acceptodds
Under review as a conference paper at ICLR 2027

Rethinking Privileged Information Effectiveness in Multimodal On-Policy Self-Distillation

Abstract

On-policy self-distillation (OPSD) can provide the teacher with privileged information (PI) unavailable to the student, potentially improving teacher supervision during training. However, it remains unclear whether PI that provides a larger direct benefit to the teacher also leads to larger gains that persist in the PI-free student. We study this question in multimodal OPSD by explicitly separating two quantities: PI access gain, the target-task performance gain obtained when a fixed model is given direct access to PI, and PI effectiveness, the target-task performance gain retained by the PI-free student after OPSD relative to a matched no-PI baseline. In controlled fine-grained perception experiments, these quantities are not consistently aligned across PI conditions: Answer yields a larger PI access gain than Hint, whereas Hint achieves higher PI effectiveness. Student-side analyses further reveal condition-specific differences in the use of task-relevant visual evidence, with Hint showing the clearest target-region localization and selective sensitivity to target-region corruption. Extending the analysis across tasks further shows substantial variation in PI effectiveness across target tasks and concrete PI conditions. Overall, our results show that greater PI access gain does not necessarily translate into greater PI effectiveness, and that PI effectiveness is task- and condition-dependent and not intrinsic to a PI form. Motivated by this task dependence, a simple task-level PI routing strategy improves aggregate held-out performance over uniform PI baselines.

Primary area:
interpretability and explainable AI
Keywords:
OPSD, distillation, VLM, Privileged Information

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_t7jyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the effects of different forms of privileged information (PI) on multimodal on-policy self-distillation (OPSD). The authors introduce two metrics, "PI Access Gain" and "PI Effectiveness," to characterize the discrepancy between the direct performance gains obtained from PI and its effectiveness as a distillation supervision signal. Through experiments across four task types, the authors demonstrate that greater PI Access Gain does not necessarily translate into higher PI Effectiveness, and that the effectiveness of different PI conditions is highly task-dependent. The paper further analyzes how students trained under different PI conditions attend to and rely on task-relevant visual regions. Finally, the authors propose a validation-based, task-level PI routing strategy as a heuristic for selecting appropriate privileged information for different tasks.

Critical strengths
  1. The paper addresses a meaningful research question in multimodal OPSD by distinguishing PI Access Gain from PI Effectiveness, highlighting that stronger privileged information does not necessarily provide more effective supervision for PI-free students.
  2. The experiments systematically compare different PI conditions across four multimodal tasks using matched No-PI OPSD baselines, revealing task-dependent differences in PI effectiveness and providing valuable empirical insights.
  3. The paper is clearly structured and well presented. The Visual Overlap and Matched Occlusion analyses offer complementary perspectives on how different PI conditions influence students' visual attribution and reliance on task-relevant evidence.
Critical weaknesses

The paper's central observation is that the direct performance gains from PI do not necessarily translate into corresponding distillation benefits. However, prior work such as ViCuR has already highlighted the limitations of answer-based privileged information and demonstrated that visually grounded privileged supervision can lead to more effective distillation. Although this paper further defines PI Access Gain and PI Effectiveness and extends the investigation across multiple tasks and PI conditions, it remains unclear whether these results provide substantially new insights or mainly offer broader empirical validation of existing findings. Furthermore, the proposed PI Routing strategy appears to be more of a validation-based heuristic for selecting task-specific PI conditions. While it brings some performance improvements, it also requires evaluating different PI conditions on a validation set in advance to determine the routing configuration. The trade-off between performance gains and computational cost therefore deserves further discussion.

Regarding differences in PI Effectiveness, the analyses in Section 4 provide some insights into the potential mechanisms behind Hint-based OPSD but do not sufficiently explain why different PI conditions lead to different distillation outcomes. In Figure 2, apart from Hint, the Visual Overlap distributions of other PI conditions and baselines do not exhibit noticeable differences that correspond to their varying distillation gains. This raises questions about whether Visual Overlap can effectively explain differences in PI Effectiveness or mainly reflects the specific characteristics of Hint as localized visual privileged information. Moreover, other baselines in Figure 3 exhibit similar sensitivity to target-region occlusion as the Hint-trained model. Although this experiment demonstrates the models' reliance on task-relevant visual evidence, such behavior is largely expected in fine-grained perception tasks and does not sufficiently explain why different PI conditions produce different distillation outcomes. Overall, these analyses primarily characterize differences in student visual behavior rather than providing a convincing mechanistic explanation for the paper's central findings.

Questions for authors
  1. How does the paper's central insight differ from ViCuR? What new insights does it provide beyond broader empirical validation across different PI conditions and tasks?

  2. Could the authors more clearly articulate the key findings and claims of the mechanistic analyses in Section 4, and explain how they support the paper's central insight regarding PI Effectiveness?

8 comments
AuthorsOpenReview ↗

Response 1

Dear Reviewer t7jy,

We sincerely thank you for your careful review and constructive feedback. We particularly appreciate your concerns regarding the novelty, mechanistic insights, and practical implications of our work. These comments have helped us further clarify our contributions and their distinctions from prior work. We respond to your concerns below.

W1: The paper's central observation is that the direct performance gains from PI do not necessarily translate into corresponding distillation benefits. [...] The trade-off between performance gains and computational cost therefore deserves further discussion.

