Generate, but Verify: Reducing Hallucination in Vision-Language Models with Retrospective Resampling

Thank you for acknowledging the contributions of our work and for the valuable feedback. We provide a few clarifications on the raised weakness below:

[W1: Justification on REVRESE’s Novelty]

While we appreciate the reviewer’s related work suggestions, we would like to clarify that REVERSE is methodologically distinct from [1–5]. Specifically, [1, 2, 4] perform multi-stage verification or refinement after full text generation, relying on separate modules such as external VLMs or object detectors. [3] applies self-correction through multi-round conversations only after the initial generation. [5] focuses on on-the-fly hallucination detection but operates purely in the LLM space without any correction mechanism.

In contrast, REVERSE unifies generation, verification, and correction within a single VLM, enabling on-the-fly correction during decoding without completing the initial output. Moreover, REVERSE introduces explicit confidence tokens, which provide interpretable, token-level self-correction, which is an approach not explored in prior work.

This unified generative-verification framework, highlighted as novel and well-motivated by reviewers PVJe, oB2j, and XTB2, leads to SOTA performance across multiple VLM hallucination benchmarks (Section 4), demonstrating clear advantages over [1–5].

[W2: Breakdown Analyses on Discriminative Hallucination Tasks]

Thank you for the suggestion. We would like to clarify that POPE is a benchmark that focuses exclusively on yes/no questions in a discriminative setting. As for “OPOPE,” we were unable to identify this benchmark in the literature and would appreciate clarification from the reviewer on its definition or source.

As noted in Lines 225–229, our paper primarily focuses on generative hallucination tasks (for example, image captioning), where the space of possible outputs is large and REVERSE's backtracking and self-correction are most beneficial. For discriminative tasks, where answers are typically binary, we have already reported aggregate results in Appendix E. Below, we now provide the official task-wise breakdown for POPE (average F1 score and 95% CI for all sub-categories), which shows only minor differences between REVERSE and the baseline LLaVA-v1.5-7B. While we include this breakdown for completeness, we emphasize that these results are less indicative of REVERSE's core contributions, which target open-ended generation rather than binary classification.

[W3: Additional Results on General LVLM Benchmarks]

As also noted by the reviewer, we include results for POPE, AMBER-D, and MME-Hall in Appendix E (Table A.1). We now additionally report results on GQA and MM-Vet, where REVERSE is comparable to LLaVA-v1.5-7B. These findings confirm that the improvements in hallucination reduction do not come at the expense of performance on standard LVLM benchmarks. We will include these additional results and discussions in the final revision. However, we would like to re-emphasize that our primary focus is on open-ended VQA and generative hallucination tasks. For these, we already report results on five generative benchmarks in the main paper and three discriminative benchmarks in the appendix, evaluated across three different VLM backbones, which we believe provide a thorough evaluation of our approach.

[W4: Discussion on Inference Overhead]

Thank you for raising this important point. We agree that efficiency is critical and have already discussed it in Lines 297–307 of the main paper. In addition, below we provide a more detailed analysis, including the number of correction rounds, hallucination score (CHAIR), and relative runtime (measured by total generated tokens of REVERSE-v1.5-7B compared to LLaVA-v1.5-7B on the AMBER-G task):

The overhead above can be separated into two primary categories, verification overhead and correction overhead:

Verification overhead. Our token-level confidence check is performed inline during generation and adds negligible runtime cost. By contrast, prior methods such as Woodpecker and LURE require additional calls to external VLMs or object detectors, which introduce substantial inference overhead.

Correction overhead. We further analyze performance across different maximum correction rounds:

1. In prior work, a single correction round roughly doubles runtime (full re-generation), and N correction rounds lead to about (N+1)× runtime.
2. In REVERSE, although we set the maximum correction rounds to 50, 37% of samples require no backtracking, and among the remaining cases, over 50% converge within a single round, keeping the overhead modest. As shown in the table above, additional correction rounds further reduce CHAIR scores while runtime overhead remains at most ~3× on average, since corrections are localized rather than full re-generations.

We will include this table and discussion in the final version. We also view the development of even more efficient correction strategies as a promising direction for future work.

Thank you for acknowledging the contributions of our work and for the valuable feedback. We provide a few clarifications on the raised weakness below:

[W1,2: Details and Discussions on Data Augmentation Procedure]

Our data-generation process is similar to FAVA [1], which uses an LLM to augment spans of the data, however differs in two significant ways - the first is scale, while FAVA is composed of 35K fine-grained hallucinations, REVERSE's dataset is composed of an order of magnitude more hallucinated samples (625K), allowing for large-scale supervised fine-tuning. The second, is that we indicate spans explicitly in the token stream (with CN and UN tokens), which allows the model to identify the spans automatically within the same general transformer architecture, rather than being required to build on an external classification mechanism.

