acceptodds
Under review as a conference paper at ICLR 2027

VISTA: Internalizing Collective Visual Experience via On-Policy Distillation for Active Multimodal Agents

Abstract

Active multimodal agents use visual tools to acquire task-relevant evidence while reasoning. Although reinforcement learning samples multiple interaction trajectories per input, outcome-based objectives primarily use the group to estimate scalar advantages, leaving complementary visual discoveries underused. We introduce VISTA, which internalizes collective visual experience through on-policy distillation by turning observations from same-input rollouts into shared supervision. *Collective visual experience distillation* (CVED) organizes these observations with their interaction context and aligns them with individual decisions, while *heterogeneity-aware policy improvement* (HAPI) reinforces successful trajectories and provides experience-guided distillation for unsuccessful attempts. An experience-conditioned teacher evaluates the student's sampled response prefixes, allowing discoveries from one trajectory to guide learning in another without replacing the student's original history or generating new target trajectories. The trained agent retains its visual tools and acts using its own interaction history. VISTA achieves the strongest average performance among the evaluated active multimodal agents of comparable size and consistently outperforms same-backbone training baselines across fine-grained perception and general reasoning tasks, demonstrating the value of collective experience for active multimodal learning.

Primary area:
foundation or frontier models, including LLMs
Keywords:
on policy distillation, multimodal large language model

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

This paper proposes VISTA, a training framework for active multimodal agents that shares visual observations across same-input rollouts. CVED constructs decision-aligned privileged teacher contexts and distills teacher distributions along student-generated response prefixes. HAPI reinforces successful trajectories and applies distillation to unsuccessful ones. The trained agent retains its visual tools and uses only its own interaction history at inference.

Critical strengths

The paper is generally clear and offers a useful perspective on reusing visual discoveries beyond scalar outcome rewards. Controlled comparisons show consistent improvements across perception and reasoning benchmarks and several Qwen backbones. The structured-context, self-only hindsight, and crossed-routing ablations provide informative evidence for the value of collective experience and differentiated supervision.

Critical weaknesses
  1. Important experience-construction details are delayed to the appendix. A concise main-text description and a concrete example would substantially improve accessibility. Please show the source trajectories, selected crops and relations, and the instantiated teacher history for one student decision, including useful evidence contributed by an unsuccessful trajectory.
  2. The mechanism behind HAPI's advantage remains only partly understood. Uniform Hybrid differs from HAPI in both reinforcement and distillation routing, not merely in whether successful trajectories receive distillation. Table 11 usefully separates these masks, but changing the distillation subset also changes its token denominator and effective weighting. Controls matching loss scales, together with gradient-agreement or teacher-disagreement analysis, would clarify why joint routing helps and whether distillation interferes with successful behavior.
  3. Experience selection and teacher reliability deserve further validation. It would be useful to measure teacher quality when all rollouts fail or provide redundant evidence, and whether privileged hindsight encourages answering before the student has acquired sufficient evidence. These diagnostics would help characterize when collective experience supplies reliable corrective guidance.
  4. The practical cost of the gains are insufficiently quantified. Training schedules are matched, but teacher inference and experience construction introduce additional computation. Please report GPU-hours, peak memory, and performance under matched total compute. Although results are averaged over three seeds, standard deviations or confidence intervals would help assess modest differences between routing and teacher-update variants.
Questions for authors

See weakness.

3 comments
AuthorsOpenReview ↗

Response to tGQm (1/2): Concrete construction and routing evidence

We sincerely thank you for your thoughtful feedback and helpful suggestions on explaining and evaluating collective experience. In response, we have clarified its construction with a worked example and added a common-denominator routing control, along with analyses of teacher guidance, observation behavior, and training cost. We have incorporated these additions into the revised manuscript and provide detailed responses below.

1. A concise construction description and worked example. To make the construction easier to follow, Section 3.2 now describes each step: deduplicate crop images while retaining source occurrences; select up to B=4 artifacts including required crop ancestors; and serialize producing actions, artifact relations, decision-relative visibility, and categorical outcomes. Generated final answers and reference answers are excluded from the added teacher context.