Q1: How does the paper's central insight differ from ViCuR? What new insights does it provide beyond broader empirical validation across different PI conditions and tasks?

1. Clarifying the Difference from ViCuR (W1, Q1)

We would first like to clarify an important detail about ViCuR: the privileged information provided to its teacher is actually a textual Visual Cue (denoted as SS), rather than visual input itself. These cues describe question-relevant visual evidence, whose underlying information remains available in the student's original image. ViCuR further introduces a Sink-Token Cross-Attention module, allowing the student to aggregate relevant evidence from its own visual tokens into an internal Cue Representation.

The key difference between the two works lies in their research focus: ViCuR investigates the recoverability of textual cues grounded in visual evidence and their student-side implementation, whereas we systematically study how different forms of PI translate teacher-side access benefits into student-side learning gains, and how this transfer relates to student behavior and task requirements.

2. Insights Beyond Broader Empirical Comparisons (W1, Q1)

Our study provides three complementary insights.

First, PI Access Gain does not necessarily predict PI Effectiveness. We distinguish the direct performance benefit obtained by a fixed Base Model accessing PI from the incremental learning gain retained by the PI-free student after OPSD, relative to a matched No-PI baseline. Our FGP results show that ANSWER yields a greater PI Access Gain, whereas HINT achieves higher PI Effectiveness. This demonstrates that the direct benefits of PI access to the teacher do not necessarily translate into greater learning gains for the student.

Second, different PI conditions are associated with differences in student-side visual behavior. Through Visual Attribution and Matched Occlusion analyses, we investigate how students trained under different PI conditions utilize task-relevant visual evidence. These analyses provide a complementary characterization of the resulting students' behavior beyond final accuracy.

Third, PI Effectiveness is task-dependent rather than intrinsic to a particular PI format. Our cross-task analysis reveals that the relative effectiveness of different PI conditions changes with task requirements. For example, visual HINT is particularly beneficial for fine-grained perception, whereas ANSWER provides greater student-side gains in knowledge-intensive perception (KIP) and multi-hop spatial reasoning (MSR). In KIP, crop-based HINT helps the teacher localize the target region but does not provide the knowledge needed to interpret the identified content. In MSR, even a Spatial HINT specifically designed for the task does not necessarily yield greater student gains. These findings suggest that PI selection should consider whether the information provided matches the actual requirements of the target task, rather than assuming a generally superior PI format.

3. Practical Implications and Limitations of PI Routing (W1)

Building on our cross-task analysis, we further explore Task-level PI Routing as a preliminary investigation of the potential of selecting PI according to task requirements. The current routing strategy is a small-scale proof of concept conducted on subsets of the four tasks. We consider three candidate conditions (No-PI, ANSWER, and HINT) and use 200 validation examples per task to select the PI condition.

We acknowledge that the current strategy introduces additional computational costs. Importantly, the current PI Routing serves as a preliminary downstream application of our finding regarding the compatibility between task requirements and PI. We hope to further explore more computationally efficient PI selection methods and investigate the trade-off between performance gains and computational costs in practical applications.

AuthorsOpenReview ↗

Response 2

W2: Regarding differences in PI Effectiveness, the analyses in Section 4 provide some insights into the potential mechanisms behind Hint-based OPSD but do not sufficiently explain why different PI conditions lead to different distillation outcomes. [...]

Overall, these analyses primarily characterize differences in student visual behavior rather than providing a convincing mechanistic explanation for the paper's central findings.

Q2: Could the authors more clearly articulate the key findings and claims of the mechanistic analyses in Section 4, and explain how these analyses support the paper's central insight regarding PI Effectiveness?

We thank the reviewer for this thoughtful comment. We would like to clarify that the purpose of Section 4 is not to establish a complete causal mechanism explaining the distillation differences across all PI conditions, but rather to investigate, from a student-side behavioral perspective, what visual evidence utilization behaviors are retained by PI-free students after training under different PI conditions. As a fine-grained perception task, FGP requires the utilization of visual evidence from target regions to answer correctly. Accordingly, we conduct two complementary analyses, Visual Overlap and Matched Occlusion, to examine whether students' visual attribution concentrates on task-relevant regions and whether the visual evidence in these regions actually affects model predictions.

1. Figure 2: Visual Attribution Reflects Differences in PI Modalities.

In FGP, ANSWER, TRAJECTORY, and EXPERIENCE provide textual PI, whereas HINT provides visual crops centered on question-relevant regions. Since Visual Overlap measures the concentration of a model's visual attribution over image regions, it primarily reflects the student's spatial localization behavior.

Specifically, ANSWER, as the most direct form of textual PI, provides the teacher with the correct answer but does not explicitly indicate where the supporting visual evidence is located in the image. HINT, in contrast, directly provides localized visual evidence. Although students trained with ANSWER also need to use the image to answer correctly, this does not necessarily mean that they develop the same target-region localization behavior. Our results show that, among the evaluated PI conditions, only HINT exhibits a clear enhancement in target-region attribution. This highlights the advantage of visual HINT over textual PI in promoting the target-region localization behavior required by FGP.

2. Figure 3: Target-Region Occlusion Verifies the Actual Impact of Visual Evidence on Predictions.

We note that sensitivity to target-region occlusion is expected for models that rely on task-relevant visual evidence in FGP. Our analysis aims to examine whether the relevant visual regions indeed contain evidence that affects model predictions. It complements Figure 2: while Figure 2 examines whether students attribute their visual attention to the target region during inference, Figure 3 further verifies the importance of these target regions for answering the questions correctly.

