acceptodds
Under review as a conference paper at ICLR 2027

Seeing What Matters: Visual-Selective On-Policy Self-Distillation for VLM Hallucination Mitigation

12 views

Abstract

Vision-language models (VLMs) still suffer from object hallucination in open-ended generation. They often mention objects, attributes, or relations that are not grounded in the image. Preference-based methods such as DPO can reduce hallucinations, but open-ended caption preferences are expensive to collect and often entangle visual faithfulness with style, length, and level of detail. Meanwhile, directly turning object cues into hard pseudo captions can amplify the teacher's linguistic bias and hallucinations. We study how to extend on-policy self-distillation from verifiable language tasks to open-ended multimodal hallucination reduction. Instead of using privileged hints that directly verify an answer, we use training-time object cues to construct a non-verifying privileged visual teacher and distill its soft token distributions on responses sampled by the student itself. We further propose Visual-Selective On-Policy Self-Distillation (VS-OPSD). By contrasting the privileged teacher under the real image and a weak-image condition, VS-OPSD identifies visually sensitive tokens and applies distillation only at those positions. Experiments across two model families and two LLaVA model scales show that VS-OPSD reduces hallucination without preference pairs while largely preserving general multimodal capability. Controlled ablations attribute the gains to concentrating supervision on visually sensitive tokens rather than simply reducing the number of distilled tokens.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Vision-Language Models, Multimodal Hallucination, On-Policy Self-Distillation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_w4PoOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Visual-Selective On-Policy Self-Distillation (VS-OPSD) to mitigate object hallucination in vision-language models. A frozen teacher receives training-time object cues and provides soft supervision on student-generated responses. The method compares teacher predictions under original and degraded images to identify visually sensitive token positions, then selectively applies sensitivity-weighted distillation. Experiments on LLaVA and Qwen models report reduced hallucination while largely preserving general multimodal capabilities, without additional inference-time computation.

Critical strengths
  1. Practical training design. VS-OPSD uses object-category cues rather than preference pairs or complete reference captions, and adds no inference-time computation. This makes it a potentially accessible approach to hallucination mitigation.

  2. Consistent improvements over Vanilla OPSD. The method reduces AMBER hallucination rates across three evaluated backbones, with improvements also observed across three paired random seeds on LLaVA-7B. These results provide evidence that the proposed supervision weighting has practical value despite the limited methodological novelty.

Critical weaknesses
  1. Limited methodological novelty. The method largely builds on existing OPSD and visual-dependence-based distillation. In particular, VA-OPD already uses differences between teacher predictions under original and degraded images to allocate supervision. The main additions are object-cue-conditioned self-distillation, a distribution-level Visual-KL score, and sensitivity-weighted token selection for open-ended hallucination mitigation. While potentially useful, please compare it with your VS-OPSD.
  2. The hallucination–coverage trade-off is insufficiently characterized. On LLaVA-7B, VS-OPSD reports lower Cover than all other mitigation methods listed in Table 1, although several external references use different settings. This raises concerns that part of the hallucination reduction comes from more conservative content selection. Appendix B.2 provides length-normalized statistics, but these do not rule out selectively omitting uncertain objects or details.
  3. The necessity of the core design choices is not sufficiently demonstrated. The ablations omit a straightforward baseline that retains all tokens and weights them by Visual-KL. Consequently, it remains unclear whether top-30% selection is necessary or whether continuous weighting alone would achieve comparable results.
Questions for authors

How does the method advance beyond OPSD and VA-OPD? Does the hallucination reduction persist when methods are compared at matched coverage? How does VS-OPSD compare with weighting all tokens by Visual-KL without top-30% selection?

2 comments
AuthorsOpenReview ↗

Thank you for the detailed review! Here is our clarification of the differences from OPSD and VA-OPD. (For your Q1)

We thank the reviewer for the detailed review and for recognizing the practical training design, the absence of inference-time computation, and the consistent improvements across backbones and random seeds. Below we address the three concerns in turn.

