acceptodds
Under review as a conference paper at ICLR 2027

Where to Apply Privilege: On-Policy Self-Distillation with Textual Privilege for Multimodal LLMs

Abstract

On-policy self-distillation (OPSD) has emerged as an effective post-training paradigm for multimodal large language models (MLLMs) that enhances visual understanding without external teacher models or reward verifiers. While vanilla OPSD adopts reference answers as textual privilege, recent variants for visual understanding primarily rely on visual privilege. However, we observe that MLLMs leverage answer-supporting evidence more effectively from targeted textual descriptions than from the original images, even when both inputs are individually sufficient for deriving the answer. This suggests substantial potential for textual privilege in self-distillation. Nevertheless, naive use of textual privilege can introduce context-specific artifacts, such as “based on the description”, even when no description is available to the student. This behavior reflects a failure to disentangle genuine improvements from traces of privileged text, which we term the privileged-context attribution problem. To address this problem, we propose Role-OPD. It adopts caption-based textual privilege to enhance the model's visual understanding. Furthermore, instead of presenting the privileged caption as user-provided context, Role-OPD places it in an intermediate assistant turn, framing it as the teacher's own visual understanding. This role-aware design suppresses the transfer of context-specific artifacts while preserving the genuine improvements from privileged captions. Experiments on six fine-grained visual understanding benchmarks show that Role-OPD achieves consistent gains and outperforms existing post-training methods.

Primary area:
foundation or frontier models, including LLMs
Keywords:
On-Policy Self-Distillation, Caption-based Textual Privilege, Role-aware Design, Multimodal LLMs

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_EnBFOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Role-OPD, an on-policy self-distillation method for multimodal LLMs. A frozen copy of the model serves as the teacher and receives a privileged textual caption during training. The paper makes two claims: 1) An “evidence-utilization gap.” Models answer far more accurately from targeted textual descriptions than from the original images, which suggests textual privilege is promising for self-distillation. 2) A “privileged-context attribution problem.” When privileged text is placed in the user turn, the student learns to produce artifacts such as “based on the description,” even though no description is available at inference time. To address the second issue, Role-OPD places the caption in an intermediate assistant turn, framing it as the teacher’s own visual understanding. Experiments on Qwen3.5-4B/9B across six fine-grained, high-resolution benchmarks and five held-out benchmarks show improvements over several baselines.

Critical strengths
  1. A concrete problem with a simple and effective solution. The paper identifies the privileged-context attribution problem, a practically relevant failure mode of textual privilege, and makes it measurable through the PCAR metric. It then resolves the problem by relocating the privileged caption to an assistant turn, which adds no inference overhead. This is the main reason I lean toward acceptance: the problem is real and clearly defined, and the fix is simple, cost-free, and easy to adopt in other self-distillation frameworks.

  2. Strong controls for the role-aware design. The ablation separating privilege form from placement is clear. The additional controls, an explicit no-attribution instruction and a three-turn all-user context, show that the gain comes from role assignment itself, not from suppressing attribution phrases or from the multi-turn structure. This is consequential because it makes the paper’s core claim about where to place privilege well supported, even though other parts of the evaluation are less rigorous.

Critical weaknesses
  1. The motivation experiment does not support its conclusion. The captions are generated with access to the ground-truth answer, so the near-ceiling accuracy in the textual setting reflects answer leakage rather than better evidence utilization. The visual setting, by contrast, uses the full image without localization such as cropping. The comparison therefore conflates modality with answer leakage. This matters for my assessment because it undermines the paper’s first contribution, the claim that textual privilege is superior to visual privilege, and with it the rationale for choosing captions in the first place.

  2. The gain over Vision-OPD cannot be clearly attributed to the role-aware design. The privileged captions come from a much larger, answer-aware model, so Role-OPD receives richer privileged content than the baselines. It is also unclear whether the on-policy baselines were trained with the same distillation recipe, even though the paper itself shows that such choices substantially affect accuracy.

