acceptodds
Under review as a conference paper at ICLR 2027

Unlocking What MLLMs Already See for Hallucination Mitigation

Abstract

Hallucinations remain a critical challenge for Multimodal Large Language Models (MLLMs), which often generate statements inconsistent with the input image. In this work, we uncover a : although an MLLM may hallucinate during generation, it can often recognize these hallucinations when its output is decomposed into atomic claims and each claim is independently verified against the image. Building on this observation, we propose a self-supervised framework called (elf-erification uided n-olicy istillation), which turns this self-verification capability into more reliable generation behavior. SVG-OPD first elicits complementary fact-oriented and probing responses, decomposes them into atomic claims, and self-verifies each claim to identify hallucinated and non-hallucinated content. These verified claims are then provided as privileged information to a frozen teacher, whose token-level distributions supervise student-generated trajectories through on-policy distillation. Unlike existing training methods, SVG-OPD requires neither external knowledge nor a stronger teacher model, instead exploiting the MLLM's own visual verification capability for hallucination-aware post-training. Extensive experiments across generative and discriminative hallucination benchmarks, model families, and scales show that SVG-OPD substantially mitigates hallucinations while broadly preserving general multimodal capabilities.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Multimodal Large Language Models, Hallucination Mitigation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_VDtSOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes to reduce the hallucinations in MLLMs by using the same MLLM as the rectifier. The claim is that MLLMs (only Qwen in the paper) have a generation-verification gap, where they may generate hallucinations, but they themselves are a good verifier if the content is a hallucination, given the input image. This property is employed to use the MLLM to generate privileged information of 'facts' and 'hallucinations' about the image. Later, a student version is trained to that is guided by the teacher (having access to the privileged info) using on-policy distillation. Experiments are conducted mainly with Qwen-VL-8B-Instruct, and cross-backbone experiments also include 4B and 7B variants of Qwen. Generally strong results are reported over SOTA relevant methods.

Critical strengths

A well written paper, that is clear about the problem (which in itself is an important one) and builds towards the method systematically by first showing empirical evidence behind the motivation.

The results (although with Qwen only) are strong across multiple metrics.

Critical weaknesses

The central idea of the paper is not highly convincing. The work is build on the premise that MLLMs hallucinate when they generate, but not as much when they verify. Whereas this promises some performance gain by reusing the base model for verification, it still remains a model that 'hallucinates'. So, verification by a hallucinating model is not ideal. Not just that, the method seems to keep using the model for other steps, e.g., creating atomic claims. This is not convincing. The only gain achieved here is that instead of using external module, we reuse the model itself - no matter how good/bad it is. The results are only on Qwen, which seems to be good at creating atomic claims and fitting in the proposed overall framework. In general, this might not be true.

Whereas the paper is well-written, it suffers a bit from over-sophistications. Eg, the teacher MLLM seems to be the same as the reference MLLM and most of the process is just prompting the model (creating fact/probe view, atomic claim creation and even the guidance signal estimation). Different symbols for reference and teacher model do not help understanding that, nor the provided description of the approach. The main technical contribution is in training the student using an existing concept.

Ultimately, the results are provide for Qwen variants only. That does not convince a reader of the generalizabiity of the main idea of the paper that MLLMs can themselves be used to curb hallucinations - Qwen, may be, but MLLMs in general; that is not well supported.

Questions for authors
  • How do the authors verify that the created sets of 'hallucinated' and 'non-hallucinated' concepts are correct?
  • It seems to the reviewer that eventually the outputs are generated using only the student model, while the reference/teacher model is no longer needed. That is, for deployment, only student model is sufficient. Is that understanding correct?
  • Can authors provide details of training the student model. That is, what computational resources, training data, etc. is required?
  • Have the authors tried the same idea with other MLLMs, eg LLaVa?
  • Have authors tried their approach on more recent generative hallucination inducing images like the ones generated by GHOST (ICLR'26)?
3 comments

We greatly appreciate your recognition of our paper's well-written, the importance of the problem we address, and the strong results across multiple metrics.