Q1 & W1: Novelty relative to OPSD and VA-OPD. Our work and the concurrent work VA-OPD both recognize that a controlled visual intervention on the teacher can identify which positions depend on the image, and can therefore guide how supervision is allocated. We discuss this shared insight in Related Work (Section 5) and provide a controlled comparison in Appendix E.1. Below we clarify the differences between the two works in four aspects.

(1) Research setting. OPSD mainly performs self-distillation on verifiable reasoning tasks using reference solutions, and VA-OPD mainly studies transferring visual understanding and reasoning from an external teacher to a student. We instead target open-ended object hallucination mitigation and build a privileged self-teacher from non-verifying object cues. This distinction is described in the Introduction and Related Work.

(2) Visual-sensitivity measure. VA-OPD measures whether visual information increases the teacher’s confidence in the generated token, and rectifies negative log-probability gains to zero. In hallucination, however, the real image may lower the probability of an incorrect token, so such negative gains should not be discarded as uninformative. Visual-KL measures the change in the candidate distribution via KL divergence. It accounts for the probability changes of all candidates, not only the sampled token, and can therefore identify positions that are sensitive to the visual intervention but whose VA is rectified to zero. This design is simple yet effective.

On 200 fixed rollouts, we score every token with both measures using the same teacher, weak image, and top-30% token budget.

Hallucinated object mentions (num) Visual-KL       VA
Covered (218 in total)76.6% (167)57.3% (125)
Missed among the 38 with VA = 016/3838/38

Visual-KL covers more hallucinated object mentions than VA (76.6% vs. 57.3%). More importantly, among the 38 mentions that VA completely misses (VA = 0), Visual-KL misses only 16, recovering 22 of them (57.9%).

(3) Rollout-level weighting. VA-OPD normalizes the mean visual advantage across multiple sibling rollouts of the same input to determine response-level distillation weights, with K = 4 by default. VS-OPSD performs sensitivity estimation, token selection, and weighting within a single rollout, without comparisons across responses.

(4) Token-level supervision allocation. VA-OPD retains the distillation loss on both the high- and low-advantage groups, whereas VS-OPSD applies a sensitivity-weighted loss only at the selected high-sensitivity positions. Since a small fraction of tokens carries most of the Visual-KL mass, direct distillation is not needed at every position (see our response to Q3).

Finally, the comparison in Appendix E.1 further shows that, in our open-ended object hallucination setting, VS-OPSD outperforms our reproduced VA-OPD under the same experimental protocol.

Our responses to Q2 and Q3 are provided in the next comment.

AuthorsOpenReview ↗

Thank you for the detailed review! We provide results on hallucination at matched coverage and an all-token weighting ablation below. (For your Q2 & Q3)

Q2 & W2: Hallucination–coverage trade-off. We acknowledge that hallucination mitigation involves a trade-off between hallucination rate and coverage. We address this concern from four aspects.

(1) Controlled comparison. In the directly comparable setting (Table 1), VS-OPSD achieves coverage similar to Vanilla OPSD on all three backbones while consistently reducing Hal:

BackboneVanilla OPSD Cover / HalVS-OPSD Cover / Hal
Qwen2.5-VL-3B61.7 / 26.961.0 / 25.0
LLaVA-1.5-7B48.2 / 25.547.7 / 22.4
LLaVA-1.5-13B49.1 / 23.849.0 / 21.1

On LLaVA-1.5-7B and Qwen2.5-VL-3B, VS-OPSD also exceeds the coverage of the base model (47.7 vs. 46.6 and 61.0 vs. 59.2).

(2) Comparison with SFT. SFT learns token by token from complete reference captions and thus produces more detailed outputs with higher coverage (49.6 on LLaVA-1.5-7B). However, because this learning is off-policy, its Hal even rises above the base model (31.4 vs. 29.8), which defeats the purpose of hallucination mitigation.