Questions for authors
  1. What fraction of the privileged captions contain the answer string or the correct option verbatim? How does the evidence-utilization comparison change if the visual setting uses evidence-centered crops and the textual setting uses captions generated without the answer?

  2. How does Role-OPD perform when the captions are generated by the student model itself, rather than by Qwen3.5-397B?

  3. Which baseline results were taken from prior work, and were OPSD and Vision-OPD trained with the same objective, teacher update strategy, and data as Role-OPD?

4 comments

We thank the reviewer for the constructive comments and suggestions. We hope our point-by-point responses below address your concerns.

[W1/Q1]: Clarification of motivation experiments

We thank the reviewer for requesting this clarification. We clarify that none of our captions contain explicit answer statements such as “The answer is A. Red” or correct-option labels. We inspected all captions and found that the answer string appears only as part of a description of visual facts rather than as a standalone answer label. The original images already contain the visual evidence needed to determine the correct answer. We use the ground-truth answer to guide caption generation to construct textual evidence sufficient to support answering the question, rather than to directly provide an answer label to the target model. The captions also contain question-irrelevant image content, so the target model must still identify and utilize the relevant evidence. Without answer guidance, the caption generator may omit critical evidence or introduce perceptual errors, conflating evidence sufficiency with the target model’s ability to utilize that evidence. Our original experiment therefore compares the utilization of answer-sufficient evidence presented in visual and textual forms, and our conclusion is restricted to this evidence-sufficient setting.

Following the reviewer’s suggestion, we conduct additional experiments using the evidence-centered crops provided by the dataset and captions generated without access to the ground-truth answer. The evidence-centered crops ensure that the visual input contains and localizes the evidence needed to answer the question. Nevertheless, their question-answering accuracy remains substantially lower than that obtained from answer-guided captions. We also manually inspect the captions generated without answer access and find that only 61.88% contain sufficient evidence to answer the corresponding question. This confirms that answer guidance primarily ensures evidence sufficiency. Despite this disadvantage, captions generated without answer access still achieve higher average accuracy than full-image inputs.

Question-answering accuracy (%) under different visual and textual evidence settings.

Evidence ModalityEvidence ConstructionQwen-3.5-4BQwen-3.5-9BGPT-5.5GLM-5.2Average
VisualFull image48.9854.0364.1461.8357.25
VisualEvidence-centered crop73.1975.0482.8680.3977.87
TextualCaption w/o answer access59.2560.6361.2261.1660.57
TextualCaption w/ answer access97.4098.0698.7898.7098.24

[W2/Q2/Q3]: Attribution of performance gains

We thank the reviewer for requesting this clarification. We clarify that Vision-OPD also uses Qwen3.5-397B to construct its privileged information. Therefore, the original comparison does not provide Role-OPD with privileged information from a stronger generator than the baseline. We further evaluate Role-OPD using captions generated by the 4B student model itself. As shown below, it achieves an average performance of 77.79, outperforming the Vision-OPD result of 76.42 reported in our paper.

Performance of Role-OPD (4B) with external and self-generated captions.

Caption GeneratorV$^\star$ZoomBenchHR-Bench 4KHR-Bench 8KMME-RealWorld ENMME-RealWorld CNAvg.
Qwen3.5-4B91.6260.7185.5081.2575.3572.3377.79
Qwen3.5-397B92.1561.1885.7581.6275.6172.6278.16

The results for the Closed-Source Models (Single Forward Pass) and Open-Source Models (Single Forward Pass) groups in Table~1 of our paper are taken from the Vision-OPD. We will clarify their provenance in the revised manuscript. In our main experiments, OPSD and Role-OPD use the same training objective, teacher-update strategy, and training data. Vision-OPD uses the same training objective and data as Role-OPD but retains the best teacher-update strategy reported in its original work. We further report results under matched settings, as shown below.

Performance of OPSD, Vision-OPD, and Role-OPD under matched settings.