W1. On the Reliability and Effectiveness of Self-Verification.

The reliability of self-verification is directly supported by the analysis in Figure 1 of the manuscript, where ground-truth annotations from AMBER-Gen, supplemented by manual labeling, show that the model correctly identifies 90.2% of hallucinated and 97.4% of non-hallucinated atomic claims.

To further assess whether these claims provide useful supervision, following Reviewer mySS, we directly evaluate the teacher conditioned on self-verified evidence. The teacher improves all eight metrics over vanilla, further supporting the effectiveness of self-verified atomic claims.

SettingCHAIR↓HalRate↓Cog↓Obj Resp.↓Obj Ment.↓AMBER-Dis Acc↑F1↑MMStar↑
Vanilla (Qwen3-VL-8B-Instruct)7.761.23.818.88.987.390.362.0
Teacher5.435.71.510.35.588.391.262.9

W2. The over-sophistications in defining teacher and reference models.

We thank you for this helpful suggestion. As you pointed out, the teacher model and the reference model share identical model weights. To avoid potential confusion, we have followed your suggestion and consistently used πT\pi_T to denote both models throughout the revised manuscript.

W3. Evaluate the idea on a non-Qwen backbone.

To further evaluate the generalizability of our SVG-OPD, we extended our experiments to LLaVA-OneVision-1.5-8B-Instruct. Encouragingly, SVG-OPD consistently improves performance across all evaluated metrics, in line with our findings on Qwen-VL series. We have included these results in Table 3 of the revised manuscript.

BackboneModelCHAIR↓HalRate↓Cog↓Obj Resp.↓Obj Ment.↓AMBER-Dis Acc↑F1↑MMStar↑
LLaVA-OneVision-1.5-8B-InstructVanilla7.562.46.621.311.083.588.366.0
LLaVA-OneVision-1.5-8B-InstructSVG-OPD5.836.84.18.94.286.790.466.6
Qwen3-VL-4B-InstructVanilla8.061.63.721.310.289.892.259.5
Qwen3-VL-4B-InstructSVG-OPD5.941.31.914.87.290.192.659.6
Qwen2.5-VL-7B-InstructVanilla5.026.21.811.36.086.590.360.9
Qwen2.5-VL-7B-InstructSVG-OPD4.323.21.58.74.587.791.261.3
Qwen3-VL-8B-InstructVanilla7.761.23.818.88.987.390.362.0
Qwen3-VL-8B-InstructSVG-OPD5.741.21.711.65.588.591.362.3

Q1. How do the authors verify that the created sets of 'hallucinated' and 'non-hallucinated' concepts are correct? Please see W1 above.

Q2. Is the student model alone sufficient for deployment, while the reference/teacher model is no longer needed?

Yes. After distillation, the student internalizes the verification-guided knowledge and performs inference independently, without requiring the reference model, teacher model, or additional verification steps. Therefore, SVG-OPD introduces no additional inference-time overhead.

Q3. Can authors provide details of training the student model. That is, what computational resources, training data, etc. is required?

We follow the experimental setup of RLSF-V (ICML 2026) [1] and train on 10K unlabeled image-instruction pairs from the RLAIF-V (CVPR 2025) [2] dataset. Our method is trained on 8 NVIDIA H100 80GB GPUs, with training taking approximately 4 hours for Qwen3-VL-8B. Please refer to the Implementation Details paragraph in Section 4.1 of the manuscript for further information.

Q4. Have the authors tried the same idea with other MLLMs? Please see W3 above.

Q5. Evaluation on GHOST images.