(3) Matched coverage experiments. To compare the methods at matched coverage, we prompted the models to produce descriptions at different levels of detail and measured hallucination at each level. Results are averaged over two training seeds (42 and 43) at step 100 on LLaVA-1.5-7B, with generation capped at 512 tokens with no truncation.

PromptVanilla OPSD Cover ↑ / Hal ↓VS-OPSD Cover ↑ / Hal ↓
Brief34.55 / 5.9034.15 / 5.75
Standard48.35 / 25.3548.20 / 24.30
Detailed49.55 / 25.6549.60 / 23.00
Comprehensive50.45 / 36.5050.60 / 34.55

VS-OPSD achieves coverage comparable to Vanilla OPSD at every level, and even slightly higher at the Detailed and Comprehensive levels, while its hallucination rate is consistently lower, with the largest reductions at these two levels (2.65 and 1.95). The hallucination reduction therefore cannot be explained by more conservative descriptions.

(4) Output quality. The general-capability results in Table 1 show that our gains are not obtained by sacrificing output quality. On LLaVA-1.5-7B, VS-OPSD largely preserves MME (1735.34 vs. 1757.46 for the base model) and ScienceQA (63.38 vs. 63.66), consistent with improved visual grounding rather than degraded generation. By contrast, DPO (300 steps) attains a high coverage of 55.0, but its general capability drops severely (MME 1348.46, ScienceQA 62.61), which shows that coverage alone does not fully reflect output quality.

Q3 & W3: Weighting all tokens by Visual-KL without top-30% selection. We thank the reviewer for suggesting this. Table 2(a) in the paper already separates the effects of selection and weighting within the selected set. We further added the all-token weighting variant suggested as a supplementary ablation, adopting the same experimental protocol as Table 1 in the paper:

MethodSelectionWeightingCHAIR ↓Cover ↑Hal ↓Cog ↓
Vanilla OPSDAllUniform5.748.225.52.6
Top-30% uniformTop 30%Uniform5.747.324.02.4
All-token weightedAllVisual-KL5.447.823.02.4
VS-OPSDTop 30%Visual-KL5.047.722.42.3

Continuous weighting alone already reduces Hal from 25.5 to 23.0. Adding top-30% selection further reduces Hal to 22.4 and CHAIR from 5.4 to 5.0 at nearly identical coverage (47.7 vs. 47.8), yielding the best results on CHAIR, Hal, and Cog. Selection alone also reduces Hal (24.0). The two components are therefore complementary: selection determines where to distill, and weighting determines how strongly to distill at each position. Top-30% selection is thus beneficial rather than redundant, and its gain does not come at the cost of coverage.

Ethics flag. We believe our work does not raise specific ethical concerns. As described in the Ethics Statement of the paper, this work uses only publicly available models and datasets and does not involve human participants, private data, or personally identifiable information. We would be happy to address any specific concern the reviewer may have.

We hope these results address the reviewer's concerns, and we would be grateful if the reviewer would reconsider the score in light of them. Please feel free to let us know if there are any further questions.

Reviewer_u1sHOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

To reduce hallucinations in VLMs, the authors propose Visual-Selective On-Policy Self-Distillation (VS-OPSD). The method compares a privileged teacher’s predictions under the original image and a weak-image condition, where the image is heavily blurred to suppress fine-grained visual information.

The student first generates an on-policy response. The teacher then evaluates the same response trajectory using both the original image and the blurred image. The difference between the two next-token distributions is measured using Visual-KL, which serves as a measure of how visually sensitive each token is.

During training, the method keeps only the top 30% most visually sensitive tokens and weights the distillation loss according to their Visual-KL scores. In this way, supervision is concentrated on tokens that depend most strongly on visual evidence rather than being applied uniformly across the entire response.

The method is mainly designed to reduce open-ended object hallucinations, especially incorrect claims about whether objects are present in the image.

Critical strengths

Well-motivated heuristic for reallocating self-distillation efforts towards visually dependent tokens. Clearly motivated extension from previous works. Comprehensive experimental results.