The Fullerton example illustrates the missing evidence. A crop from an unsuccessful student trajectory contains "THE FULLERTON" but omits "HOTEL," the word distinguishing the answer options. A peer first localizes the building and then captures the full inscription. At the same student-sampled prefix, the teacher sees the original student history plus the peer views and their acquisition context; the student still sees only its original history. JSD transfers the resulting conditional distribution without replacing the sampled trajectory.

Appendix H.4 provides the source observations, the actions that produced them, their ancestry relationships, and a readable schematic of the teacher context for this example. The source-pool experiment provides complementary evidence for the value of observations from failed attempts: using full-group observations increases average accuracy from 83.29 to 83.83 compared with successful-source-only observations under the same B=4 cap. In a 200-group selector diagnostic with 765 artifacts selected per condition, 115 full-group artifacts come exclusively from unsuccessful trajectories. This complements the worked example with quantitative evidence for failed-source contributions (Appendix H.3).

2. Joint routing remains beneficial when the denominator is controlled. To address the normalization concern, we fix lambda = 0.10, the RL denominator, training data, optimization budget, and teacher updates, and normalize every distillation term by the full-group response-token count N.

Routing configurationThree-task Avg. ± sample SD
Uniform hybrid, N82.87 ± 0.12
RL mask only, N82.55 ± 0.18
Distillation mask only, N83.22 ± 0.11
HAPI, shared N83.67 ± 0.17
HAPI, original N-minus83.83 ± 0.22

Joint routing exceeds Uniform by 0.80 points and also outperforms both single-mask controls. The original denominator adds 0.16 points. The revised main-text analysis and Appendix H.1 (Table 12) report this control alongside the original factorial ablation.

The routing design assigns complementary roles to the two learning signals: verified successful behavior receives outcome-based reinforcement, while unsuccessful attempts receive token-level guidance informed by the group's additional observations. Since teacher and student condition on different information, they can prefer different continuations even along a successful trajectory. HAPI directs this corrective signal to unsuccessful attempts while preserving reinforcement of verified success. The common-denominator experiment supports this joint allocation without requiring the assumption that every teacher disagreement is harmful.

AuthorsOpenReview ↗

Response to tGQm (2/2): Teacher guidance, premature answers, and cost

3. Teacher guidance from unsuccessful or redundant experience. An unsuccessful trajectory can still contain relevant evidence, even if the student fails to use it to produce a correct final answer. CVED preserves useful observations even from unsuccessful sources, and source/action relations allow the teacher to interpret them jointly. With binary rewards, an all-failure group has zero group-relative task-reward advantage, whereas complementary observations can still provide a distillation signal. The successful-source-only versus full-group comparison in Appendix H.3 empirically supports retaining failed-source evidence in mixed-success groups.

The key condition is whether the shared observations supply evidence missing from the student's history. Redundant or insufficient observations may leave that gap unresolved. Categorical feedback and acquisition context help the teacher interpret the available evidence, but do not guarantee correct guidance. Appendix H.3 clarifies this condition: collective experience can reuse discoveries that outcome-only supervision leaves underused, with its value depending on the information those discoveries provide.

4. Premature answering under CVED and Oracle ROI. We compare the two teacher-context variants from Table 3 on the same 100-question subset requiring additional evidence. HAPI routing and the student interface are fixed, and necessary evidence is annotated independently of the methods. A response is premature if the student answers before acquiring that evidence.

Teacher contextPremature answersPremature and incorrect
Oracle ROI18/10015/100
CVED10/1007/100

On this subset, CVED reduces both rates by eight percentage points. The teacher receives collective observations with acquisition context and decision-relative alignment; oracle crops remain training-only in the Oracle ROI control. This comparison tests the complete teacher-context design, while Table 3's identical-observation CVED-versus-Unstructured comparison tests evidence organization. Appendix H.3 (Table 17) reports the diagnostic and its definitions.

5. Variability and teacher updates. Appendix H.5 provides task-level S1-S3 results and sample SD for the key ablations. HAPI exceeds Uniform in all three original runs (83.83 ± 0.22 versus 82.87 ± 0.12). For teacher updates, the six-task perception average is 73.81 ± 0.19 for EMA and 73.53 ± 0.59 for the frozen teacher. EMA's modest average advantage is task-dependent: it improves V*Bench/ZoomBench, while the frozen teacher scores higher on HR-8K/MME-RW-CN. Both teachers receive newly collected experience, so this comparison measures the additional effect of updating teacher weights.