MethodParam SizeV$^\star$ZoomBenchHR-Bench 4KHR-Bench 8KMME-RealWorld ENMME-RealWorld CNAvg.
OPSD4B86.3954.3282.3879.6274.8172.5275.01
Vision-OPD4B88.4860.4782.8879.5074.7270.6676.12
Role-OPD4B92.1561.1885.7581.6275.6172.6278.16

Thank you for the detailed responses. My concerns, especially those regarding the motivation, experiments, and attribution of the gains, have been addressed. Therefore, I will keep my positive score.

Dear Reviewer EnBF,

We are glad that our response has addressed your concerns. We sincerely appreciate your time and thoughtful feedback.

Best regards

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

The paper studies the placement of training-only textual privilege in on-policy self-distillation for multimodal LLMs. Role-OPD uses answer-supporting image captions and places them in an intermediate assistant turn in the teacher context, while the student receives only the image and question. The teacher is frozen and initialized from the same checkpoint as the student, which is trained using token-level Jensen–Shannon divergence on student-generated trajectories. Experiments on Qwen3.5-4B and 9B report improved average visual-understanding performance and a lower Privileged-Context Attribution Rate (PCAR). The paper also evaluates explicit no-attribution instructions, a matched three-turn all-user context, held-out benchmarks, and autonomous image-description quality.

Critical strengths

1. The role-aware context design is supported by targeted ablation experiments. On the 4B model, moving from caption-injected OPSD to Role-OPD improves the average score from 75.54 to 78.16 and reduces PCAR from 73.36% to 0.84%.

2. The method offers practical benefits without introducing additional inference-time components. The paper reports average improvements at both model sizes. Results on five held-out benchmarks, together with human and LLM preference evaluations of autonomous image descriptions, provide additional evidence of useful transfer. Despite the small implementation change and the use of an existing distillation objective, these results demonstrate practical value.

Critical weaknesses

1. The role-placement effect is supported, but the necessity of detailed image captions remains unclear. Compared with standard OPSD, the proposed method changes both the content and placement of privileged information: reference answers are replaced with answer-supporting image captions, and the teacher adopts a role-aware context. However, the paper does not include the crucial control of placing the reference answer itself in an intermediate assistant turn. It therefore remains unclear whether detailed captions are necessary or whether role reassignment benefits textual privilege more generally.

2. The proposed mechanism and the robustness of the gains need stronger validation. Lower PCAR indicates less explicit attribution to unavailable context, but does not necessarily imply more faithful use of visual evidence. The metric is assessed by a single LLM judge without reported human validation. The paper does not report multi-seed results or uncertainty estimates, particularly for the modest generalization gains over Vision-OPD in Appendix G (Table 4): 0.72 points for 4B and 0.52 points for 9B.

Questions for authors

1. Does detailed caption privilege remain beneficial after controlling for role placement? A comparison of reference answers and detailed captions in the same intermediate assistant turn, alongside the OPSD and Vision-OPD baselines in Table 1, would help isolate the effects of privileged content and role placement. The comparison should use matched training samples, teacher configurations, and training settings.

2. How well do the evaluations support the proposed mechanism and the robustness of the gains? Please provide human validation of the PCAR judge and clarify how reduced explicit attribution relates to more faithful use of visual information. Multi-seed results or uncertainty estimates for the main role-placement comparison and Appendix G (Table 4) would help assess the stability of the improvements.

4 comments

We thank the reviewer for the constructive comments and suggestions. We hope our point-by-point responses below address your concerns.

[W1/Q1]: Disentangling Privileged Content and Role Placement

We thank the reviewer for suggesting this controlled comparison. We clarify that the benefit of our role-aware design generalizes beyond caption privilege. We introduce a controlled comparison that fixes the privilege placement to the same intermediate assistant turn and varies only the privilege type. As shown below, under matched settings, Role-OPD consistently outperforms OPSD and Vision-OPD with reference answer, hint, rationale, and caption privileges, while caption achieves the highest average performance. Detailed captions are therefore not required for the role-aware design to be effective, but they provide an effective form of privilege that fits the interaction structure well. When generating the intermediate assistant turn, the model has observed the image but not the question, making caption generation natural, whereas producing an answer, hint, or rationale at this stage is less semantically coherent. Thus, the role-aware design benefits textual privilege more generally, while captions provide a natural and effective form of privilege for this interaction structure.