Critical weaknesses

The degraded-image or token selection scheme overlaps with VA-OPD but authors are not hiding that overlap and articulated the difference and experimentally compared against it.

Questions for authors

NA

1 comment
AuthorsOpenReview ↗

Thanks for your valuable review! We provide a supplementary clarification and additional visualizations below for a more complete picture of our work.

We sincerely thank the reviewer for recognizing the motivation of our work and the comprehensiveness of our experiments, and for the time and effort devoted to reviewing our paper. We provide a brief clarification and additional visualizations below for a more complete picture of our work.

Supplementary clarification. The key difference from VA-OPD lies in how visual sensitivity is measured. VA scores only the sampled token and rectifies negative gains to zero. In hallucination, however, the real image may lower the probability of an incorrect token, so such positions receive VA = 0 and are treated as visually insensitive. Visual-KL instead measures the change in the full candidate distribution, and can therefore still identify these positions. To verify this, we scored every token of 200 fixed rollouts with both measures, using the same teacher, weak image, and top-30% budget:

Hallucinated object mentions (num) Visual-KL       VA
Covered (218 in total)76.6% (167)57.3% (125)
Missed among the 38 with VA = 016/3838/38

Visual-KL covers substantially more hallucinated object mentions than VA, and recovers 22 of the 38 mentions that VA rectifies to zero. This distribution-level design is simple, yet it captures visually sensitive hallucinations that a sampled-token score misses.

Additional visualizations. We have added Appendix G (Figures 9–10), which shows complete responses for complex multi-object and spatial-relation scenes, highlighting the selected tokens and the corresponding corrections.

We thank the reviewer again. And we are happy to answer any further questions.

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

The authors proposed the Visually-Selective On-Policy Self-distillation (VS-OPSD) for open-ended image captioning tasks. (from abstract) Authors applied object cues to construct privileged visual teacher. (from contribution 1) Based on the observation of highly non-uniform of visual supervision across response tokens, the proposed solution identifies visually sensitive tokens by the measurement between the real-image teacher and weak-image teacher. Then distills these tokens selectively. (from contribution 2) The visual-KL is the core innovation. Visual-KL measures KL between real-image and the blurred image. The proposed VS-OPSD distills on the top 30% sensitive tokens.(from contribution 3 + method)

Critical strengths
  1. The motivation is strong. The top 30% of tokens represent 88% of the visual signal, thus, selective distillation is meaningful.
  2. The strategy of separating visual dependence from generic distillation difficulty is the paper’s strongest idea. The weak-image teacher keeps the same cues and prefix, so Visual-KL reflects the image alone.
Critical weaknesses
  1. The scope is narrow. Training uses COCO object cues only, and evaluation covers object-level hallucination only.
  2. As the author mentioned, both selection and weighting contribute. But the method focuses on selection. The paper is built on the premise that selecting the top 30% of tokens is the main contribution.
Questions for authors

Can we add an ablation study about contribution attribution?

2 comments
AuthorsOpenReview ↗

Thank you for the insightful review! We add a complete contribution-attribution ablation and clarify the training and evaluation scope.

We thank the reviewer for the thoughtful review and for recognizing our core design. Below we address the question and the weaknesses in turn.

Q & W2: Contribution attribution of selection and weighting. The ablation in the paper already isolates part of this effect. In Table 2(a), Random-30 increases Hal from 25.4 (Vanilla OPSD) to 27.3, showing that sparsification alone does not help. On the same Visual-KL top-30% token set, Top-30 (uniform) obtains 25.1 Hal, while sensitivity weighting (VS-OPSD) reduces it to 22.4, showing that weighting brings a substantial additional gain beyond selection.