Compared to ZINA [2] (which is also much smaller at only 20K hallucinations), REVERSE's dataset is much more flexible in the types of hallucinations generated. Since ZINA relies on a graph-based parsing system, only specific hallucinations can be generated, instead of the more open-ended set generated by REVERSE.

We appreciate the reviewer’s suggestions and will include the above discussions in the appendix and related work.

[W3: Comprehensive Image Captioning Metrics]

Thank you for the helpful suggestion. To evaluate caption quality, AMBER-G already includes a “coverage” metric that measures how many relevant objects are mentioned in addition to the hallucination metric, CHAIR.

For MSCOCO, we follow your advice and also report CLAIR [R-1], an LLM-as-a-judge metric that has been shown to correlate better with human judgments compared to traditional automatic metrics such as CIDEr and SPICE. CLAIR scores range from 0 to 1, with higher values indicating better caption quality.

These results show that REVERSE achieves a ~30% reduction in object hallucination while maintaining comparable caption quality to the baseline. We will include these discussions in the final revision.

[R-1] CLAIR: Evaluating Image Captions with Large Language Models

[Q1: The selection of VLM backbones]

We selected the VLM backbones based on two main considerations. First, most prior hallucination-mitigation work has been implemented on LLaVA-v1.5, making it the standard baseline for comparison. Second, to ensure relevance to current research trends, we added LLaVA-MORE (LLaVA with Llama-3.1) and Qwen-2.5-VL, which are among the most widely used and actively discussed open-source VLMs in the community. As the community has largely moved away from early models such as MiniGPT (used in [3]) toward these newer backbones, we chose this set to provide a representative and up-to-date view of the field.

[Q2,3: Clarification on Our Data Augmentation Process]

The full data generation pipeline is detailed in Appendix B. In summary, we generate one negative sample per positive example using a combination of rule-based algorithms and LLM prompting:

• Type-aware rules: We first categorize QA pairs by type. For simple cases (e.g., yes/no, numbers), we apply straightforward rules such as flipping answers or sampling distractors.
• LLM-generated negatives: For more complex cases (e.g., open-ended descriptive answers), we use GPT-4o to generate diverse and semantically plausible negatives, following the structured prompts described in Appendix B.
• Quality control: We remove trivial negatives (e.g., those involving only pronouns) and rebalance positive/negative order to prevent positional bias.

As a result, the data used for REVERSE is fundamentally different from that used in existing methods such as FAVA [1] and ZINA [2], both in scale (where REVERSE contains 1.3M samples, compared to 35K in FAVA and 20K in ZINA), as well as diversity (since REVERSE’s training data consists of a wide range of hallucination types across attributes, objects, world entities and scenes that go beyond the graph-based replacement framework in prior work).

Thank you for acknowledging the contributions of our work and for the valuable feedback. We provide a few clarifications on the raised weakness below:

[W1+2, Q1+2 Discussions on the Inference Overhead]

Thank you for raising this important point. We agree that efficiency is critical and have already discussed it in Lines 297–307 of the main paper. In addition, below we provide a more detailed analysis, including the number of correction rounds, hallucination score (CHAIR), and relative runtime (measured by total generated tokens of REVERSE-v1.5-7B compared to LLaVA-v1.5-7B on the AMBER-G task):

The overhead above can be separated into two primary categories, verification overhead and correction overhead:

Verification overhead. Our token-level confidence check is performed inline during generation and adds negligible runtime cost. By contrast, prior methods such as Woodpecker and LURE require additional calls to external VLMs or object detectors, which introduce substantial inference overhead.

Correction overhead. We further analyze performance across different maximum correction rounds:

1. In prior work, a single correction round roughly doubles runtime (full re-generation), and N correction rounds lead to about (N+1)× runtime.
2. In REVERSE, although we set the maximum correction rounds to 50, 37% of samples require no backtracking, and among the remaining cases, over 50% converge within a single round, keeping the overhead modest. As shown in the table above, additional correction rounds further reduce CHAIR scores while runtime overhead remains at most ~3× on average, since corrections are localized rather than full re-generations.

We will include this table and discussion in the final version. We also view the development of even more efficient correction strategies as a promising direction for future work.

[Minor Issue 1: Generalization to Video or Multi-Image QAs]

While the current paper focuses on single-image QA using the LLaVA instruction dataset, the method is not limited to this setting. In principle, it can generalize to multi-image QA, video QA, or even pure LLM-based QA by applying the same data augmentation strategy to the corresponding instruction-tuning data and appropriately adjusting the hyperparameters. We will add, and expand on this discussion in our section on future work.

[Minor Issue 2: Failure cases]

We appreciate the reviewer’s suggestion and will include representative failure cases in the final revision (unfortunately, images cannot be included in the rebuttal due to policy restrictions). We note, however, that such failure cases are often detectable by our method. Unlike approaches without an explicit retry mechanism, REVERSE can route these cases to a fallback refusal response (e.g., “I’m sorry, I can’t understand the question”), which helps mitigate the risk of incorrect outputs. We provide preliminary trials of this mechanism in Appendix F.2 for open-ended QA.