Performance comparison of OPSD, Vision-OPD, and Role-OPD under matched settings.

MethodParam SizeV$^\star$ZoomBenchHR-Bench 4KHR-Bench 8KMME-RealWorld ENMME-RealWorld CNAvg.
OPSD4B86.3954.3282.3879.6274.8172.5275.01
Vision-OPD4B89.0159.8883.6280.6274.8670.5376.42
Role-OPD (Answer)4B92.1560.8385.5081.5075.1171.8477.82
Role-OPD (Hint)4B91.1059.2985.5080.6374.2272.1877.15
Role-OPD (Rationale)4B91.6258.8285.7581.5075.0272.4777.53
Role-OPD (Caption)4B92.1561.1885.7581.6275.6172.6278.16

[W2/Q2]: Mechanism validation and robustness

We thank the reviewer for highlighting the need for stronger mechanism validation and robustness analysis. We evaluate PCAR on V$^\star$ using human annotators and three independent model judges. As shown in Table 1, the human annotators and all three model judges report similar PCAR for each method. We clarify that PCAR is not a direct measure of visual faithfulness. Instead, it diagnoses a specific distillation failure mode. The teacher may explicitly attribute its prediction to privileged information unavailable to the student, and the student may inherit this attribution pattern through distillation. Because the student observes only the original image and question, this pattern can cause it to refer to context absent from its input rather than ground its response in the available visual evidence. Role-OPD presents the privilege as the teacher’s own understanding, reducing both explicit attribution to privileged context and its transfer to the student. Lower PCAR therefore indicates that the distilled student is less affected by unavailable-context attribution and that its supervision is better aligned with the evidence available in its input. This supports more faithful use of visual information.

Table 1: PCAR (%) evaluated by human annotators and various model judges on V$^\star$ Bench.

MethodHumanQwen3.5-397BGLM-5.2GPT-5.5
OPSD90.0587.4388.4888.48
Caption-injected OPSD68.5970.6871.7370.16
Role-OPD0.520.521.570.52

To assess the robustness of the performance gains, we conduct multi-seed experiments with seeds 40, 41, and 42 for Role-OPD, OPSD, and Vision-OPD using the 4B model. We report the mean and standard deviation on the fine-grained visual perception benchmarks and held-out benchmarks in Tables 2 and 3, respectively. Due to the limited rebuttal period and computational cost, we currently report multi-seed results only for the 4B setting and these two baselines. We will include results for additional baselines and the 9B model in the revised manuscript.

Table 2: Multi-seed results on fine-grained visual perception benchmarks.

MethodParam SizeV$^\star$ZoomBenchHR-Bench 4KHR-Bench 8KMME-RealWorld ENMME-RealWorld CNAvg.
OPSD4B86.740.8986.74_{0.89}55.030.5155.03_{0.51}82.790.2982.79_{0.29}79.670.3679.67_{0.36}74.820.0474.82_{0.04}72.310.1572.31_{0.15}75.230.2475.23_{0.24}
Vision-OPD4B89.180.2589.18_{0.25}59.570.2459.57_{0.24}83.410.2983.41_{0.29}80.750.1180.75_{0.11}74.840.0274.84_{0.02}70.490.1170.49_{0.11}76.370.0976.37_{0.09}
Role-OPD4B91.970.25\mathbf{91.97}_{\mathbf{0.25}}61.140.15\mathbf{61.14}_{\mathbf{0.15}}85.580.12\mathbf{85.58}_{\mathbf{0.12}}81.710.12\mathbf{81.71}_{\mathbf{0.12}}75.600.04\mathbf{75.60}_{\mathbf{0.04}}72.500.11\mathbf{72.50}_{\mathbf{0.11}}78.080.10\mathbf{78.08}_{\mathbf{0.10}}

Table 3: Multi-seed results on held-out benchmarks.