Following the reviewer's suggestion, we further conducted a complete contribution-attribution ablation, adopting the same experimental protocol as Table 1 in the paper:

MethodSelectionWeightingCHAIR ↓Cover ↑Hal ↓Cog ↓
Vanilla OPSDAllUniform5.748.225.52.6
Top-30% uniformTop 30%Uniform5.747.324.02.4
All-token weightedAllVisual-KL5.447.823.02.4
VS-OPSDTop 30%Visual-KL5.047.722.42.3

Both components contribute. Selection alone reduces Hal from 25.5 to 24.0, and weighting alone reduces it to 23.0. Combining the two yields the best Hal (22.4), together with the lowest CHAIR and Cog. The two components are therefore complementary: selection determines where to distill, and weighting determines how strongly to distill at each position. This ablation further supports our design of allocating supervision by Visual-KL. As the reviewer points out, both weighting and selection contribute to the results and act synergistically; we will emphasize this conclusion in the revision.

W1: Narrow scope.

Training cues. We use COCO for training because it is widely used and provides reliable ground-truth object annotations. To test whether the method depends on these specific labels, Table 3(c) replaces the COCO ground-truth cues with objects automatically detected by Grounding DINO. Under the same detector cues, VS-OPSD reduces Hal from 27.9 (Vanilla OPSD) to 24.7, only 0.8 above its ground-truth-cue counterpart. The method therefore does not rely on a particular source of annotations. This also suggests that large-scale automatically annotated data could serve as training data for our method, which is a promising direction for future work.

Evaluation. We agree that this work primarily targets object hallucination. However, VS-OPSD concentrates supervision on positions where the image changes the teacher's prediction, which strengthens the model's reliance on visual information beyond object names. Appendix D.1 (Figure 7) shows that the selected tokens extend beyond the provided cues to attributes, counts, and actions. In addition, the new Appendix G (Figures 9–10) shows corrected descriptions in complex scenes involving multiple objects and spatial relations. We further evaluated spatial-relation understanding on VSR:

LLaVA-1.5-7BZero-shot Acc ↑Random Acc ↑
Base64.9868.34
Vanilla OPSD66.6168.38
VS-OPSD67.1068.79

Although VS-OPSD is designed for object hallucination, it also outperforms both the base model and Vanilla OPSD in spatial-relation judgment, yielding a modest improvement. We believe that adapting the privileged cues, for example by incorporating spatial or relational information, so that VS-OPSD can directly target spatial and relational hallucination is a promising direction for future research.

The ablation study addressed my major concern. My final rating is 3 (Weak acceptance).

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

This paper introduces VS-OPSD, a visual-selective on-policy self-distillation approach designed to reduce hallucinations in open-ended vision-language generation. The key idea is to leverage privileged object cues and estimate token-level visual dependence by contrasting the predictions of a teacher conditioned on the original image with those conditioned on a degraded image. Based on this visual sensitivity signal, the method selectively distills a subset of informative tokens. The experimental results demonstrate consistent hallucination mitigation across several VLM configurations, with relatively limited degradation in general capabilities.

Critical strengths
  1. Well-motivated and intuitive methodology. The paper proposes a reasonable approach to concentrating supervision on visually sensitive tokens.
  2. Promising empirical improvements. The proposed method consistently improves hallucination-related metrics over vanilla, while maintaining competitive performance on general multimodal benchmarks.
  3. Comprehensive ablation studies. The paper includes useful analyses of token selection, sensitivity weighting, token budgets, visual degradation strategies, and robustness to noisy object cues.
Critical weaknesses
  1. VS-OPSD requires evaluating both real-image and weak-image teacher distributions during training. The additional computational cost is not thoroughly quantified, making it difficult to assess the efficiency trade-off relative to simpler distillation strategies.
  2. The method uses a fixed top-30% selection ratio, which is empirically motivated but may not be optimal for different models or response lengths.
  3. Visual-KL measures the sensitivity of teacher predictions to image degradation, but visually sensitive tokens are not necessarily hallucinated or incorrect tokens. Further analysis would help clarify when this distinction affects training effectiveness.
Questions for authors
  1. Could the authors train VS-OPSD under larger generation budgets (e.g., 256 or 512 tokens)? It would be particularly interesting to examine how hallucination rates and the effectiveness of Visual-KL-based token selection change with response length.