Following your suggestion, we additionally evaluate our SVG-OPD on the recent GHOST (ICLR 2026)[3] benchmark, which uses adversarially generated images to induce object hallucinations. The results are as follows:

ModelHallucination Rate ↓
Qwen3-VL-8B-Instruct53.37
+SVG-OPD49.62
↳ Improvement+ 3.75
LLaVA-OneVision-1.5-8B-Instruct38.94
+SVG-OPD35.90
↳ Improvement+ 3.04

Our SVG-OPD reduces hallucination rates by 3.75 and 3.04 percentage points, respectively, demonstrating its effectiveness on the adversarially generated images.

[1] RLSF-V: Mitigating Hallucinations in MLLMs via Fuzzy Semantic Self-Feedback (ICML 2026)

[2] Rlaif-v: Open-source ai feedback leads to super gpt-4v trustworthiness (CVPR 2025)

[3] GHOST: Hallucination-Inducing Image Generation for Multimodal LLMs (ICLR 2026)

Thank you for your thoughtful responses and further evaluations. Whereas the reviewer still maintains that a hallucinating model rectifying itself with self-verification is not ideal in practice, as the model itself becomes the bottleneck, author responses generally address the reviewer concerns. Hence, the reviewer is willing to update the rating to weak accept.

We sincerely thank you for taking the time to review our response. Your insightful suggestions have played a pivotal role in enhancing the overall quality of our paper. We deeply appreciate your recognition of our efforts to address the concerns and your willingness to update the rating once again.

Reviewer_PXm2OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies visual hallucinations in Multimodal Large Language Models (MLLMs). The authors identify a generation-verification gap: although MLLMs may hallucinate during generation, they can often recognize these hallucinations when their responses are decomposed into atomic visual claims and verified individually against the image. Based on this observation, the authors propose SVG-OPD, which elicits image descriptions through two complementary prompting strategies, decomposes them into atomic claims, and verifies each claim to construct positive and negative evidence. The method then performs on-policy distillation, using a frozen version of the original MLLM conditioned on the verified evidence as the teacher and a trainable version as the student. This process aims to reduce hallucinations without relying on external supervision.

Experiments on several Qwen-VL models demonstrate that SVG-OPD improves performance on multimodal hallucination benchmarks, particularly on open-ended generation tasks, while largely preserving general multimodal capabilities. The authors also conduct ablation studies to investigate the contributions of different design components.

Critical strengths
  • The proposed method is intuitive and reasonably straightforward. As a training-based approach, it incorporates self-verification signals into the generation policy without introducing additional inference-time overhead.

  • SVG-OPD achieves consistent improvements across multiple hallucination benchmarks, while largely preserving general multimodal capabilities.

Critical weaknesses

I have some concerns about the novelty of the proposed generation-verification gap, which is highlighted as one of the main contributions. The underlying observation has already been exploited to some extent in prior work, such as TACO (EACL 2026), which also leverages atomic-level verification to identify hallucinations in MLLM-generated responses. Although this paper provides a more explicit quantitative characterization of the gap, the main novelty appears to lie in how the self-verification results are utilized for training, rather than in the observation itself.

Questions for authors
  • Evidence construction: The current method constructs positive evidence exclusively from the fact-oriented view and negative evidence exclusively from the probing view. However, hallucinated claims may also appear in the fact-oriented view, and valid claims may appear in the probing view. It seems reasonable to include these claims as well, provided they pass the corresponding verification criteria. Have the authors experimented with using both positive and negative claims from both views?

  • Treatment of uncertain claims: The verifier classifies claims as entailed, contradicted, or unknown, but only contradicted claims are used as negative evidence, while unknown claims are discarded. Since claims that cannot be verified from the image may also be considered hallucinations, could the authors explain the rationale behind excluding unknown claims? Have they investigated whether including them as negative evidence affects performance?

  • Backbone diversity: Although the method is evaluated on three backbones, all of them belong to the Qwen-VL family. Given that SVG-OPD relies on the model's own visual verification capability, have the authors experimented with other MLLM families to further validate its generalizability?

3 comments

We greatly appreciate your recognition of the intuitive and reasonably straightforward nature of our method, as well as its consistent improvements across multiple hallucination benchmarks.