3. Overall Findings and Their Relationship to PI Effectiveness.

Importantly, our OPSD objective matches the next-token distributions between the teacher and student, without explicitly distilling or aligning Attention Maps. Therefore, the visual attribution patterns analyzed in Section 4 are not directly supervised during training, but are behaviors observed after OPSD training. Furthermore, all analyses are conducted on the held-out FGP test set, where students receive only the original images and questions, without access to any PI. Thus, we examine the visual behaviors actually retained by the resulting PI-free students after training under different PI conditions.

The relevance of these analyses to PI Effectiveness is that they provide a complementary characterization of the learning differences induced by different PI conditions from a student-side behavioral perspective. Figure 2 reveals differences in students' visual attribution behaviors, while Figure 3 examines the actual impact of the corresponding visual evidence on predictions. Considering the requirements of FGP, these results suggest that visual HINT, compared with the evaluated textual PI conditions, is more effective in promoting the target-region localization behavior required by the task. This also indicates that even when the teacher has access to the correct answer provided by ANSWER, it does not necessarily mean that the student will develop the same visual evidence utilization behavior as a student trained with visual HINT after OPSD.

Therefore, Section 4 characterizes the differences in student-side learning outcomes associated with different PI conditions, providing complementary evidence for understanding the transfer of PI benefits from teacher-side information access to student-side learning gains.

AuthorsOpenReview ↗

A friendly reminder

Dear Reviewer t7jy,

Thanks again for your valuable comments and precious time. We genuinely hope you could have a look at the new results and clarifications and kindly let us know if they have addressed your concerns. We would appreciate the opportunity to engage further if needed.

Best,

Authors of Paper 16735

Reviewer_4D7ROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies privileged information (PI) in multimodal on-policy self-distillation (OPSD). It separates two quantities: PI access gain, the improvement of a fixed base model when it is directly given PI, and PI effectiveness, the improvement of the PI-free student after OPSD relative to a matched No-PI OPSD baseline. In a controlled fine-grained perception (FGP) setting with Qwen3.5-4B, the authors find that ANSWER yields the largest access gain while HINT yields the largest effectiveness. Attention-rollout and matched-occlusion analyses suggest that the HINT student attends to the target region. Extending to knowledge-intensive perception, multi-hop spatial reasoning and geometric reasoning, the paper reports that the ranking of PI conditions changes across tasks, and proposes a task-level PI routing strategy selected on validation data.

Critical strengths
  1. The question is very interesting. Separating the direct benefit of PI to the teacher from the benefit retained by the PI-free student is a useful lens for OPSD.
  2. The matched No-PI OPSD control is a sensible way to isolate the incremental contribution of PI from the effect of self-distillation itself.
  3. The paper covers several PI forms (multimodal, textual and visual) and different task families which is very nice. The paper also reports external benchmarks in addition to in-domain results.
Critical weaknesses
  1. Single base model. All experiments use one model, Qwen3.5-4B. However, reported from existing literatures as well, different model families and scales may behave very differently, especially since PI access gain depends directly on the capability of the base model. This is the biggest concern i have: the central claim of the paper is that PI effectiveness is condition-dependent and task-dependt, yet the most basic condition, the base model, is never varied. If there are different models, checkpoints and size (model parameterization), would the central claim of the paper (PI gain does not indicate PI effectiveness) remains hold? The related work cited by the authors (https://arxiv.org/pdf/2609.20612) also reports that the benefit of privileged references varies across models. In its current form the study reads as a case study of one model, which makes it hard to draw generalizable conclusions.
  2. Checkpoint selection. All main results use a fixed step-100 checkpoint, with no justification and no validation-based selection. The conclusions appear sensitive to this choice. This choice requires more justification and what will happen if train longer? Table 8 it shows step. 200 but there is not No-PI OPSD reference included.
  3. No explanation of the main phenomenon. If the observed mismatch between PI access gain and PI effectiveness is real (need to first answer my raised 1st weekness), why does it occur? The paper characterizes the students but does not propose or test any hypothesis. I think a hypothesis and some validation experiments are important to make the paper complete and to let it extend beyond a phenomenon on the surface.
  4. The HINT analysis is somewhat circular. HINT gives the teacher exactly the target region (if i understand correctly), so it is expected that the resulting student attends more to that region, right? (figure 2 and 3) These analyses explain why HINT works well on FGP, but they do not address the main question of why access gain and effectiveness diverge.
  5. Clarity of the method section. The method needs a careful rewrite. For example, pi_T is used in the OPD objective (Sec. 2.1) before it is defined, and only introduced in Sec. 2.2. The symbol T is used both for the sequence length and as the teacher subscript in the same equation. The method part writing is confusing.
  6. Trade-off with general capabilities. Almost all PI-conditioned students perform worse than the original backbone on external benchmarks. Does this indicate like a catastrophic forgetting? It would be more helpful to talk about this trade-off and the potential cause of it.
Questions for authors
  1. Does the ANSWER vs HINT reversal on FGP hold for other model families or for different model sizes?
  2. Why was step 100 chosen? Could the authors report PI effectiveness along the full training trajectory, and add the No-PI OPSD reference for GeoR at step 200?
  3. What do the authors believe causes the mismatch between PI access gain and PI effectiveness? Can an experiment be designed to distinguish between candidate explanations?
  4. Why does No-PI OPSD underperform the backbone on FGP and GeoR, and how does PI-conditioned OPSD compare with SFT or GRPO in terms of the trade-off between in-domain gains and general capabilities?
7 comments
AuthorsOpenReview ↗

Response 1

Dear Reviewer 4D7R,

W1: Single base model. All experiments use one model, Qwen3.5-4B. [...] In its current form, the study reads more like a case study of a single model, making it difficult to draw generalizable conclusions.