6. Cost and common-budget performance. On eight H100 GPUs, Qwen3-VL-4B training uses 150/165/180 GPU-hours for GRPO/OPSD/VISTA, with similar peak memory of approximately 80 GB per GPU. Final three-visual-task averages are 80.57/81.61/83.83. At a common 150-GPU-hour cap they are 80.57/81.53/83.75, so VISTA retains gains of 3.18/2.22 points over GRPO/OPSD. At 100 GPU-hours, VISTA already reaches 82.09. These are three-run means using the last checkpoint within each cap, with the same initialization, data, and tool interface. Section 4.9 and Appendix H.2 (Table 13 and Figure 6) report the full comparison.

The new results and analyses are highlighted in blue in the revised manuscript.

We hope these responses and new results have addressed your concerns. Thank you again for your valuable suggestions and for helping us improve the manuscript. We look forward to your further feedback!

AuthorsOpenReview ↗

Follow-up on our response to Reviewer tGQm

Dear Reviewer tGQm,

We appreciate your constructive feedback on our submission. We have uploaded our responses and additional analyses addressing each of your points. As the discussion phase is drawing to a close, we would welcome your thoughts on whether our reply resolves your concerns, and we remain happy to provide any further clarification.

Thank you again for your time,

The authors

Reviewer_7ZKuOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

VISTA uses visual observations from multiple tool-using rollouts for the same image question, beyond their outcome rewards. CVED organizes crops and their interaction context for a training-time teacher. Teacher and student score the same student-sampled prefixes, but the student sees only its own history. HAPI applies GRPO to successful trajectories and CVED to unsuccessful ones; both contribute observations. Only the tool-using student remains at inference. Experiments cover fine-grained visual QA and general multimodal reasoning, with ablations of experience source, context organization, and loss allocation.

Critical strengths
  1. Table 10 isolates the source of visual experience. The three-benchmark average rises from 81.70% with self-only hindsight to 83.83% with full-group observations, while categorical outcome feedback, HAPI routing, and the artifact budget stay fixed. This is my main reason for leaning toward acceptance.
  2. The main design choices have controls. Table 3 compares structured and unstructured context with the same observations and budget; Tables 4 and 11 compare loss routing. Tables 2 and 7 cover different sizes and Qwen backbones.
Critical weaknesses
  1. Small gains lack variance and cost context. In Table 2, VISTA leads the strongest other distillation method by 1.31 (4B) and 0.89 (8B) points on the six-benchmark average; HAPI leads the uniform hybrid by 0.96 points in Table 4. The paper reports three-seed means but no per-seed scores or error ranges. CVED also requires teacher inference, and Appendix G notes the lack of a matched-compute comparison. Stability and cost-effectiveness remain unclear.
  2. Final-answer accuracy does not isolate active observation. The paper does not report crop relevance, tool-call counts, or how often the agent fails after seeing decisive evidence. Table 10 supports the value of peer observations, but cannot separate better crop choices from better reasoning after a crop.
Questions for authors
  1. Please report per-seed scores for Tables 2, 4, and 10 and total training costs for VISTA and the strongest same-backbone baselines. Does Table 2's sub-two-point advantage persist at similar compute?
  2. Which observation behaviors change? Please relate crop relevance, useful-crop rate, or tool-call count to final-answer accuracy.
  3. In groups with both successes and failures, how does CVED perform using only successful trajectories' observations versus all observations, at the same artifact budget?
  4. With the selected crops fixed, how does accuracy change when categorical outcome feedback or interaction relations are removed separately?
3 comments
AuthorsOpenReview ↗

Response to 7ZKu (1/2): Stability and compute-controlled results

We sincerely thank you for your constructive feedback and your positive assessment of the collective-versus-self-only comparison. Following your suggestions, we have expanded the evaluation to include complete results across three runs, comparisons at common compute budgets, and analyses of observation behavior and experience sources. These additions are included in the revised manuscript and discussed below.

1. Complete three-run results for Tables 2, 4, and 10. Appendix H.5 includes every task-level score for all trained methods at both scales in Table 2 and every configuration in Tables 4 and 10. The summary below uses six tasks for Table 2 and three visual tasks for Tables 4/10. We compute each run's task average before calculating sample SD.