MethodParam SizeMMVPCV-BenchMMStarPOPEVisualProbeAvg.
OPSD4B78.560.4278.56_{0.42}87.980.4887.98_{0.48}77.510.9377.51_{0.93}88.151.0788.15_{1.07}49.951.3749.95_{1.37}76.430.3776.43_{0.37}
Vision-OPD4B78.220.6978.22_{0.69}87.740.7387.74_{0.73}79.000.2579.00_{0.25}89.150.10\mathbf{89.15}_{\mathbf{0.10}}53.800.1653.80_{0.16}77.580.1277.58_{0.12}
Role-OPD4B79.890.83\mathbf{79.89}_{\mathbf{0.83}}88.810.10\mathbf{88.81}_{\mathbf{0.10}}79.870.52\mathbf{79.87}_{\mathbf{0.52}}88.720.1188.72_{0.11}54.460.36\mathbf{54.46}_{\mathbf{0.36}}78.350.24\mathbf{78.35}_{\mathbf{0.24}}
AuthorsOpenReview ↗

Brief Follow-up on Our Response

Dear Reviewer aSRi,

Thank you again for your time and constructive feedback. As the discussion period is coming to a close, we wanted to briefly follow up on our responses.

We addressed your concerns by adding a controlled comparison that fixes role placement while varying the privileged-information type, showing that the role-aware design generalizes to reference answers, hints, and rationales, with captions achieving the highest average performance. We also provided human and multi-model validation of PCAR, clarified its relationship to visual faithfulness, and reported multi-seed results for both the main and held-out evaluations at the 4B scale.

We hope these additional results and clarifications address your concerns, and we would be grateful if you could take them into account in your final assessment. Please let us know if any questions remain.

[W2/Q2]: Mechanism validation and robustness

To assess the robustness of the performance gains, we conduct multi-seed experiments (seeds 40, 41, 42) for Role-OPD, OPSD, and Vision-OPD. Below we report the mean and standard deviation on fine-grained visual perception and held-out benchmarks for the 9B model.

Multi-seed results on fine-grained visual perception benchmarks.

MethodParam SizeV$^\star$ZoomBenchHR-Bench 4KHR-Bench 8KMME-RealWorld ENMME-RealWorld CNAvg.
OPSD9B91.450.6591.45_{0.65}57.630.3957.63_{0.39}86.000.2086.00_{0.20}82.090.3682.09_{0.36}75.160.2075.16_{0.20}73.110.36\mathbf{73.11}_{\mathbf{0.36}}77.570.1077.57_{0.10}
Vision-OPD9B91.970.2591.97_{0.25}64.900.39\mathbf{64.90}_{\mathbf{0.39}}87.540.1587.54_{0.15}85.130.1885.13_{0.18}73.410.2273.41_{0.22}70.400.5170.40_{0.51}78.890.1278.89_{0.12}
Role-OPD9B96.510.25\mathbf{96.51}_{\mathbf{0.25}}62.480.2562.48_{0.25}88.210.12\mathbf{88.21}_{\mathbf{0.12}}85.840.21\mathbf{85.84}_{\mathbf{0.21}}76.760.15\mathbf{76.76}_{\mathbf{0.15}}72.760.2472.76_{0.24}80.430.12\mathbf{80.43}_{\mathbf{0.12}}

Multi-seed results on held-out benchmarks.

MethodParam SizeMMVPCV-BenchMMStarPOPEVisualProbeAvg.
OPSD9B81.550.3281.55_{0.32}88.240.3188.24_{0.31}81.470.1981.47_{0.19}88.990.1188.99_{0.11}51.260.5751.26_{0.57}78.300.0978.30_{0.09}
Vision-OPD9B81.891.1381.89_{1.13}88.540.6988.54_{0.69}82.910.2882.91_{0.28}89.820.09\mathbf{89.82}_{\mathbf{0.09}}56.960.6456.96_{0.64}80.020.3880.02_{0.38}
Role-OPD9B83.440.68\mathbf{83.44}_{\mathbf{0.68}}88.960.33\mathbf{88.96}_{\mathbf{0.33}}83.400.84\mathbf{83.40}_{\mathbf{0.84}}89.570.0989.57_{0.09}58.131.02\mathbf{58.13}_{\mathbf{1.02}}80.710.36\mathbf{80.71}_{\mathbf{0.36}}
Reviewer_oJpDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Role-OPD, an on-policy self-distillation method for improving fine-grained visual understanding in MLLMs. It places privileged image captions in an intermediate assistant turn rather than the user message, reducing privileged-context attribution artifacts while improving visual question-answering performance.