We thank the reviewer for pointing out that model scale and model family may affect the relationship between PI Access Gain and PI Effectiveness. We acknowledge that our current experiments are primarily based on Qwen3.5-4B, without systematic evaluation across different model families and scales. Due to limited computational resources, we are unable to conduct additional comprehensive cross-model experiments.

By keeping the backbone fixed while systematically varying PI conditions and task requirements, we examine the relationship between teacher-side PI Access Gain and student-side PI Effectiveness, and reveal that the two do not necessarily align. We believe this distinction, together with our cross-task experimental results, provides valuable insights into understanding and evaluating PI in multimodal OPSD. Further systematic investigation of PI Effectiveness across different model families and scales remains an important direction for future work.

W2: Checkpoint selection. All main results use a fixed step-100 checkpoint, with no justification and no validation-based selection. The conclusions appear sensitive to this choice. This choice requires more justification and what will happen if training is continued for longer? [...]

Q2: Why was step 100 chosen? Could the authors report PI Effectiveness along the full training trajectory, and add the No-PI OPSD reference for GeoR at step 200?

We thank the reviewer for raising this concern.

1. Rationale for Selecting Step-100.

Step-100 was adopted as a fixed checkpoint for all experiments, rather than selected as the best checkpoint based on validation performance. Under our current training setup, the model is already close to convergence after 100 steps. We chose this checkpoint to balance sufficient training with computational efficiency and the potential risk of overfitting from prolonged training, while ensuring a consistent training budget across different PI conditions.

2. Step-200 Results and Additional No-PI Baseline.

To further examine the effects of longer training, we additionally report the step-200 results for FGP and GeoR in Tables 6 and 7, respectively, in the appendix. Overall, the main experimental findings remain consistent at step-200. Furthermore, to address the missing baseline pointed out by the reviewer, we have added the GeoR step-200 No-PI OPSD results to Table 7, providing the corresponding reference for comparison.

W3: No explanation of the main phenomenon. If the observed mismatch between PI Access Gain and PI Effectiveness is real, why does it occur? The paper characterizes the students but does not propose or test any hypothesis. I think a hypothesis and some validation experiments are important to make the paper complete and to let it extend beyond a phenomenon on the surface.

W4: The HINT analysis is somewhat circular. If I understand correctly, HINT gives the teacher exactly the target region, so it is expected that the resulting student attends more to that region, right? These analyses explain why HINT works well on FGP, but they do not address the main question of why access gain and effectiveness diverge.

Q3: What do the authors believe causes the mismatch between PI Access Gain and PI Effectiveness? Can an experiment be designed to distinguish between candidate explanations?

We thank the reviewer for raising this important question. To further understand the discrepancy between PI Access Gain and PI Effectiveness, we investigate two complementary aspects: teacher-side conditional supervision and student-side visual behavior. One possible explanation is that a teacher's ability to obtain the correct answer through PI does not necessarily mean that it can provide effective corrective supervision under the actual prefixes generated by the student. Meanwhile, different PI conditions may also influence the task-relevant behaviors ultimately retained by the student. To investigate these aspects, we conduct additional EMA Teacher Supervision experiments and further clarify the behavioral analyses in Section 4.

1. Teacher-side Conditional Supervision: EMA Teacher Supervision Analysis (W3, Q3)

During actual OPSD training, the EMA teacher provides supervision conditioned on student-generated prefixes. Therefore, the teacher's direct answerability with PI is not necessarily equivalent to its conditional supervision performance under the student's actual generation states.

AuthorsOpenReview ↗

Response 2

To investigate this difference, we use the step-100 Answer and Hint EMA checkpoints with their corresponding PI inputs to conduct paired evaluations on the original rollout prefixes from 624 FGP questions. We further group these prefixes according to whether the original rollout produced a correct or incorrect final answer.

Table 1: Comparison of Answer EMA and Hint EMA on Original Rollout Prefixes.

Prefix GroupCorrect TrajectoriesIncorrect TrajectoriesAll Trajectories
Original Rollouts+0.0254 (302)+0.4904 (322)+0.2653 (624)

Values report the average ΔlogP = logP(Answer EMA) − logP(Hint EMA) for the correct-answer continuation under the same prefixes. The numbers before the slash indicate the number of evaluated trajectories. Positive ΔlogP values indicate that Answer EMA assigns a higher log-probability to the correct-answer continuation than Hint EMA.

The results show that Answer EMA achieves higher average log-probabilities for correct-answer continuations across all three groups, consistent with the overall trend of PI Access Gain.

We further examine 18 trajectories from 12 questions containing visually verifiable errors. Among the 17 valid independent answer positions, Answer EMA provides greater relative support for the correct option than Hint EMA in 16/17 positions. However, in 16/17 positions, Answer EMA still assigns a lower probability to the correct option than to the incorrect option selected by the student.

