- 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 –, compared with – 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 to , while recall increases from to . 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 to . 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 to . 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.
- 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.
- 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 ? This would help explain the observed reduction in object recall and assess the trade-off between hallucination mitigation and content preservation.