acceptodds
Under review as a conference paper at ICLR 2027

Don't Grade Your Own Exam: Iterative External-Detector Verification and Rewind Decoding for LVLM Hallucination

Abstract

Large vision-language models (LVLMs) often generate fluent responses that are not fully grounded in visual evidence. Most existing hallucination-mitigation methods rely on the generator's own internal representations to revise its predictions, leaving the correction process exposed to the same biases that caused the original error. External vision models offer a complementary source of evidence that can verify generated claims independently of the generator. Building on this idea, we introduce SIEVE, a training-free decoding framework that performs selective, iterative verification during generation. Whenever the model produces a verifiable claim about the image, an external verifier evaluates that claim against the visual evidence. If the claim is rejected, SIEVE rolls generation back to the rejected word, suppresses it at the same decoding position, and resumes from the model's next-best continuation. The regenerated claim is verified again, allowing correction to continue until the output is supported by the evidence rather than simply deleting unsupported content. We instantiate SIEVE for object hallucination, where established benchmarks and reliable external verifiers are readily available, using an object detector as the verifier. Across LLaVA-1.5, InstructBLIP, and Qwen2.5-VL, SIEVE reduces object hallucinations by 42-80% while preserving caption length and fluency.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LVLM, Hallucination, VLM

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_p8t9OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SIEVE, a training-free framework for mitigating object hallucinations in large vision-language models. Unlike approaches relying primarily on the generator's internal information, SIEVE uses an external object detector to verify generated object claims against visual evidence. When a claim is rejected, generation rewinds to the beginning of the rejected word, blocks the corresponding token at that position, and resumes decoding. The regenerated claims are verified iteratively until accepted. The authors further introduce a Rescue–Damage criterion for selecting the external verifier and employ Learn-then-Test (LTT) to calibrate the verification threshold. Experiments across three LVLMs demonstrate substantial reductions in object hallucinations.

Critical strengths

1. Simple and effective framework. SIEVE combines external verification with iterative rewind decoding without requiring additional training. The approach is straightforward and achieves promising results across different LVLMs.

2. Good experimental performance. The method substantially reduces object hallucinations while maintaining reasonable decoding costs. The comparisons and ablation studies provide useful evidence for its effectiveness.

Critical weaknesses

1. Limited novelty and insufficient validation of iterative verification. The main approach combines existing ideas, including external visual verification and iterative correction. Although the combination is interesting, its novelty appears somewhat limited. Moreover, Table 6 examines different post-rejection actions but does not isolate the benefit of iterative verification itself. A direct comparison with one-shot external verification under the same evidence provider and decoding settings would better establish the contribution of the proposed mechanism.

2. Strong dependence on the external Evidence-provider. The effectiveness of SIEVE depends heavily on the accuracy and category coverage of the external object detector. For example, Table 4 reports a Rescue rate of 0.50 and a Damage rate of 0.17 for Faster R-CNN, indicating that some unsupported claims may remain undetected while valid claims may be incorrectly rejected. The robustness of the framework to verifier errors and unsupported object categories is not sufficiently investigated.

Questions for authors

1. Limitations of the external evidence provider. How does SIEVE handle object categories outside the detector's supported vocabulary? Could verifier errors or unsupported categories lead to additional hallucinations during regeneration? A more detailed analysis of these cases would be helpful.

2. Contribution of iterative verification. Could the authors provide a direct comparison between one-shot external verification and iterative rewind decoding using the same evidence provider and decoding settings? This would help clarify the independent contribution of iterative correction.

3. Repeated rejections and termination. What happens when multiple consecutive candidate objects are rejected? Is there a maximum number of rewinds or a fallback mechanism to ensure termination?

4. LTT calibration and hyperparameter selection. Could the authors clarify the calibration sample size, statistical testing procedure, confidence parameter, and candidate threshold selection used in LTT? A sensitivity analysis of the Damage tolerance and calibration settings would also help assess the robustness of the method.

5. The manuscript appears to omit the AI Use Statement and others required by the ICLR 2027 submission guidelines.

Reviewer_a8YnOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper is motivated by the observation that LVLMs are unreliable at detecting their own hallucinations. It therefore introduces an independent external detector to verify object mentions during generation and iteratively rewind and regenerate whenever unsupported content is detected. Experiments on LLaVA-1.5, InstructBLIP, and Qwen2.5-VL show substantial reductions in object hallucination across COCO-500 and COCO-Dense.