This finding indicates that even stronger relative support for the correct answer does not necessarily mean that the teacher can effectively correct the student's current erroneous state.

2. Student-side Visual Behavior: Behavioral Transfer of HINT (W3, W4)

Our OPSD objective matches the next-token distributions between the teacher and student, without explicitly distilling or aligning Attention Maps. Therefore, the visual attribution patterns analyzed in Section 4 are not directly supervised during training, but are behaviors observed after OPSD training.

Furthermore, all analyses are conducted on the held-out FGP test set, where students receive only the original images and questions, without access to any PI. Thus, we examine the visual behaviors actually retained by PI-free students after training under different PI conditions.

Correctly solving FGP requires utilizing fine-grained visual evidence from the target region. Figure 2 shows that, among the evaluated PI conditions, only HINT leads to a noticeable enhancement in target-region visual attribution, whereas ANSWER does not exhibit a comparable change. The newly added matched No-PI control further shows that Hint-trained students exhibit substantially stronger target-region attribution than No-PI students, both at generation start and immediately before the final answer.

Figure 3 further examines the actual impact of target-region visual evidence on model predictions through Matched Occlusion analysis. The results show that blurring the target region substantially reduces model accuracy, whereas matched background blur has a much smaller effect. This indicates that the target region contains important visual evidence required for FGP.

Therefore, our analysis does not merely show that students attend to the target region after HINT provides it to the teacher. Rather, it reveals that, without explicit Attention Distillation and without access to PI during testing, students trained with HINT retain stronger target-region visual attribution. In contrast, although ANSWER provides the teacher with the correct answer, it does not lead to a comparable enhancement in the student's visual behavior.

3. Relationship to PI Effectiveness

Taken together, the EMA experiments reveal the distinction between direct answerability and conditional supervision from the teacher's perspective, while Section 4 characterizes the visual behaviors retained by students trained under different PI conditions.

On FGP, although ANSWER achieves higher PI Access Gain, HINT yields higher PI Effectiveness (+19.87 vs. +10.74 percentage points). Combined with the above analyses, we suggest that the advantage of HINT lies not only in providing additional visual information to the teacher, but also in helping the student retain the task-relevant visual evidence utilization behavior required by FGP.

These experiments provide complementary empirical evidence from teacher-side supervision and student-side behavior, offering further insights into the discrepancy between PI Access Gain and PI Effectiveness.

AuthorsOpenReview ↗

Response 3

W6: Trade-off with general capabilities. Almost all PI-conditioned students perform worse than the original backbone on external benchmarks. Does this indicate like a catastrophic forgetting? It would be more helpful to talk about this trade-off and the potential cause of it.

Q4: Why does No-PI OPSD underperform the backbone on FGP and GeoR?

We thank the reviewer for raising this important concern. We observe that the performance changes induced by OPSD are not limited to PI-conditioned settings. Even under No-PI settings, students may perform worse than the original backbone on FGP and GeoR. This suggests that OPSD training itself may alter the original parameter state and generation behavior of the model, without necessarily improving performance across all tasks.

Furthermore, the performance degradation of PI-conditioned students on external benchmarks reflects a potential trade-off between task-specific learning gains and the preservation of general capabilities.

1. Potential Forgetting and Behavioral Changes.

To further understand this phenomenon, we examine the generation trajectories of students trained under different PI conditions. We observe that some students still produce expressions such as "Given the reference answer..." or "Given the reference trajectory..." during inference, even though no PI is provided.

Notably, these outputs do not reproduce the specific content of the PI. Instead, they retain the expression patterns used by the teacher under PI-conditioned settings. This suggests that students may learn not only the task-relevant supervision provided by PI, but also transfer response patterns associated with the training-time PI context to inference settings where the corresponding PI is unavailable.

2. Implications for General Capability Degradation.

These observations suggest that PI-conditioned OPSD may lead students to overfit to training-time conditional supervision patterns, thereby altering their original generation behavior. Such behavioral shifts may be related to the observed degradation on external benchmarks and may also reflect a certain degree of forgetting.

However, we consider this a potential explanation rather than directly attributing the observed degradation to catastrophic forgetting. Further investigation is needed to understand the long-term effects of PI-conditioned OPSD on the model's original capabilities and generation behavior.

W5: The method needs a careful rewrite. For example, πT\pi_T is used in the OPD objective (Sec. 2.1) before it is defined, and only introduced in Sec. 2.2. The symbol TT is used both for the sequence length and as the teacher subscript in the same equation. The method part writing is confusing.

We thank the reviewer for pointing out these issues. We have revised the notation in Section 2.1 to improve clarity and consistency. Specifically, we now define the teacher policy πT\pi_T before its first use in the OPD objective and replace the sequence-length notation TT with LL throughout the objective to avoid ambiguity. We have also checked the consistency of the notation in the subsequent OPSD formulation.

Reviewer_VFWgOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies whether the performance improvement a teacher obtains from privileged information (PI) translates into improvements for a student that does not have access to PI at inference time. The authors distinguish PI access gain from PI effectiveness, with the latter measured against a matched No-PI distillation baseline. They compare several PI conditions, including answers, trajectories, experience summaries, and visual hints. The main finding is that ANSWER provides the largest direct benefit to the teacher in fine-grained perception, while HINT produces the largest student improvement. The preferred PI condition also varies across tasks. The paper further examines students' use of relevant visual regions and introduces a task-level PI routing strategy based on validation performance.