ComparisonS1 Avg.S2 Avg.S3 Avg.Mean ± SD
Table 2, 4B Imagine-OPD79.1679.2879.8279.42 ± 0.35
Table 2, 4B VISTA80.8980.6580.6480.73 ± 0.14
Table 2, 8B Imagine-OPD81.2181.2081.4281.27 ± 0.13
Table 2, 8B VISTA82.1282.1282.2582.16 ± 0.07
Table 4, Uniform hybrid82.9982.7582.8582.87 ± 0.12
Table 4, HAPI83.9983.5883.9183.83 ± 0.22
Table 10, Self-only81.6180.7682.7381.70 ± 0.99
Table 10, Collective83.9983.5883.9183.83 ± 0.22

VISTA exceeds the strongest competing distillation method, Imagine-OPD, in each run at both scales, with mean gains of 1.31 and 0.89 points. HAPI outperforms Uniform in all three runs, as does collective experience compared with self-only hindsight. These results show that both gains are consistent across the evaluated runs.

2. Performance at common compute budgets against GRPO and OPSD. On Qwen3-VL-4B with eight H100 GPUs, final training costs are 150/165/180 GPU-hours for GRPO/OPSD/VISTA, with three-task averages of 80.57/81.61/83.83. Peak memory is approximately 80 GB per GPU for each method. At common cumulative budgets:

GPU-hour capGRPOOPSDVISTA
5079.1979.3679.70
10079.4980.8182.09
15080.5781.5383.75

The methods share initialization, training data, hardware, and the student tool interface. Each entry uses the last available checkpoint within its cap and reports a three-run mean. At 150 GPU-hours, VISTA retains gains of 3.18 and 2.22 points over GRPO and OPSD; its 100-GPU-hour result also exceeds their 150-GPU-hour results. Section 4.9 and Appendix H.2 report per-task scores and the checkpoint protocol.

Together, these results support gains over GRPO/OPSD at common compute budgets on the three visual benchmarks and consistent gains over Imagine-OPD across six tasks under the same training schedule.

AuthorsOpenReview ↗

Response to 7ZKu (2/2): Observation behavior and experience controls

3. Which observation behaviors improve? We evaluate GRPO, OPSD, and VISTA on a fixed set of 200 questions and label crop relevance and evidence sufficiency independently of the final model answer.

MethodRelevant crop rateDecisive-evidence coverageError given decisive evidence
GRPO57.02%69.00%17.39%
OPSD61.11%73.00%15.07%
VISTA66.06%77.00%12.99%

Relevant crop rate is the fraction of crops that help locate the target or determine a required attribute or relation. Decisive-evidence coverage is the fraction of questions for which the accumulated observations suffice to distinguish the answer. Conditional error is the error rate on those evidence-covered questions.

On this diagnostic set, VISTA acquires relevant evidence more consistently, with coverage 8 and 4 points higher than GRPO and OPSD. It also makes fewer errors after acquiring decisive evidence, with conditional error lower by 4.40 and 2.08 points. These measures separately characterize evidence acquisition and subsequent errors, with conditional error evaluated on each method's own evidence-covered subset. Section 4.9 and Appendix H.3 (Table 14) provide the definitions and results.

4. Do unsuccessful trajectories contribute useful observations? We compare successful-source-only and full-group experience in mixed-success groups under the same per-group artifact cap B=4.

Source poolV*HR-4KHR-8KAvg. ± SD
Successful sources only87.7883.0079.0883.29 ± 0.14
Full group88.4883.5079.5083.83 ± 0.22

Full-group experience improves all three tasks, with a 0.54-point average gain. In a separate 200-group selector diagnostic, both conditions select 765 artifacts in total; 115 full-group artifacts originate exclusively from unsuccessful trajectories (15.03%). These results support the value of retaining observations from unsuccessful trajectories under the shared budget cap. Appendix H.3 (Table 15) reports the comparison.

5. Contributions of outcome feedback and interaction relations. We remove these channels independently while keeping the artifact-selection procedure and budget unchanged.

