VLM’s self-evaluation signals into a lightweight reward model for online hallucination correction, while streaming UniProbe uses internal representations from a frozen backbone to detect, reject, and re-decode potentially hallucinated tokens. RETRACT’s combination of self-distillation, a linear post-commit verifier, and local retraction offers a potentially useful simplification. However, although these related methods are acknowledged, neither is included in the experimental comparisons. The comparison with an adapted TruthPrInt baseline does not fully establish the advantage over these closer alternatives. A focused comparison with one representative method, reporting both hallucination reduction and inference cost, would help determine whether the proposed differences constitute a substantial advance.

The transfer from explicit verification to the lightweight verifier is insufficiently characterized. The reported AUROC advantage in Appendix B.5 concerns explicit object-existence queries to the backbone, whereas deployment relies on a linear verifier trained to approximate those judgments. The post-commit ablation supports the usefulness of candidate-specific representations, but does not directly establish how much of the teacher’s verification advantage survives distillation or how often grounded objects are incorrectly rejected at the selected threshold. This matters because hallucination reduction is accompanied by lower object recall; for example, Qwen CHAIR recall decreases from 66.9066.90 to 64.1364.13. Similar caption lengths do not resolve this concern. A held-out assessment of the deployed verifier’s detection performance and false-rejection rate would clarify the effectiveness of the central distillation step and the resulting coverage trade-off.

Questions for authors

The paper presents solid work, and I would be open to raising my score if the authors can satisfactorily address my questions.

  1. Novelty and comparison with closely related methods — clarification. PSRD distills self-evaluation signals for online correction, while streaming UniProbe uses internal representations to detect and reject potentially hallucinated tokens. What is RETRACT’s principal technical advantage over these approaches, beyond using a simpler verifier? Could you clarify why these methods were not included in the experiments? A focused comparison with one representative method on a single backbone and benchmark, including inference cost, would help establish the practical significance of the differences.
  2. Effectiveness of the distilled verifier — limited additional analysis. Appendix B.5 establishes the advantage of explicit verification over generation-token confidence. How much of this advantage is retained by the deployed linear verifier? Could you report its detection performance on held-out object events and the false-rejection rate for grounded objects at the selected threshold τ\tau? This would help explain the observed reduction in object recall and assess the trade-off between hallucination mitigation and content preservation.
Reviewer_2WvQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper tries to addresses object hallucination in vision-language models by exploiting the gap between their generation and explicit verification capabilities. The authors propose RETRACT, a decoding framework that distills object-existence judgments from a frozen VLM into a lightweight verifier and uses post-commit verification with local retraction to correct unsupported object candidates during generation. Experiments across three VLM backbones and multiple hallucination benchmarks demonstrate consistent reductions in object hallucinations while largely preserving generation quality and introducing only modest inference overhead.

Critical strengths
  1. The paper identifies a substantial discrepancy between VLMs' object generation and explicit verification capabilities through paired object-event analysis. This observation provides a clear and well-supported motivation for incorporating self-verification into autoregressive decoding.
  2. RETRACT integrates lightweight self-distilled verification with post-commit candidate assessment and local retraction, enabling unsupported objects to be corrected before influencing subsequent generation. The framework operates without modifying the backbone parameters or repeatedly invoking explicit verification queries.
Critical weaknesses
  1. Although KV-cache masking is intended to maintain consistency in the generation context, Table 4 shows that removing this component achieves lower hallucination rates across all three backbones. Further analysis would help clarify its necessity and potential benefits.
  2. Table 1 shows a decrease in object recall across all three backbones when using RETRACT. A more detailed analysis of correct and erroneous retractions would help clarify the trade-off between hallucination reduction and preservation of valid visual content.
  3. RETRACT shares similarities with existing self-verification-based hallucination mitigation methods. A more detailed discussion of the methodological differences and direct comparisons would help better highlight the novelty and advantages of the proposed framework.
Questions for authors
  1. Table 4 shows that removing KV-cache masking consistently achieves lower hallucination rates across all three backbones. Could the authors explain why KV-cache masking is still necessary and provide further evidence of its practical benefits?
  2. Table 1 shows that RETRACT reduces object recall across all three backbones. Could the authors provide a more detailed analysis of correct and erroneous retractions to clarify whether hallucination reduction comes at the expense of valid visual content?
  3. RETRACT shares similarities with existing self-verification-based hallucination mitigation methods. Could the authors further elaborate on the key methodological differences?
Reviewer_sGQxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes RETRACT, a decoding framework designed to mitigate object hallucinations in vision-language models (VLMs). Motivated by the observation that VLMs can verify object existence more reliably than their generation probabilities suggest, the authors train a lightweight linear verifier using self-generated verification signals. During inference, the verifier evaluates candidate object tokens based on their post-commit hidden representations and selectively retracts potentially hallucinated objects through KV-cache masking and token reselection. Experiments across three VLM backbones and multiple hallucination benchmarks demonstrate improvements over several existing decoding methods with relatively low inference overhead.