Critical strengths
  • The paper is clearly motivated by the transfer of privileged-context attribution artifacts from teachers to students.
  • The presentation is clear, and the method is easy to follow.
  • Role-OPD is a simple yet practical design that improves performance without additional inference overhead.
Critical weaknesses
  • Related-work coverage and comparisons are incomplete, particularly regarding ViCuR’s textual visual cues and hint leakage, and H-OPD’s image-description-based teacher conditioning.
  • The role-aware design is evaluated primarily with captions. Its effectiveness with other privileged information, such as hints, reference answers, and rationales, remains explained.
  • The method still relies on captions generated by an external strong model, making its gains dependent on caption quality and weakening the claim of external-teacher independence. For self-distillation method it should be a great concern.
  • The mechanism behind the role-aware design remains insufficiently explored. The existing controls establish its effectiveness but do not explain how role assignment changes the teacher’s predictive distribution.
Questions for authors
  • Does assigning reference answers, hints, or rationales to the assistant role yield similar improvements?
  • How sensitive is performance to the caption generator’s strength and to captions generated without access to reference answers?
  • What are the teacher accuracies under caption-injected OPSD and Role-OPD? Does the role-aware design improve teacher correctness, or mainly change how its supervision transfers to the student?
  • How does role assignment affect teacher accuracy, confidence, and teacher–student distribution alignment beyond suppressing explicit attribution? I'll be glad to raise my score if major concerns mentioned can be well resolved.
4 comments

We thank the reviewer for the constructive comments and suggestions. We hope our point-by-point responses below address your concerns.

[W1]: About incomplete related-work coverage and comparisons

We thank the reviewer for pointing us to ViCuR and H-OPD. ViCuR and Role-OPD address the same general problem of teacher-student mismatch caused by teacher-only privileged information, but take different approaches. ViCuR replaces answer-side hints with textual visual cues whose underlying evidence remains accessible in the image, and introduces a cue-recovery module to help the student recover this evidence from visual tokens. Role-OPD instead changes how textual privilege is presented, placing the caption in an intermediate assistant turn to frame it as the teacher’s own visual understanding and suppress the transfer of privileged-context attribution. H-OPD uses task-relevant image descriptions to provide visual semantics to a text-only teacher and dynamically combines supervision from vision-language and text-only teachers at the token level. In contrast, Role-OPD does not require multiple teachers or token-level combination of teacher supervision.

Since H-OPD has not publicly released its code or model weights, we are unable to evaluate it under our experimental setup. Therefore, we report below additional comparative experiments with ViCuR only, using the same experimental setup.

Performance comparison of ViCuR and Role-OPD under matched settings.

MethodParam SizeV$^\star$ZoomBenchHR-Bench 4KHR-Bench 8KMME-RealWorld ENMME-RealWorld CNAvg.
ViCuR4B81.6849.4786.2580.3868.5162.7771.51
Role-OPD4B92.1561.1885.7581.6275.6172.6278.16

We believe that the comparison with ViCuR provides further evidence of the effectiveness of our method. In the revised manuscript, we will expand the discussion of related work, including ViCuR and H-OPD, and add the corresponding comparative experiments to the experimental section.

[W2/Q1]: Role-aware design with diverse privileged information

We thank the reviewer for requesting this clarification. We clarify that assigning reference answers, hints, or rationales to the assistant role also yields consistent improvements in average performance, as shown in the table below. We adopt image captions as privileged information because, when the model receives only an image, an image description is a natural assistant response. In contrast, generating a reference answer, hint, or rationale before receiving the question is less natural. Consistent with this role alignment, image captions also achieve the highest average performance.