Thank you for acknowledging the contributions of our work and for the valuable feedback. We provide a few clarifications on the raised weakness below:

[W1: Missing references and benchmarks]

We thank the reviewer for pointing out these benchmarks. We would like to clarify that:

1. MM-Hal [3]: Results are already reported in Table 3, where REVERSE achieves up to 10% gains across all VLM backbones.

2. POPE [4] and other discriminative tasks: Results and discussions are provided in Table A.1 and Appendix Section E, where REVERSE is comparable to or better than all reported baselines across different model VLM backbones.

3. Additional benchmarks (requested by the reviewer): We have further evaluated REVERSE on HallusionBench, GQA, and MM-Vet, and the results are as follows:

On general VQA benchmarks such as HallusionBench, GQA, and MM-Vet, REVERSE achieves performance comparable to the LLaVA-v1.5-7B baseline, showing that our gains on hallucination benchmarks (e.g., MM-Hal) do not come at the expense of general VLM capabilities. Besides, while HallusionBench is sometimes treated as a visual hallucination benchmark, it mainly evaluates performance on visually misleading inputs rather than the type of language-prior or bias-driven object hallucination that our work specifically addresses. Thanks for the valuable feedback, and we will add these results and a corresponding discussion to the final version.

[Q2: Diversity of the Generated Synthetic Data]

Our complete data generation pipeline is described in Appendix B. In brief, we generate one negative sample per positive example using a combination of rule-based algorithms and LLM prompting, with a strong focus on diversity and quality:

1. Diverse negative generation. We first classify QA pairs into multiple types. For simple cases (e.g., yes/no or numerical questions), we use rule-based transformations such as flipping answers or generating plausible but incorrect numbers. For complex cases (e.g., long-form, open-ended answers), we employ GPT-4o with the prompt described in Appendix B to produce varied negative examples that differ in both content and phrasing.

2. Quality control. To further improve diversity and prevent artifacts, we skip trivial edits (e.g., pronoun swaps) and re-distribute examples to balance positive/negative order to ensure that the negative examples are non-trivial and informative.

These steps ensure that our training data covers a wide range of error types while maintaining balance. We will make this clearer in the final version.

[Q3: Sensitivity of Threshold Tuning]

While the threshold could potentially raise generalization concerns, we would like to clarify that:

1. A single universal threshold works well across domains. For example, a threshold of 0.003 enables REVERSE with LLaVA-series models to achieve SOTA results on AMBER-G, MSCOCO-CHAIR, MM-Hal, and HaloQuest. Similarly, for Qwen-2.5-VL, a threshold of 0.01 yields SOTA performance across these diverse tasks.

2. The threshold provides interpretable control, which is unique to our method. As shown in Figure 5 and Lines 286–296, smaller thresholds (e.g., 0.0001) can significantly reduce hallucination for conservative use cases such as medical QA (even surpassing GPT-4V), while larger thresholds allow greater expressiveness for creative tasks like storytelling. To our knowledge, REVERSE is the only approach that exposes this trade-off through a single, user-adjustable parameter; prior methods effectively fix this threshold implicitly without offering user control.

We appreciate the reviewer’s feedback on this and will include this clarification in the final version. We also agree that enabling the model to adaptively select thresholds during decoding is a promising direction for future work and will add this to our limitations section.

[Q4: Detailed Discussions on Inference Overhead due to Iterative Correction]

Thank you for raising this important point. We agree that efficiency is critical and have already discussed it in Lines 297–307 of the main paper. In addition, below we provide a more detailed analysis, including the number of correction rounds, hallucination score (CHAIR), and relative runtime (measured by total generated tokens of REVERSE-v1.5-7B compared to LLaVA-v1.5-7B on the AMBER-G task):

The overhead above can be separated into two primary categories, verification overhead and correction overhead:

Verification overhead. Our token-level confidence check is performed inline during generation and adds negligible runtime cost. By contrast, prior methods such as Woodpecker and LURE require additional calls to external VLMs or object detectors, which introduce substantial inference overhead.

Correction overhead. We further analyze performance across different maximum correction rounds:

1. In prior work, a single correction round roughly doubles runtime (full re-generation), and N correction rounds lead to about (N+1)× runtime.
2. In REVERSE, although we set the maximum correction rounds to 50, 37% of samples require no backtracking, and among the remaining cases, over 50% converge within a single round, keeping the overhead modest. As shown in the table above, additional correction rounds further reduce CHAIR scores while runtime overhead remains at most ~3× on average, since corrections are localized rather than full re-generations.

We will include this table and discussion in the final version. We also view the development of even more efficient correction strategies as a promising direction for future work.

[W2: Limitations]

Based on the clarification above, we will include further discussion on sensitivity of threshold tuning and potential runtime overhead to the limitation section.