W1. Novelty of our method compared with TACO.

We thank the reviewer for highlighting TACO (EACL 2026) [1], which also identifies the generation–verification gap. We have added a citation to this concurrent work in the revised paper. The key distinction lies in how verification is leveraged. TACO uses atomic verification and confidence aggregation to refine outputs at inference time. In contrast, we transfer the model's verification capability into its generation policy through self-supervised training, enabling verification-free inference. Furthermore, we explicitly construct both positive and negative atomic evidence as complementary training signals, which is absent from TACO.

Q1. Evidence construction.

Following your suggestion, we investigate the alternative of using all self-verified positive and negative claims from both the fact and probing views (hereafter referred to as the both-views variant) on Qwen3-VL-8B-Instruct:

MethodCHAIR↓HalRate↓Cog↓Obj Resp.↓Obj Ment.↓AMBER-Dis Acc↑F1↑MMStar↑
default5.741.21.711.65.588.591.362.3
both-views7.356.52.913.16.388.391.162.2
↳ Gain vs. default-1.6-15.3-1.2-1.5-0.8-0.2-0.2-0.1

Our default strategy outperforms the both-views variant across all metrics, demonstrating the effectiveness of view-specific evidence selection over combining all verified claims.

Q2. Treatment of uncertain claims.

Following your suggestion, we conduct the ablation in which unknown claims are also incorporated into the negative evidence set. The results are summarized below.

MethodCHAIR↓HalRate↓Cog↓Obj Resp.↓Obj Ment.↓AMBER-Dis Acc↑F1↑MMStar↑
Ours (default)5.741.21.711.65.588.591.362.3
Unknown as Negative6.943.42.213.76.487.590.862.3

As shown in the table above, treating unknown claims as negative evidence is less effective than our default approach. These results suggest that the reliability of negative evidence is more important than simply increasing its coverage. We therefore exclude unknown claims to avoid introducing potentially misleading supervision into the distillation process.

Q3. Backbone diversity.

To further evaluate the generalizability of our SVG-OPD, we extended our experiments to LLaVA-OneVision-1.5-8B-Instruct. Encouragingly, SVG-OPD consistently improves performance across all evaluated metrics, in line with our findings on Qwen-VL series. We have included these results in Table 3 of the revised manuscript.

BackboneModelCHAIR↓HalRate↓Cog↓Obj Resp.↓Obj Ment.↓AMBER-Dis Acc↑F1↑MMStar↑
LLaVA-OneVision-1.5-8B-InstructVanilla7.562.46.621.311.083.588.366.0
LLaVA-OneVision-1.5-8B-InstructSVG-OPD5.836.84.18.94.286.790.466.6
Qwen3-VL-4B-InstructVanilla8.061.63.721.310.289.892.259.5
Qwen3-VL-4B-InstructSVG-OPD5.941.31.914.87.290.192.659.6
Qwen2.5-VL-7B-InstructVanilla5.026.21.811.36.086.590.360.9
Qwen2.5-VL-7B-InstructSVG-OPD4.323.21.58.74.587.791.261.3
Qwen3-VL-8B-InstructVanilla7.761.23.818.88.987.390.362.0
Qwen3-VL-8B-InstructSVG-OPD5.741.21.711.65.588.591.362.3

[1] Taming Object Hallucinations with Verified Atomic Confidence Estimation (EACL 2026)

Thank you for the detailed response and additional experiments. My experimental concerns have been adequately addressed. I believe my rating of Weak Accept best reflects my overall assessment of the paper now.

Thank you for your support and insightful suggestions. We are happy to hear that all your concerns have been addressed. Your feedback helped us improve the clarity and intuitive explanation of our method, ultimately strengthening the work. We sincerely thank you for taking the time to review our response once again.

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

This paper proposes SVG-OPD, a self-supervised method for reducing MLLM hallucinations. The main idea is that the model can often detect its own hallucinated claims when they are checked separately against the image.