Critical strengths
  1. The research question is clear and practically relevant. A teacher's ability to answer more accurately with additional information does not guarantee that the student will learn more from its supervision.

  2. The experiments report both benefits and limitations. The evaluation covers four task families and multiple external benchmarks, including negative results on geometric reasoning and broader generalization.

Critical weaknesses
  1. The additional insight beyond prior work is not sufficiently clear. ViCuR[1] already discusses why visual evidence recoverable from the student's original input can be useful privileged information. Zhang et al.[2] already analyzes the additional contribution of PI using matched reference-free distillation controls. The main conclusion remains an empirical observation that different tasks favor different PI conditions. The paper does not yet offer a more specific explanation that would help readers choose PI for a new task.

  2. The connection between student behavior and the incremental benefit of PI is incomplete. PI access gain measures complete answers from a fixed base model, while training uses an evolving EMA teacher on student-generated prefixes. High accuracy with ANSWER does not establish that the teacher provides better corrections during training, leaving the link to useful student supervision unclear.

[1] ViCuR: Visual Cues as Recoverable Privilege for Multimodal On-Policy Distillation [2] What Does Privileged Information Add to On-Policy Self-Distillation?

Questions for authors
  1. When a student has already generated incorrect intermediate content, does a teacher given the correct answer still provide more useful corrections than a teacher given a visual hint?
  2. How would the matched No-PI model perform in Figures 2 and 3? With this control included, which changes in visual behavior could be attributed specifically to PI?
  3. Is the ranking reversal between FGP and KIP mainly caused by their different task requirements, or by the fact that the KIP crop provides almost no additional help to the teacher?
  4. How does the routing strategy compare with using a mixture of all three conditions for every task, without selecting conditions by task?
9 comments
AuthorsOpenReview ↗

Response 1

Dear Reviewer VFWg,

We sincerely thank you for your careful review and constructive feedback. We particularly appreciate your concerns regarding the novelty and practical implications of our work. Your comments have helped us clarify the central research question of this study and its distinctions from prior work. We respond to your concerns below.

W1: The additional insight beyond prior work is not sufficiently clear. ViCuR [1] already discusses why visual evidence recoverable from the student's original input can be useful privileged information. Zhang et al. [2] already analyzes the additional contribution of PI using matched reference-free distillation controls. The main conclusion remains an empirical observation that different tasks favor different PI conditions. The paper does not yet offer a more specific explanation that would help readers choose PI for a new task.

1. Our Contributions

We acknowledge that the observation that different tasks favor different PI conditions is one of our empirical findings and contributions, but it does not encompass the full scope of our work. Our study centers on a fundamental question: Can the direct benefits of PI access to the teacher translate into incremental learning gains retained by the PI-free student? Our contributions to this question are threefold:

(1) PI Access Gain vs. PI Effectiveness: We distinguish the teacher-side benefits of accessing PI from the student-side learning gains retained after training, and reveal that the two may not align.

(2) Student-side Behavioral Analysis: We investigate how different PI conditions influence students' utilization of task-relevant visual evidence through visual attribution and occlusion analyses.

(3) Cross-task PI Effectiveness Analysis: Across four multimodal tasks with different capability requirements, we systematically examine how PI effectiveness depends on task requirements and concrete PI conditions, and explore its implications for PI selection.

Our cross-task analysis also provides preliminary guidance for selecting PI in new tasks. Our results suggest that PI selection should account for whether the information provided matches the actual requirements of the target task, rather than relying solely on the form of PI. For example, FGP primarily requires fine-grained visual discrimination, making visual HINT directly useful for providing visual guidance. In contrast, although crop-based HINT in KIP helps the teacher localize the target region, it does not provide the knowledge needed to interpret the identified content, resulting in relatively limited gains. In MSR, even a Spatial HINT specifically designed for the task does not necessarily yield greater student gains.

Building on these findings, we further explore Task-level PI Routing, which provides an initial demonstration of the potential value of task-specific PI selection. We acknowledge that our current routing strategy remains a simple exploration based on a limited set of candidate conditions and validation data. Developing more reliable and generalizable task–PI matching strategies remains an important direction for future work.

2. Relationship to Prior Work

(1) Relationship to ViCuR.

ViCuR studies Recoverable Visual Privilege by providing the teacher with textual visual cues (denoted as SS) that describe question-relevant visual evidence. Its central idea is that the underlying visual evidence described by these cues remains available in the student's original image, allowing the student to recover and utilize this information. Thus, ViCuR primarily investigates the recoverability of textual visual cues and their utilization by the student.

Our study differs in scope. We investigate not only textual PI but also visual PI directly provided to the teacher, including localized image crops, spatial annotations, and auxiliary geometric constructions. We further compare these with textual PI conditions such as ANSWER, TRAJECTORY, and EXPERIENCE. While ViCuR focuses on Recoverable Visual Privilege, our study systematically investigates the role of PI in multimodal OPSD across different PI forms, student-side behaviors, and cross-task effectiveness.

(2) Relationship to Zhang et al.

Zhang et al. is a concurrent work that primarily investigates the additional contributions of different answer-matched textual reference representations over reference-free distillation in mathematical reasoning. We acknowledge that matched No-PI OPSD is an essential experimental control for measuring the incremental benefits of PI, but it is an experimental setting in our study rather than our central contribution.