Critical strengths
  1. Empirical improvements: The proposed method reduces object hallucinations across three VLM backbones and multiple benchmarks. The experiments cover several representative decoding baselines and demonstrate the potential effectiveness of verification-guided correction.

  2. Efficient inference-time correction: RETRACT introduces relatively small inference overhead by employing a lightweight verifier and local token retraction without modifying the backbone parameters. This is a practically appealing design.

Critical weaknesses
  1. The proposed framework combines a learned verifier with selective token rejection and resampling. However, similar ideas have already been explored in verification-guided hallucination mitigation, including REVERSE [1], TruthPrInt [2], PSRD [3], and UniProbe [4].
    In particular, PSRD distills self-evaluation signals into a lightweight reward model for online decoding, while UniProbe employs a learned hallucination detector with token-level rejection and resampling. Although these approaches are discussed in the related work, quantitative comparisons with PSRD and UniProbe are missing. Given that RETRACT also requires training an additional verifier, the current comparisons are insufficient to establish its distinctive advantages over closely related learning-based approaches.unclear.

    • [1] Wu, T.-H., Lee, H., Ge, J., Gonzalez, J. E., Darrell, T., and Chan, D. M. (2025). Generate, but Verify: Reducing Hallucination in Vision-Language Models with Retrospective Resampling. Advances in Neural Information Processing Systems (NeurIPS), 38. https://doi.org/10.52202/085713-2206
    • [2] Duan, J., Kong, F., Cheng, H., Diffenderfer, J., Kailkhura, B., Sun, L., Zhu, X., Shi, X., and Xu, K. (2025). TruthPrInt: Mitigating Large Vision-Language Models Object Hallucination Via Latent Truthful-Guided Pre-Intervention. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp. 7372–7382. https://doi.org/10.1109/ICCV51701.2025.00692
    • [3] Zhang, Y., Sun, C., Chen, K., Bai, X., Xiang, Y., and Zhang, M. (2026). Mitigating Multimodal Hallucination via Phase-wise Self-reward. arXiv preprint arXiv:2604.17982. https://arxiv.org/abs/2604.17982
    • [4] Samuel, D., Bar-Shalom, G., Frasca, F., Fetaya, E., Ziser, Y., Chechik, G., and Maron, H. (2026). UniProbe: A Learnable Token-Level Hallucination Detector for Large VLMs using Multi-Structural Internal Representations. arXiv preprint arXiv:2608.10835. https://arxiv.org/abs/2608.10835
  2. Insufficient validation of the learned verifier. The central motivation relies on the generation–verification gap observed through explicit object-existence queries to the frozen VLM. However, the actual retraction decisions in Eqs. (4)–(6) are made by a separately trained linear verifier. The paper does not report whether this learned verifier preserves the explicit verification capability, e.g., through AUROC/AUPRC or false-positive/false-negative rates. Furthermore, the sensitivity of hallucination mitigation performance to verifier errors is not investigated.

  3. Generation quality and coverage: Although object recall and coverage are reported, RETRACT consistently reduces object recall relative to Vanilla decoding in Table 1. Comparable output lengths alone do not establish preservation of descriptive quality. More comprehensive evaluation of the hallucination–coverage trade-off is needed.

  4. Offline training cost: The proposed approach requires collecting object events, querying the backbone for verification targets, and training a separate verifier for each backbone. Although the training configuration is described in Appendix B.4, the associated computational cost is not quantified. The efficiency comparison focuses mainly on online inference latency, leaving the full computational overhead unclear.

Questions for authors
  1. Effectiveness of the distilled verifier: Could the authors report the learned linear verifier's AUROC/AUPRC and false-positive/false-negative rates on held-out object events, alongside the explicit VLM verification scores? This comparison is essential to establish whether the generation–verification gap motivating RETRACT is actually preserved by the distilled verifier.

  2. Role of KV-cache masking: Why does the variant without KV-cache masking consistently achieve better CHAIR-i and CHAIR-s scores in Table 4? Given that preventing rejected candidates from affecting subsequent generation is a central contribution, how do the authors justify retaining this component?

  3. Comparison with closely related methods: How does RETRACT compare with PSRD and UniProbe, particularly regarding hallucination reduction, coverage preservation, and computational efficiency under comparable training assumptions?

  4. Hallucination–coverage trade-off and training cost: Can the authors provide hallucination–recall trade-off curves or comparisons at matched object recall, together with additional caption-quality metrics? Additionally, what are the total offline verification-query and verifier-training costs for each backbone?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.