  2. Could the authors provide additional qualitative visualizations showing the selected tokens and their corresponding hallucination corrections, particularly for complex scenes involving multiple objects or spatial relationships?

1 comment
AuthorsOpenReview ↗

Thank you for the constructive review! We add experiments on training budgets, more visualization analysis, and point-by-point responses to the weaknesses.

We thank the reviewer for the positive assessment and constructive suggestions. Below we will address the questions and weaknesses.

Q1: Different training budgets. We did not train with 512 tokens because, in our captioning setting, the default 128-token budget already covers most responses.

(1) Hallucination Rates. We additionally trained both methods with 64-, 128-, and 256-token rollout budgets, following the same protocol as our main experiments (Table 1):

Training budgetMethodCHAIR ↓Cover ↑Hal ↓Cog ↓
—Base7.146.629.83.3
64Vanilla OPSD6.147.826.02.7
64VS-OPSD5.748.325.02.8
128Vanilla OPSD5.748.225.52.6
128VS-OPSD5.047.722.42.3
256Vanilla OPSD5.747.925.22.7
256VS-OPSD5.348.123.62.3

VS-OPSD consistently reduces hallucination over Vanilla OPSD under all three budgets. We observe smaller gains under the 64-token training budget. This may be because shorter training rollouts contain fewer hallucinations, resulting in weaker corrective supervision.

(2) Effectiveness of Visual-KL. We also examined whether Visual-KL selection remains effective as responses grow longer. On a probe of 500 COCO-val images, we group captions by length and apply the same top-30% selection:

Response length LLResponsesVisual-KL mass captured ↑Hallucinated mentions covered ↑
0<L≤640 < L \le 642288.44%0/1 (0.00%)
64<L≤12864 < L \le 12839391.09%256/356 (71.91%)
L>128L > 1288592.93%138/189 (73.02%)

The 0% coverage for L≤64L \le 64 is not informative: only one hallucinated mention is detected in this group, so the estimate has very high variance.

Overall, the selected 30% of tokens carry most of the Visual-KL mass and cover most hallucinated object mentions, showing that the selection strategy remains effective across response lengths.

Q2: Qualitative visualizations. Figure 7 (Appendix D.1) already shows that the selected tokens extend beyond the cues to entities, attributes, counts, and actions, and Figure 8 (Appendix D.2) shows that teacher-alignment gains concentrate on the selected positions. Also, we added Appendix G (Figures 9–10), which shows complete responses for complex multi-object and spatial-relation scenes, highlighting the selected tokens and the corrections relative to the base model.

Although our method mainly targets object hallucination, we also evaluated spatial-relation understanding on VSR (Visual Spatial Reasoning; Liu et al., 2023):

LLaVA-1.5-7BZero-shot Acc ↑Random Acc ↑
Base64.9868.34
Vanilla OPSD66.6168.38
VS-OPSD67.1068.79

VS-OPSD yields a modest improvement in spatial-relation judgment over both the base model and Vanilla OPSD.

W1: Computational cost. This is quantified in Appendix E.2 (Table 12). Relative to Vanilla OPSD, VS-OPSD increases step time by only 0.9% and peak reserved memory by 4.7%, since the weak-image teacher adds one gradient-free forward pass. The weak-image branch is used only during training, so VS-OPSD adds no parameters, latency, or computation at inference.

W2: Fixed top-30% ratio. Table 2(c) in the main paper shows that budgets of 10%, 30%, and 50% yield Hal of 23.3, 22.4, and 23.0, all better than Vanilla OPSD. The gain therefore holds over a broad range rather than at a finely tuned point. Our core insight is that a small fraction of tokens carries most of the visual signal; 30% is an empirical choice, not a requirement. The length analysis in Q1 further shows that this concentration holds across response lengths. Adapting the ratio to the model or response length is a promising direction, which we leave to future work.