Teacher-context channelsThree-task Avg. ± SD
Neither82.72 ± 0.05
Feedback only83.06 ± 0.05
Relations only83.38 ± 0.09
Both83.83 ± 0.22

Removing feedback, relations, or both reduces average accuracy by approximately 0.46, 0.77, or 1.11 points. The training variants in Table 16 share the artifact-selection procedure and budget. The original CVED-versus-Unstructured comparison in Table 3 provides the complementary control with identical selected observations and visual budget, isolating the value of organizing that evidence. Together, these controls support the contributions of both the collective visual observations and the context used to organize them.

We hope these responses and new results have addressed your concerns. Thank you again for your valuable suggestions and for helping us improve the manuscript. We look forward to your further feedback!

AuthorsOpenReview ↗

Follow-up on our response to Reviewer 7ZKu

Dear Reviewer 7ZKu,

We appreciate your constructive feedback on our submission. We have uploaded our responses and additional analyses addressing each of your points. As the discussion phase is drawing to a close, we would welcome your thoughts on whether our reply resolves your concerns, and we remain happy to provide any further clarification.

Thank you again for your time,

The authors

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

This paper studies how multiple interaction trajectories sampled by an active multimodal agent for the same input can provide richer training supervision. VISTA has two components: CVED aggregates visual observations and interaction links across trajectories to construct teacher contexts aligned with the current decision, and performs distillation on prefixes sampled by the student; HAPI applies GRPO to successful trajectories and distillation to failed ones. The trained student relies only on its own interaction history and visual tools. Experiments cover fine-grained visual question answering, general multimodal reasoning, and multiple Qwen backbones, with ablations on experience sources, context organization, and supervision allocation.

Critical strengths
  1. A clear problem with practical value in reusing experience across trajectories. The method uses visual observations from existing rollouts to provide additional supervision. The main configuration does not require giving the teacher reference answers or final answers from other trajectories, nor does it require generating new teacher target trajectories. The extra context used during training leaves the student's inference interface unchanged.

  2. Controlled experiments provide substantial support for overall effectiveness. Tables 2 and 7 show gains across model sizes and backbones. More directly, Table 10 holds group-level outcome feedback and other settings constant while extending single-trajectory hindsight to collective experience, raising mean accuracy from 81.70% to 83.83%. In Table 3, CVED also outperforms Unstructured with the same selected observations and visual budget. These results support the value of visual information from other trajectories and how it is organized.

Critical weaknesses
  1. HAPI's routing gains are not sufficiently separated from loss reweighting. Appendix E already provides a full factorial ablation of the two masks. According to Appendix B.3, Uniform hybrid normalizes the distillation term by the number of tokens in the entire group, NN, whereas HAPI uses the number of tokens in failed trajectories, N−N^-; the RL term still uses the full-group denominator. Thus, at the same λ\lambda, the distillation coefficients for failed tokens differ by a factor of N/N−N/N^-, which varies with the proportion of failed tokens. Table 11 supports the full configuration, but does not yet establish whether the gains come from allocating supervision by success or failure or from changes in effective loss weighting. I suggest a comparison with a shared distillation denominator and separate tuning of λ\lambda for Uniform hybrid.

  2. Gains supporting the core components lack statistical uncertainty estimates. The paper states that all results are averaged over three random seeds, but reports no standard deviations, individual runs, or confidence intervals. HAPI improves mean accuracy over Uniform hybrid by 0.96 percentage points (Table 11), while CVED improves over Oracle ROI by 0.58 percentage points (Table 3). These differences directly affect the assessment of the component designs. Results for each seed and their variability are needed to assess the stability of these gains.

Minor issues. Appendix C acknowledges that the same training schedule does not imply the same computational cost, but does not quantify the overhead of teacher forward passes and context construction. Please report GPU time and peak GPU memory. Also clarify how the final checkpoint is selected and whether hyperparameters are chosen using an independent validation set.

Questions for authors
  1. Does HAPI retain its advantage in Table 11 if both HAPI and Uniform hybrid normalize distillation by the total number of tokens in the group, or if λ\lambda is tuned separately for Uniform hybrid?
  2. Could you provide the results for all three random seeds and their variability for Tables 3 and 11 to show whether the gains from the key components are stable?
3 comments
AuthorsOpenReview ↗

Response to DJVY (1/2): Routing controls and three-run stability

We sincerely thank you for your thoughtful review and constructive suggestions. In response to your concerns about routing, stability, and training cost, we have added a common-denominator control, complete results across three runs, and comparisons at common compute budgets. We have incorporated these results and the accompanying clarifications into the revised manuscript and respond to each point below.