The method verifies atomic claims generated by the model and uses the verified positive and negative claims as guidance for a frozen teacher. The student is then trained using on-policy distillation and does not need this extra information at inference time.

Experiments on several hallucination and general multimodal benchmarks show reduced hallucination while mostly preserving overall model performance.

Critical strengths
  1. The main idea is interesting and practical. The method uses the model’s own ability to verify its responses and does not need human labels or a stronger external model. This makes the approach more self-contained.

  2. The results are generally strong on different hallucination benchmarks, while the general multimodal performance is mostly preserved. The paper also includes several useful ablations, which helps support the effectiveness of the proposed components.

Critical weaknesses
  1. The novelty is not fully convincing to me. Several parts of the method, including self-verification, atomic claim checking, and distillation from the model’s own feedback, are related to previous work. The main new part seems to be how these components are combined, so the paper should position this more clearly.

  2. Some of the ablations also weaken the main story. Most of the improvement seems to come from the verified positive claims, while the hallucinated negative claims add only a smaller gain. Also, off-policy distillation performs very close to the proposed on-policy version. This makes it less clear which parts of the proposed method are really necessary.

Questions for authors

Can you clarify the novelty compared with previous self-verification and self-feedback methods, especially works that already use atomic claims or distill the model’s own feedback? What is the main technical difference of SVG-OPD?

From Table 4, most of the improvement seems to come from the positive verified claims (E^+), while the negative hallucinated claims (E^-) provide a smaller gain. Can you explain more clearly what role (E^-) is playing in the final method?

The off-policy distillation result is very close to the proposed on-policy version and is even slightly better on one Object HalBench metric. Can you clarify why on-policy distillation is important here?

Can you report the teacher performance when it directly uses the verified evidence? This would help show how much of the teacher’s improvement is actually transferred to the student.

Can you clarify whether the hallucination reduction is partly coming from more conservative outputs? Reporting AMBER Cover, average response length, or number of mentioned objects would help understand this.

Future work suggestions: It would be useful to evaluate the method on a non-Qwen backbone to better support the generalization claim.

2 comments

We sincerely appreciate your recognition of our interesting and practical idea, the self-contained nature of our approach, and its strong performance on different hallucination benchmarks.

W1. Clarify the novelty compared with previous self-verification and self-feedback methods.

We thank you for highlighting the need to clarify our novelty relative to prior self-verification and self-feedback methods. We clarify the key distinctions in the following aspects:

  1. Prior self-verification methods. Following your suggestion, we further investigated the related work and found that the concurrent TACO (EACL 2026) [1] also observes the self-verification phenomenon in MLLMs during reasoning. We have added the corresponding citation.

  2. Distinction from TACO. TACO performs self-verification at inference time to refine responses, introducing substantial inference-time overhead. In contrast, our SVG-OPD internalizes verification signals into the generation policy via self-distillation, enabling direct response generation without additional inference-time overhead. Moreover, our proposed fact and probing views elicit complementary non-hallucinated and hallucinated claims to provide richer distillation evidence, further enhancing performance.

  3. Distinction from prior self-feedback methods. Unlike methods such as RLSF-V (ICML 2026) [2], which derives self-feedback from the model's own uncertainty to construct response-level preferences for DPO, SVG-OPD instead elicits explicit claim-level visual feedback through self-verification against the image. This feedback is provided as privileged context to the teacher, which provides token-level supervision on student-generated trajectories via on-policy distillation

W2. The role of non-hallucinated claims E+E^+ and hallucinated claims E−E^-.

Non-hallucinated claims E+E^+ provide positive guidance on what the model should generate, while hallucinated claims E−E^- provide complementary negative guidance on what the model should avoid generating. Our experiments show that E+E^+ alone can substantially mitigate hallucinations by reinforcing visually grounded content. Building on this, E−E^- further provides targeted guidance against plausible but visually unsupported claims, helping suppress residual hallucinations that positive guidance alone may not fully address.