W3: Visually sensitive tokens vs. hallucinated tokens. We agree that the two are not equivalent, and VS-OPSD does not aim to detect hallucinated tokens. As stated in Section 3.2, Visual-KL “measures visual sensitivity, not token correctness,” identifying positions where the image changes the teacher’s prediction.

These positions involve not only object names but also attributes, counts, and actions. Appendix D.1 shows that only 7.35% of selected words exactly match the provided cues, while 55.03% are non-cue content words. Thus, rather than specifically correcting the logits of hallucinated words, we distill the real-image teacher at positions that carry visual signal, teaching the student to rely more on visual evidence, that is, to "see what matters."

Nevertheless, the selected positions still cover approximately 72–73% of hallucinated object mentions in responses (Q1). The additional examples in Appendix G illustrate the selected visual concepts alongside corrected descriptions, providing an interpretable view of this supervision-allocation strategy.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response: Clarifications, Supplementary Analyses, and Main Findings

We thank all reviewers for their thoughtful and constructive feedback. We are encouraged that the reviewers found the motivation clear and well grounded (R1, R2, R3), the separation of visual dependence from generic distillation difficulty to be a strong idea (R2), the experiments comprehensive (R1, R3), and the training design practical, with no inference-time cost and consistent gains across backbones and random seeds (R4).

Clarifications. Several of the reviewers' questions concern aspects already covered by explanations and experiments in the paper. For these, we point to the relevant sections and clarify them in the individual responses:

  • Token budget (R1): ablation over 10%, 30%, and 50% selected tokens (Table 2(c)).
  • Visual sensitivity vs. hallucination (R1): Visual-KL measures visual sensitivity rather than token correctness (Section 3.2, Appendix D.1).
  • Training cost (R1): step time and memory relative to Vanilla OPSD (Appendix E.2).
  • Robustness to automatically detected cues (R2): Grounding DINO cues in place of COCO ground truth (Table 3(c)).
  • Relation to VA-OPD (R3, R4): discussion in Related Work and a controlled comparison (Section 5, Appendix E.1).

Supplementary analyses. Beyond these clarifications, and following the reviewers' suggestions, we provide the following supplementary experiments and analyses to further validate our method (details in the individual responses):

  • Selection and weighting ablation (R2, R4): a complete attribution study, including the all-token Visual-KL weighting variant, under the same protocol as Table 1.
  • Generation budgets (R1): training with 64-, 128-, and 256-token rollout budgets, and the effectiveness of Visual-KL selection across response lengths.
  • Hallucination at matched coverage (R4): prompts with four levels of detail.
  • Visual-KL vs. VA diagnostic (R3, R4): signal-level comparison on the same rollouts.
  • VSR evaluation (R1, R2): spatial-relation judgment.
  • Qualitative cases (R1, R2, R3): new Appendix G (Figures 9–10) with complex multi-object and spatial-relation scenes.

Main findings. These additional experiments further support the effectiveness and reliability of our method, as summarized below:

  1. Selection and weighting are complementary. Selection alone reduces Hal from 25.5 to 24.0 and weighting alone to 23.0, while combining them gives the best result (22.4).
  2. The hallucination reduction does not come from saying less. At matched coverage (e.g., Cover 49.60 vs. 49.55), VS-OPSD still lowers Hal from 25.65 to 23.00.
  3. Visual-KL captures hallucinated positions that VA misses. Under the same 30% budget, Visual-KL covers 76.6% of hallucinated object mentions versus 57.3% for VA, and recovers 22 of the 38 hallucinated mentions that receive VA = 0.
  4. Robustness across generation budgets. VS-OPSD outperforms Vanilla OPSD under all three training budgets, and Visual-KL selection remains effective across response lengths.

Revisions. The current revised PDF only adds Appendix G, and all other content remains unchanged. We will consider further revisions to the main text based on the reviewers' feedback during the discussion.

We thank the reviewers and the AC again for their time and effort. Please feel free to let us know if there are any further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.