1. HAPI retains its advantage after matching the distillation denominator. As you noted, the original Uniform hybrid and HAPI use different distillation denominators. To isolate the effect of routing, we fix lambda = 0.10, the RL denominator, training data, optimization budget, and teacher updates, and divide the distillation term by the full-group response-token count N in all four routing configurations.

ConfigurationThree-task Avg. (mean ± sample SD)
Uniform hybrid, N82.87 ± 0.12
RL mask only, N82.55 ± 0.18
Distillation mask only, N83.22 ± 0.11
HAPI, shared N83.67 ± 0.17
HAPI, original N-minus83.83 ± 0.22

Under the common denominator, joint routing outperforms Uniform hybrid by 0.80 percentage points and also exceeds either mask alone. The original normalization adds a further 0.16 points. This supports a benefit from routing beyond the change in scaling introduced by the original denominator. The revised main-text HAPI analysis and Appendix H.1 (Table 12) report this control, including per-task variability.

2. Stability across three runs. Appendix H.5 reports every task-level score for S1-S3 and sample SD for Tables 3 and 11. We compute the mean and sample SD across the three run-level task averages.

Original comparisonS1 Avg.S2 Avg.S3 Avg.Mean ± SD
Uniform hybrid82.9982.7582.8582.87 ± 0.12
HAPI83.9983.5883.9183.83 ± 0.22
Unstructured81.8782.3983.0382.43 ± 0.58
Oracle ROI82.6083.1184.0483.25 ± 0.73
CVED83.9983.5883.9183.83 ± 0.22

HAPI exceeds Uniform hybrid by 1.00, 0.82, and 1.06 points. CVED exceeds Unstructured in every run under the same selected observations and visual budget. CVED also achieves a 0.58-point average gain over Oracle ROI, outperforming it in two of three runs. The consistent gains over Unstructured support the value of organizing the same observations, while the comparison with Oracle ROI shows greater variation across runs. These are the original-normalization results; Table 12 separately reports the shared-N control.

AuthorsOpenReview ↗

Response to DJVY (2/2): Training cost and model selection

3. Training cost and performance at common GPU-hour budgets. To clarify the computational cost, we report results for Qwen3-VL-4B on eight NVIDIA H100 GPUs, using the same initialization, training data, and student tool interface across methods. All averages below cover V*Bench, HR-Bench-4K, and HR-Bench-8K.

MethodTraining GPU-hoursFinal-checkpoint Avg.
GRPO15080.57
OPSD16581.61
VISTA18083.83

Peak memory is approximately 80 GB per GPU across the three methods. At the final scheduled checkpoint, VISTA improves over GRPO by 3.26 points with 20% more reported training compute, and over OPSD by 2.22 points with 9.09% more compute. The advantage also persists at common cumulative GPU-hour caps:

Budget cap (GPU-hours)GRPOOPSDVISTA
5079.1979.3679.70
10079.4980.8182.09
15080.5781.5383.75

Each entry uses the last available checkpoint within the stated cap and reports mean accuracy across three runs. At 150 GPU-hours, VISTA exceeds GRPO and OPSD by 3.18 and 2.22 points. At 100 GPU-hours, it already exceeds both baselines at 150 GPU-hours. Section 4.9 and Appendix H.2 (Table 13 and Figure 6) report these results and the per-task scores.

4. Checkpoint and hyperparameter selection. The main comparisons use the final scheduled checkpoint for every method and run. Hyperparameters are selected using an independent validation set. For the fixed-budget comparison, we use the last checkpoint within each GPU-hour cap; checkpoint selection is independent of test accuracy.

The parameter ablations also show consistent gains across the tested settings: every tested nonzero distillation weight improves all three visual tasks over the zero-distillation setting, and larger rollout groups improve or preserve accuracy across the tested range. The experimental setup and Appendix H.2 clarify the selection protocol; the main paper reports sensitivity to each parameter.

We hope these responses and new results have addressed your concerns. Thank you again for your valuable suggestions and for helping us improve the manuscript. We look forward to your further feedback!