Performance of Role-OPD (4B) under different privileged-information types and role placements.

Privileged InformationRoleV$^\star$ZoomBenchHR-Bench 4KHR-Bench 8KMME-RealWorld ENMME-RealWorld CNAvg.
Reference answerUser86.3954.6784.5080.6374.8170.6875.28
Reference answerAssistant92.1560.8385.5081.5075.1171.8477.82
HintUser89.5354.6784.3880.5074.7071.4275.87
HintAssistant91.1059.2985.5080.6374.2272.1877.15
RationaleUser89.5355.2783.1378.3874.8471.7175.48
RationaleAssistant91.6258.8285.7581.5075.0272.4777.53
CaptionUser86.3958.9383.0080.1275.0272.4075.98
CaptionAssistant92.1561.1885.7581.6275.6172.6278.16

[W3/Q2]: Sensitivity to caption generation settings

We thank the reviewer for requesting this clarification. First, our claim of external-teacher independence refers to the distillation process. Role-OPD does not require a separate, substantially larger model to serve as the teacher during training. Qwen3.5-397B is used only for offline construction of caption-based privileged information, as is also the case for Vision-OPD, and is not the teacher model from which the student is distilled. We will clarify this scope in the revised manuscript. Second, constructing caption-based privileged information does not require a substantially stronger generator, but its evidence sufficiency is affected by answer access. Without reference-answer conditioning, the generator must independently infer the answer and identify the supporting visual evidence, potentially omitting necessary evidence or introducing perceptual errors. As shown below, removing answer access reduces teacher accuracy from 94.26% to 58.63%. Answer conditioning is therefore a deliberate data-construction choice that improves the evidence sufficiency of the privileged captions.

Teacher accuracy (Qwen-3.5 4B) with and without answer access.

Answer AccessTeacher Accuracy (%)
×58.63
√94.26

Using Qwen3.5-4B as the training base model, we further generate captions with Qwen3.5-4B itself and evaluate captions generated by Qwen3.5-397B without answer access. As shown below, replacing Qwen3.5-397B with Qwen3.5-4B under answer conditioning decreases the average performance by only 0.37 points, showing limited sensitivity to generator scale. In contrast, removing answer access decreases the average performance by 1.72 points, as captions generated without answer conditioning may fail to establish a sufficient informational advantage for the teacher. These results indicate that Role-OPD is more sensitive to the evidence sufficiency of the captions than to the scale of the caption generator.

Performance of Role-OPD (4B) under different caption generation settings.

Caption GeneratorAnswer AccessV$^\star$ZoomBenchHR-Bench 4KHR-Bench 8KMME-RealWorld ENMME-RealWorld CNAvg.
Qwen3.5-4B√91.6260.7185.5081.2575.3572.3377.79
Qwen3.5-397B×87.9657.0484.8881.5075.0272.2576.44
Qwen3.5-397B√92.1561.1885.7581.6275.6172.6278.16

[W4/Q3/Q4]: Mechanism of role-aware design

We thank the reviewer for requesting this clarification. First, the role-aware design has little effect on teacher accuracy and primarily changes how supervision transfers to the student. As shown below, caption-injected OPSD and Role-OPD achieve similar teacher accuracies of 93.96% and 94.26%, respectively. When the caption is placed in the user turn, it becomes additional external input available only to the teacher. The teacher therefore produces supervision conditioned on both the original input and privileged evidence, whereas the student must approximate this supervision using only the image and question. This mismatch can make the teacher distribution less transferable. When the caption is placed in an intermediate assistant turn, it is represented as the teacher’s own generation from the image rather than additional user-provided evidence. The teacher and student therefore retain the same external user-provided input, while the teacher additionally retains the caption as part of its own generation trajectory. Although this does not eliminate their information asymmetry, it reduces the mismatch between their conditioning structures, making the teacher’s predictive distribution better aligned with the distribution that the student can learn from its available input.

Teacher accuracy under caption-injected OPSD and Role-OPD.