AuthorsOpenReview ↗

Response 2

In contrast, we extend the investigation to multimodal OPSD, systematically comparing different forms of textual and visual PI, distinguishing teacher-side PI Access Gain from student-side PI Effectiveness, and examining the transfer of PI benefits and its task dependence through student-side behavioral analyses and cross-task experiments.

3. Summary

In summary, our work goes beyond the empirical observation that different tasks favor different PI conditions. We systematically investigate the transfer of PI benefits from teacher-side access to student-side learning, examining this process through student visual behavior and cross-task analyses, while also exploring its practical implications for task-level PI selection. We believe these aspects constitute the main additional insights of our work beyond the aforementioned studies.

W2: The connection between student behavior and the incremental benefit of PI is incomplete. PI access gain measures complete answers from a fixed base model, while training uses an evolving EMA teacher on student-generated prefixes. High accuracy with ANSWER does not establish that the teacher provides better corrections during training, leaving the link to useful student supervision unclear.

Q1: When a student has already generated incorrect intermediate content, does a teacher given the correct answer still provide more useful corrections than a teacher given a visual hint?

We thank the reviewer for pointing out the insufficient connection between teacher-side supervision and student-side learning gains. We further investigate this issue through two complementary aspects: the conditional supervision provided by EMA teachers on student-generated prefixes and the visual behaviors ultimately retained by students after training under different PI conditions.

1. EMA Teacher Supervision: Conditional Supervision and Error Correction under the Same Student Prefixes (W2, Q1)

During actual OPSD training, the EMA teacher provides supervision conditioned on student-generated prefixes. Therefore, the performance gain obtained by a fixed base model with direct access to PI does not necessarily reflect the conditional supervision provided by the teacher during training.

To investigate this difference, we use the step-100 Answer and Hint EMA checkpoints, together with their corresponding PI inputs, to conduct paired evaluations on the same set of Backbone-generated rollout prefixes from 624 FGP questions. We further group these prefixes according to whether the original rollout produced a correct or incorrect final answer.

Table 1: Comparison of Answer EMA and Hint EMA on Backbone-generated Prefixes.

Prefix GroupNumberΔlogP
Correct final answers302+0.0254
Incorrect final answers322+0.4904
All prefixes624+0.2653

Here, ΔlogP = logP(Answer EMA) − logP(Hint EMA), representing the difference in average log-probability assigned by the two EMA teachers to the correct-answer continuation. Positive values indicate a higher log-probability assigned by Answer EMA.

The results show that Answer EMA assigns higher average log-probabilities to correct-answer continuations across all groups, including prefixes from rollouts with incorrect final answers. This indicates that ANSWER still provides a stronger conditional supervision signal favoring the correct answer, even on prefixes where the student has produced an incorrect final answer.

However, whether stronger support for the correct answer translates into effective error correction requires further examination. We therefore inspect 18 trajectories from 12 questions containing visually verifiable errors. Among the 17 positions with valid independent answer choices, Answer EMA provides stronger relative support for the correct option than Hint EMA in 16/17 positions. Nevertheless, in 16/17 positions, Answer EMA still assigns a lower probability to the correct option than to the incorrect option originally selected by the student. Hint EMA exhibits the same behavior in all 17 positions.

These results indicate that although ANSWER provides stronger relative signals favoring the correct answer than HINT on erroneous prefixes, this advantage is insufficient to shift the teacher's preference toward the correct option in most of the examined positions. Therefore, stronger conditional support for the correct answer is not equivalent to effective corrective supervision.

AuthorsOpenReview ↗

Response 3

2. Student-side Visual Behavior and Its Relationship to PI Effectiveness (W2)

Our OPSD objective matches the teacher and student next-token distributions without explicitly distilling or aligning Attention Maps. All visual behavior analyses are conducted on the held-out FGP test set, where students receive only the original images and questions, without access to PI.

Among the evaluated PI conditions, only HINT leads to a noticeable enhancement in target-region visual attribution. Visual Overlap characterizes whether students concentrate their visual attribution on task-relevant target regions, while Matched Occlusion further examines whether visual evidence in these regions actually affects model predictions by comparing target-region occlusion with matched background occlusion. The results show that blurring the target region substantially reduces accuracy, whereas matched background blur has little effect.

These two analyses are complementary, examining visual localization and the actual contribution of visual evidence, respectively. Together, they indicate that HINT is associated with stronger target-region localization and visual evidence utilization retained by students during PI-free inference. In contrast, although ANSWER provides the teacher with the correct answer, it does not lead to a comparable enhancement in student visual attribution.

3. Overall Findings

Taken together, the EMA experiments show that stronger conditional support for correct answers from the teacher does not necessarily translate into effective corrective supervision. Meanwhile, the student-side analyses indicate that HINT better facilitates the retention of the target-region localization behavior required by FGP.

These findings suggest that PI effectiveness depends not only on the teacher's ability to support the correct answer, but also on whether the supervision helps the student acquire and retain task-relevant behaviors.

Q2: How would the matched No-PI model perform in Figures 2 and 3? With this control included, which changes in visual behavior could be attributed specifically to PI?

We thank the reviewer for suggesting this control. We have added the matched step-100 No-PI OPSD baseline to both the visual-overlap and matched-occlusion analyses in Appendix E, using the same 624 held-out FGP questions and the original evaluation protocols.