AuthorsOpenReview ↗

Follow-up on our response to Reviewer DJVY

Dear Reviewer DJVY,

We appreciate your constructive feedback on our submission. We have uploaded our responses and additional analyses addressing each of your points. As the discussion phase is drawing to a close, we would welcome your thoughts on whether our reply resolves your concerns, and we remain happy to provide any further clarification.

Thank you again for your time,

The authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response for VISTA

We sincerely thank all reviewers for their time, thoughtful feedback, and constructive suggestions. We appreciate their recognition of the collective-experience perspective and the controlled comparisons in our work. Their comments have helped us identify where the method requires clearer explanation and where additional experimental evidence would be most useful.

In response, we have conducted additional experiments, expanded the reporting of results across three runs, and clarified the construction and use of collective visual experience. These revisions address the main questions about routing, training cost, observation behavior, and the contributions of different experience sources. We summarize the key changes below and provide detailed responses in the individual reviewer replies.

Reviewer ConcernRelevant Reviewer(s)Our Actions and Revisions
1. Routing benefits and normalization controlsDJVY, tGQmWe added a common-denominator experiment comparing Uniform hybrid, each individual routing mask, and joint HAPI routing. With the distillation denominator matched, HAPI retains a 0.80-point advantage over Uniform hybrid and outperforms both single-mask controls. These results help distinguish the routing benefit from the effect of normalization (main-text HAPI analysis; Appendix H.1, Table 12).
2. Stability across runs and teacher updatesDJVY, 7ZKu, tGQmWe expanded the reporting of task-level results for all three runs and sample standard deviations for the main training comparisons and key ablations. The results show consistent gains for HAPI over Uniform and collective experience over self-only hindsight across the evaluated runs. We also clarify the variability of the Oracle ROI comparison and the task-dependent effects of EMA versus frozen teacher weights (Appendix H.5).
3. Training cost and performance under common compute budgetsDJVY, 7ZKu, tGQmWe added training GPU-hours, peak GPU memory, and comparisons at common cumulative compute budgets for GRPO, OPSD, and VISTA on Qwen3-VL-4B across three visual benchmarks. At a common 150-GPU-hour budget, VISTA achieves 83.75% average accuracy, exceeding GRPO and OPSD by 3.18 and 2.22 points, respectively (Section 4.9; Appendix H.2, Table 13 and Figure 6).
4. Evidence acquisition and premature answering7ZKu, tGQmWe added a 200-question observation diagnostic measuring crop relevance, decisive-evidence coverage, and errors after acquiring decisive evidence. We also compared CVED and Oracle ROI on a 100-question subset requiring additional evidence, where CVED reduces both premature answering and premature incorrect answering by eight percentage points (Section 4.9; Appendix H.3, Tables 14 and 17).
5. Contributions of unsuccessful trajectories and contextual information7ZKu, tGQmWe compared successful-source-only and full-group experience in mixed-success groups under the same artifact budget cap. Full-group experience improves average accuracy from 83.29% to 83.83%. A separate selector diagnostic identifies useful selected artifacts originating exclusively from unsuccessful trajectories. We also added feedback and relation ablations using the same selection procedure and budget, complementing the existing CVED-versus-Unstructured control with identical observations (Appendix H.3, Tables 15 and 16).
6. Concrete construction of collective experience and conditions for useful guidancetGQmWe clarified the construction process and added a worked Fullerton example showing the student’s missing evidence, the complementary peer observations, and their presentation to the teacher. We also explain how unsuccessful trajectories can contribute useful observations and why the value of teacher guidance depends on whether the shared experience supplies missing evidence (Section 3.2; Appendix H.3–H.4).
7. Checkpoint and hyperparameter selectionDJVYWe clarified that the main comparisons use the final scheduled checkpoint, hyperparameters are selected using an independent validation set, and common-budget evaluations use the last available checkpoint within each cap. The parameter ablations further characterize sensitivity across the tested settings (experimental setup; Appendix H.2).

The main findings are summarized in the revised manuscript, with detailed results and analyses provided in Appendix H. New results and analyses are highlighted in blue to facilitate review.

We hope these responses, additional experiments, and clarifications have addressed the reviewers’ concerns. We are grateful for the valuable suggestions that have helped us improve the manuscript, and we look forward to further discussion!

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.