As shown in the table below, both hallucinated and non-hallucinated claims generally improve performance over vanilla, while their combination yields the best overall results.

ModelCHAIR↓HalRate↓Cog↓Obj Resp.↓Obj Ment.↓AMBER-Dis Acc↑F1↑MMStar↑
Vanilla (Qwen3-VL-8B-Instruct)7.761.23.818.88.987.390.362.0
Both (ours)5.741.21.711.65.588.591.362.3
↳ Gain vs. Vanilla+2.0+20.0+2.1+7.2+3.4+1.2+1.0+0.3
w/o non-hallucinated claims7.757.53.116.37.787.690.661.9
↳ Gain vs. Vanilla+0.0+3.7+0.7+2.5+1.2+0.3+0.3-0.1
w/o hallucinated claims6.245.32.112.05.888.391.062.2
↳ Gain vs. Vanilla+1.5+15.9+1.7+6.8+3.1+1.0+0.7+0.2

W3. The role of on-policy distillation.

While off-policy distillation demonstrates the effectiveness of self-verification guidance, on-policy distillation better aligns teacher supervision with the student's evolving generation distribution, reducing the training–inference mismatch (ICLR 2024) [3], particularly since the student generates independently without teacher access at inference time.

To further investigate this distinction, we evaluated both variants on six general multimodal benchmarks (see Appendix A for dataset descriptions). Experimental results show that on-policy distillation achieves the best performance on five out of six benchmarks and consistently outperforms Vanilla. These findings demonstrate that although off-policy distillation may be effective, on-policy training delivers more consistent improvements, making it our preferred design choice.

MethodRealWorldQA ↑MMMU-val ↑MME ↑MathVista ↑MMBench ↑MM-Vet v2 ↑
Vanilla (Qwen3-VL-8B-Instruct)6̲9̲.̲2̲8̲51.892401.3844.708̲4̲.̲9̲8̲67.49
Off-policy69.155̲2̲.̲1̲1̲2̲4̲0̲5̲.̲5̲4̲46.208̲4̲.̲9̲8̲6̲8̲.̲1̲6̲
On-policy69.4152.332411.244̲6̲.̲1̲0̲85.1468.84

Q1-Q3. Please see W1-W3 above.

Q4. Report the teacher performance when it directly uses the verified evidence.

Following your suggestion, we further evaluate the teacher model when directly conditioned on self-verified atomic evidence, with the results reported in the table below. The student achieves performance comparable to the teacher across benchmarks, despite generating responses directly without explicit atomic evidence extraction at inference time. These results suggest that our method effectively internalizes the model's verification guidance into generation policy, enabling it to generalize to unseen inputs during inference.

SettingCHAIR↓HalRate↓Cog↓Obj Resp.↓Obj Ment.↓AMBER-Dis Acc↑F1↑MMStar↑
Vanilla (Qwen3-VL-8B-Instruct)7.761.23.818.88.987.390.362.0
Teacher5.435.71.510.35.58̲8̲.̲3̲9̲1̲.̲2̲62.9
Student5̲.̲7̲4̲1̲.̲2̲1̲.̲7̲1̲1̲.̲6̲5.588.591.36̲2̲.̲3̲

Notably, the teacher and student differ in their inference-time requirements. While the teacher performs atomic evidence extraction, incurring substantial computational overhead, the student generates responses directly without additional inference-time computation (184 minutes for the teacher versus 21 minutes for the student in the evaluation reported above).

Q5. Whether the hallucination reduction is partly coming from more conservative outputs?

Following your suggestion, we report AMBER Cover (i.e., Recall) and average response length. Although our method generates shorter responses than Vanilla, object recall remains largely preserved.

settingsCover (Recall) ↑Average response length
Vanilla (Qwen2.5-VL-7B)64.683.4
Ours64.274.1

To further investigate whether the reduction in hallucinations stems from more conservative generation, we conducted a controlled analysis using budget-forcing strategy, as shown in Figure 5 of the manuscript, comparing both methods under identical output-token budgets. The results demonstrate that our method consistently achieves a better recall–hallucination trade-off, maintaining comparable while reducing hallucinations. This suggests that the improvement is not merely due to shorter responses but also reflects more reliable visual objects.