MethodParam SizeTeacher Accuracy (%)
Caption-injected OPSD4B93.96
Role-OPD4B94.26

To further examine how role assignment changes the teacher’s predictive distribution, we evaluate teacher accuracy, confidence, and teacher-student alignment under controlled user-role and assistant-role placements. As shown below, the two role assignments yield nearly identical teacher accuracy (94.23% vs. 94.26%) and confidence (0.8982 vs. 0.9018). In contrast, placing the caption in the assistant role reduces the mean teacher-student token JSD from 0.0580 to 0.0326, a 43.8% relative reduction. For accuracy and confidence, we prepend the fixed prefix ``The correct option is:'' and normalize the next-token logits over the four answer choices. Accuracy is determined by the option with the highest probability, and confidence is measured as the average probability assigned to the ground-truth option. For distribution alignment, we apply teacher forcing to identical student-generated responses and compute the full-vocabulary next-token Jensen--Shannon divergence at each response position, averaged across tokens. These results show that role assignment has little effect on teacher accuracy or confidence but substantially improves the alignment between the teacher and student predictive distributions.

Effects of caption role assignment on teacher accuracy, confidence, and teacher–student distribution alignment (Qwen-3.5 4B).

Caption RoleTeacher Accuracy (%)Teacher ConfidenceTeacher–Student Token JSD ↓
User94.230.89820.0580
Assistant94.260.90180.0326
AuthorsOpenReview ↗

Brief Follow-up on Our Response

Dear Reviewer oJpD,

Thank you again for your time and constructive feedback. As the discussion period is coming to a close, we wanted to briefly follow up on our responses.

We addressed your concerns through additional comparisons with ViCuR, controlled experiments across different privileged-information types, and evaluations with student-generated and answer-agnostic captions. We also clarified the scope of external-teacher independence and further analyzed the mechanism through teacher accuracy, confidence, and teacher–student distribution alignment. These additional results support the generality of the role-aware design, its limited dependence on caption-generator scale, and its effect on supervision transfer.

We hope these results and clarifications address your concerns, and we would be grateful if you could take them into account in your final assessment. Please let us know if any questions remain.

Comments on the paper

1 comment
AuthorsOpenReview ↗

A Brief Summary of the Rebuttal Phase

Dear AC and Reviewers,

We sincerely thank all reviewers for their constructive feedback and engagement during the rebuttal and discussion phases. During the rebuttal, we provided additional controlled experiments and analyses addressing the reviewers’ main concerns. Reviewer EnBF has confirmed that our responses addressed their concerns and that they will retain their positive score.

Specifically, our discussions focused on the following aspects:

  • Evidence-Utilization Motivation and Experimental Controls (EnBF): We clarified that answer conditioning is used to ensure that captions contain sufficient visual evidence rather than to provide explicit answer labels. We further evaluated evidence-centered image crops and captions generated without answer access, directly addressing concerns about answer leakage and unequal evidence localization.

  • Disentangling Role Placement, Privileged Content, and Caption Generation (EnBF, oJpD, aSRi): Across our additional controlled experiments, we separately examined privilege type, role placement, and caption-generator scale. The results show that assistant-role placement benefits reference answers, hints, rationales, and captions, and that Role-OPD retains its performance when captions are generated by the student model itself. We also clarified baseline provenance and training configurations to better isolate the sources of the gains.

  • Mechanism and Evaluation Validation (oJpD, aSRi): Human annotators and three independent model judges produced consistent PCAR evaluations. We further showed that role assignment has little effect on teacher accuracy or confidence but substantially reduces teacher–student distribution mismatch. These analyses support our explanation that Role-OPD primarily improves the transferability of teacher supervision.

  • Robustness and Related-Work Comparisons (oJpD, aSRi): We reported multi-seed results for both 4B and 9B models on the main and held-out benchmarks, supporting the stability of the improvements. We also added a matched comparison with ViCuR, clarified the differences from ViCuR and H-OPD, and refined the scope of our external-teacher-independence claim.

Once again, we thank the AC and all reviewers for their time and valuable feedback.

Best regards,

Authors of Submission 27997

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.