With this control, the visual-attribution analysis shows that Hint leads to substantially stronger target-region attribution than No-PI, both at generation start (0.2718 vs. 0.1096) and immediately before the final answer (0.2734 vs. 0.1099). Answer shows no comparable increase. This indicates that the enhanced target-region attribution is specifically associated with Hint PI.

The occlusion analysis provides complementary evidence: target-region blur substantially reduces accuracy for both Hint and No-PI students, whereas matched background blur has little effect. Specifically, No-PI accuracy is 44.87% on the original image, 44.71% under background blur, and 29.17% under target-region blur. The importance of the target region is not unique to PI. The PI-associated difference is that Hint-trained students exhibit stronger attribution to the target region.

Q3: Is the ranking reversal between FGP and KIP mainly caused by their different task requirements, or by the fact that the KIP crop provides almost no additional help to the teacher?

We thank the reviewer for raising this question. We attribute the ranking reversal primarily to the different task requirements. In KIP, HINT is constructed by cropping the question-specific target-object region using the dataset-provided bounding box. This provides the teacher with explicit localization of the relevant visual content, helping it focus on the target object.

The key difference lies in what is required beyond visual localization. FGP primarily requires resolving fine-grained visual details, which directly benefits from improved visual access. KIP, however, additionally requires external or commonsense knowledge to interpret the identified content and derive the answer. Thus, although the crop provides useful visual grounding, it does not supply the knowledge needed to complete the task. Thus, the KIP crop does provide additional help to the teacher by explicitly localizing the relevant visual content, but this assistance is limited to visual grounding and translates into only a modest PI access gain, as KIP additionally requires knowledge-based interpretation to derive the answer.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary for Submission 16735

Dear PC, AC and Reviewers,

We sincerely thank all reviewers for their careful assessment and constructive feedback. We would like to further highlight the broader scientific findings and research implications of our work, which we believe are valuable for the development of visual OPSD.

1. Scientific Insights Beyond Existing Visual OPSD Work

Existing research on visual OPSD has often focused on specific PI designs and their mechanisms on individual datasets. While these studies provide valuable insights into individual PI designs, their dataset-specific findings make it difficult to determine whether the benefits of a particular PI condition generalize across different task requirements, or to identify the factors that may explain why different PI conditions lead to different outcomes.

We believe that PI itself deserves to be treated as a subject of systematic investigation in visual OPSD. Understanding the role of PI requires looking beyond whether it improves teacher performance, and examining whether such gains translate into learning benefits retained by the student, as well as how its effectiveness varies with task requirements. Motivated by this perspective, we distinguish between teacher-side PI Access Gain and student-side PI Effectiveness, examine student visual processing behaviors, and compare different PI conditions across multiple task families.

We hope this study encourages future visual OPSD research to pay greater attention to PI itself, through systematic comparisons of different PI conditions across diverse datasets and task requirements, together with analyses at multiple levels. We believe that understanding PI itself is an important direction for advancing visual OPSD.

2. Core Contributions

Overall, our work makes three main contributions:

  • A new perspective on PI evaluation: We distinguish teacher-side PI Access Gain from student-side PI Effectiveness, showing that the two are not necessarily aligned.
  • A systematic study of PI effectiveness: We compare different textual and visual PI conditions across four visual task families using matched No-PI OPSD baselines, investigating the task-dependent effectiveness of PI.
  • Multi-level analysis and application exploration: We investigate teacher-side conditional supervision and student-side visual evidence utilization, and further explore task-level PI Routing as a preliminary application to multi-task training.

3. Additional Experiments and Rebuttal Updates

Building on the reviewers' valuable feedback, we conducted additional experiments and analyses and further refined the manuscript:

  • EMA Teacher Supervision: We conducted paired evaluations of Answer EMA and Hint EMA on the same student-generated prefixes from 624 FGP questions. The results show that Answer EMA assigns higher average log-probabilities to correct-answer continuations than Hint EMA, including on prefixes from incorrect rollouts. However, among 17 positions with valid independent answer choices in visually verifiable erroneous trajectories, Answer EMA still assigns a lower probability to the correct option than to the incorrect option originally selected by the student in 16 positions. This indicates that although ANSWER provides stronger relative support for the correct answer than HINT, this advantage is generally insufficient to shift the teacher's preference toward the correct option.
  • Matched No-PI and Visual Behavior Analyses: We added the matched step-100 No-PI baseline to the visual attribution and matched occlusion analyses, clarifying which behavioral differences are associated with PI conditions.
  • Layer-wise Visual Processing Analysis: We further highlight the residual-attention rollout analysis presented in Appendix Fig. 7, which reveals distinct visual processing patterns across network layers for students trained with HINT and textual PI.
  • PI Routing: We added a Uniform Mixed-PI baseline under the same training budget. Task-level Routing achieves a 2.00-percentage-point improvement in average accuracy over this baseline. We view this experiment as a preliminary validation of task-specific PI selection, and hope this exploration can provide an initial contribution toward advancing OPSD in multi-task settings.
  • Additional Experiments and Manuscript Improvements: We added step-200 results and the previously missing GeoR step-200 No-PI baseline, revised the method notation, and further discussed the potential trade-off between task-specific learning gains and general capabilities.

We hope this work will encourage future visual OPSD research to move beyond individual PI designs and toward a more systematic understanding of how different forms of PI interact with different task requirements.

We sincerely thank all reviewers and Area Chairs for their time and consideration.

Best regards,

The authors of Submission 16735

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.