Q6. Evaluate the method on a non-Qwen backbone.

To further evaluate the generalizability of our SVG-OPD, we extended our experiments to LLaVA-OneVision-1.5-8B-Instruct. Encouragingly, SVG-OPD consistently improves performance across all evaluated metrics, in line with our findings on Qwen-VL series. We have included these results in Table 3 of the revised manuscript.

BackboneModelCHAIR↓HalRate↓Cog↓Obj Resp.↓Obj Ment.↓AMBER-Dis Acc↑F1↑MMStar↑
LLaVA-OneVision-1.5-8B-InstructVanilla7.562.46.621.311.083.588.366.0
LLaVA-OneVision-1.5-8B-InstructSVG-OPD5.836.84.18.94.286.790.466.6
Qwen3-VL-4B-InstructVanilla8.061.63.721.310.289.892.259.5
Qwen3-VL-4B-InstructSVG-OPD5.941.31.914.87.290.192.659.6
Qwen2.5-VL-7B-InstructVanilla5.026.21.811.36.086.590.360.9
Qwen2.5-VL-7B-InstructSVG-OPD4.323.21.58.74.587.791.261.3
Qwen3-VL-8B-InstructVanilla7.761.23.818.88.987.390.362.0
Qwen3-VL-8B-InstructSVG-OPD5.741.21.711.65.588.591.362.3

[1] Taming Object Hallucinations with Verified Atomic Confidence Estimation (EACL 2026)

[2] RLSF-V: Mitigating Hallucinations in MLLMs via Fuzzy Semantic Self-Feedback (ICML 2026)

[3] On-policy distillation of language models: Learning from self-generated mistakes (ICLR 2024)

Comments on the paper

1 comment

Dear Reviewers, AC, SAC, and PC,

We sincerely thank all reviewers for their insightful comments and constructive suggestions, as well as for recognizing the interesting and practical idea (Reviewer mySS), intuitive method and inference efficiency (Reviewer PXm2), well-written paper and the importance of the problem we address (Reviewer VDtS), and strong hallucination mitigation performance (Reviewers mySS, PXm2, VDtS).

During the rebuttal, we carefully addressed the reviewers' concerns and strengthened our manuscript with the following updates:

  1. Cross-Backbone Generalization: Added more MLLM families experiments, demonstrating consistent improvements beyond Qwen-VL (Table 3).
  2. Additional Challenging Benchmark: Evaluated on GHOST (ICLR 2026) benchmark, confirming effectiveness on adversarially generated hallucination-inducing images (Table 11).
  3. Teacher–Student Analysis: Compared the performance of the evidence-conditioned teacher and the distilled student, demonstrating that the student effectively internalizes self-verification guidance while avoiding additional inference-time overhead (Table 9).
  4. Hallucination–Coverage Analysis: Added object coverage, response-length statistics, and controlled output-budget experiments, demonstrating that our SVG-OPD achieves a favorable recall–hallucination trade-off (Figure 5, Table 10).
  5. Expanded Ablations: Examined alternative evidence construction strategies, the treatment of uncertain claims, and on-policy versus off-policy distillation, providing additional justification for our methodological choices (Tables 7–8, Appendix C.3–C.4).
  6. Related Work and Clarifications: Following the reviewers' suggestions, added and discussed RLSF-V (ICML 2026) and concurrent work TACO (EACL 2026), further clarifying their relationship to our approach (Section 2).
  7. Point-by-Point Responses: Provided comprehensive responses to specific reviewer comments.

We highlight them in blue in the latest revision. As the ICLR public discussion phase is ending soon, we kindly encourage you to share any feedback on our submission while there's still time. We'd be happy to address any concerns or provide clarifications to assist with your evaluation. Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.