Critical strengths

Simple method with clear intervention logic. Once an unsupported object name is detected, the decoding rolls back to the beginning of the word, locally masks the offending span, and regenerates, no model parameter update is required.

Critical weaknesses
  1. The paper addresses object hallucination solely by introducing an additional model. While this is effective from an engineering standpoint, as an academic paper it lacks any analysis of the phenomenon itself, which in my view makes the novelty insufficient.
  2. The reliance on an external model, together with the absence of a description of the detection model, is a real concern.
  3. The Low Recall Performance in Table1. Table1 shows that as the threshold increases, improvements in hallucination metrics come at the cost of a clear drop in Recall. Lower thresholds (e.g., τ=0.3\tau = 0.3) better preserve real objects, but the corresponding advantage on hallucination metrics is limited. The authors therefore need to further demonstrate that these gains are not obtained primarily by sacrificing coverage of real content. We suggest comparing hallucination rates at comparable Recall levels, or presenting the full hallucination-rate--Recall trade-off curve. As in object detection, where precision and recall must be balanced, a lower hallucination rate alone is insufficient to show an improvement in overall description quality.
Questions for authors
  1. Beyond adding an external model, does the paper offer any analysis of why object hallucination occurs?
  2. How to select the detection model which the method rely on, if the model is in a totally unkown environment. Can it be described in sufficient detail?
  3. Given the sharp Recall drop in Table 1, do the hallucination gains still hold at comparable Recall levels (or what does the full trade-off curve look like)?
Reviewer_FkhxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SIEVE, a training-free method for reducing object hallucinations in vision-language models. Instead of relying on the model's own confidence, SIEVE uses an external object detector to check each object mention during generation. When an object is not supported by the image, the method rewinds to that word, blocks it, and generates an alternative. The process continues until the new claim passes verification.

Critical strengths
  1. I like the idea of using an external visual verifier rather than relying only on the VLM's own predictions. The rewind-and-regenerate approach is also intuitive, since it gives the model another chance to produce a supported description instead of simply deleting an unsupported object.

  2. The experimental results are quite strong. The authors evaluate three different VLMs, compare against several existing methods, and report hallucination, recall, caption length, and decoding cost. I also appreciate the ablations showing why external verification and position-specific word blocking are useful.

Critical weaknesses
  1. My main concern is the novelty of the proposed correction strategy and whether the experiments fully isolate its contribution. Previous work such as REVERSE has already explored detecting hallucinations during generation and going back to regenerate the affected text. A more recent work, Verifier-Guided Decoding (VGD), also uses object-level verification, KV-cache rollback, and local regeneration, although it relies on an internal verifier rather than an external detector. I understand that SIEVE uses a different verification mechanism, but the overall correction process seems quite similar. The ablations in Tables 5 and 6 are helpful, but they mainly compare different rejection rules and simple regeneration strategies. I would like to see a stronger comparison with a simpler detector-based correction method using the same Faster R-CNN verifier, so it is clearer how much benefit comes specifically from the iterative rewind strategy.

  2. Another concern is the trade-off between reducing hallucinations and retaining correct visual information. The paper emphasizes that SIEVE preserves caption length and fluency, but this does not necessarily mean it preserves useful content. For example, on LLaVA in Table 1, increasing the threshold to 0.854 reduces CHAIRs from 50.2 to 15.0, but object recall also drops from 80.2 to 65.8. Similar trends appear on InstructBLIP and Qwen. Even at the more moderate threshold of 0.300, recall decreases on all three models. I understand that stronger filtering naturally introduces this trade-off, but I think it deserves more attention. It would be useful to analyze how often the detector incorrectly rejects a valid object and whether the regenerated text actually recovers that information. Also, most of the main evaluation relies on COCO object categories, which are well aligned with the external detector. It remains unclear how well the method would work for objects outside that vocabulary or in a more open-ended setting.

Questions for authors

For the risk-controlled threshold selection, Table 7 shows that the LLaVA COCO configuration with a target risk of 0.10 has an observed evaluation risk of 0.126. I understand that a statistical guarantee does not require every evaluation sample to satisfy the target, but could the authors clarify the calibration sample size, confidence parameter, and how the guarantee should be interpreted in this case?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.