acceptodds
Under review as a conference paper at ICLR 2027

Live-OPSD: On-Policy Self-Distillation for Proactive Streaming Large Video Models

10 views

Abstract

Streaming multimodal assistants must make response decisions while events are still unfolding. They need to recognize when the available evidence warrants a response, produce an appropriate continuation, and remain silent otherwise. Supervised fine-tuning learns from recorded assistant turns, but does not directly train on the response prefixes produced by the evolving policy itself. We introduce Live-OPSD, an on-policy self-distillation method that uses privileged hints available only during training to supervise these prefixes. The policy first generates a continuation from the streaming context alone. The same model then re-evaluates that continuation with a hint describing the relevant evidence, and the resulting distributions guide the unhinted policy. On Qwen3-Omni-30B-A3B, Live-OPSD attains an average score of 59.64 on LiveProBench, 5.26 points above the strongest continued-SFT baseline, and stays competitive on OVO-Bench (65.84). These results indicate that guiding the policy on its own outputs with privileged hints improves both when and how a streaming assistant responds.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
on-policy distillation, streaming video understanding, multimodal models, real-time interaction

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

Live-OPSD trains proactive streaming models on their own response continuations at recorded contexts. The same model, conditioned on privileged hints, supplies distillation targets, while Decision-Aware Selection prioritizes training points. Evaluation on two Qwen-Omni backbones shows improved average LiveProBench scores over continued SFT.

Critical strengths
  1. Privileged-context distillation provides token-level feedback on current-model continuations without an additional inference-time teacher or gate.
  2. Matched-backbone comparisons and divergence/selection ablations support the approach.
Critical weaknesses
  1. The source of the gains is insufficiently isolated. Live-OPSD receives externally annotated hints unavailable to the SFT baselines. Without hint and rollout controls, the contribution of on-policy distillation cannot be separated from additional privileged supervision. Hint generation also uses full-clip access, and the annotation-model identity is not reported.

  2. The on-policy scope and baseline coverage are limited.

    • Only the current continuation is generated on policy; earlier assistant turns remain recorded. This leaves deployment-history errors outside the training distribution.
    • Comparisons omit timing-oriented RL, including StreamPro's GRPO stage. CB-SFT includes only its supervised objective, so the results do not establish an advantage over policy-optimization alternatives.
Questions for authors
  1. Can you compare no-hint distillation, hinted distillation on reference prefixes, and full Live-OPSD to isolate the effects of hints and on-policy generation? How are hints checked for future-information leakage?
  2. How does performance change when earlier assistant turns are model-generated, particularly after premature or missed responses?
  3. Can you compare against a timing-reward RL baseline using the same backbone, data, initialization, and compute budget?
3 comments
AuthorsOpenReview ↗

Responce to Reviewer qmWh (Part 1/3)

We thank the reviewer for the constructive suggestions. We respond to each concern below.

1. Privileged hints and prefix source (Weakness 1, Question 1)

In our current rollout setup, preceding assistant turns come from preconstructed interaction histories, while the current policy generates the continuation at the target decision point. To isolate the effect of prefix source, we compare hinted distillation on preconstructed histories with the same hinted distillation on self-generated preceding histories. Both settings use the same point-specific hints and dynamically updated actor as the teacher, with the remaining training settings matched as closely as possible.

We do not include a no-hint distillation control. In our current setup, removing the hint while using the current policy as both teacher and student would produce no meaningful teacher–student learning signal. We also do not have a stronger external proactive streaming model with a compatible response protocol to use as a no-hint teacher (See Point 4). We therefore do not attribute the full gain over continued SFT to prefix source alone. The prefix comparison below isolates the effect of preceding-history source, while the hint-component ablation examines which information in the hint is useful.

The use of preconstructed preceding turns is intentional: it lets us study the target-turn distillation signal without earlier actor errors changing the context for later turns. This design was inspired by SOD [1], which reports that errors in multi-step agent trajectories can cascade, increase student–teacher divergence, and make later token-level supervision less reliable. This motivates our controlled-history design.

Table R1. Prefix-source comparison on LiveProBench (Qwen3-Omni-30B-A3B, %).

VariantES Resp.ES Sil.ES Geo.ORWRAASSDCAvg.
Live-OPSD (preconstructed preceding turns)40.2899.5963.3461.9370.0068.4959.7834.3359.64
- Self-generated prefix39.8197.9762.4562.3972.5068.4944.5732.8457.20

The “- Self-generated prefix” control keeps the same point-specific hints and dynamically updated actor as the teacher, but replaces the preconstructed preceding turns with actor-generated turns. The target continuation is generated by the current policy in both settings. The LiveProBench average is 2.44 points lower with self-generated preceding turns.

Our hints describe the currently available evidence and indicate whether the assistant should respond or remain silent. To examine the contribution of these hint components, we compare the full hint with variants that remove the response/silence recommendation or the stream description.

Table R2. Hint-component ablations on LiveProBench (Qwen3-Omni-30B-A3B, %).

VariantES Resp.ES Sil.ES Geo.ORWRAASSDCAvg.
Live-OPSD (full hint)40.2899.5963.3461.9370.0068.4959.7834.3359.64
Without response/silence recommendation39.8197.1562.1960.5563.7560.9253.2620.9053.60
Without stream description31.9497.5655.8348.1752.5050.4260.8710.4546.37

Details of data generation. The external annotator is Qwen3.5-Omni. When generating a hint for decision point t, its video input contains only the stream available up to t; it is not given future frames or the complete clip. Full-clip access is used only in a separate post-generation review of hint quality and temporal appropriateness. The student remains restricted to the current streaming context, and the deployed model receives no hints. The continued-SFT baselines use the same 3,712 training examples and their reference responses.

The 5,000 candidate examples for Live-OPSD were randomly sampled from the SFT data. We retained 3,712 examples for OPSD training, reserved 50 for testing, and discarded 1,238 after example-level quality filtering. If any turn in an example contained erroneous or unsuitable data or a hint, we discarded the entire example.

Live-OPSD data filtering and train/test split.

Filtering outcomeNumber of examplesPercentage
Candidate SFT examples5,000100.00%
Retained for OPSD training3,71274.24%
Reserved for the test set501.00%
Discarded after review1,23824.76%

[1] Q. Zhong, M. Zheng, M. Song, X. Lin, J. Sun, H. Jiang, X. Wang, and J. Fang. “SOD: Step-wise On-policy Distillation for Small Language Model Agents.” arXiv:2605.07725, 2026.

AuthorsOpenReview ↗

Responce to Reviewer qmWh (Part 2/3)

2. Earlier assistant turns generated by the model (Weakness 2, Question 2)

The effect of self-generated preceding turns is already evaluated by the “- Self-generated prefix” control in Table R1. Compared with Live-OPSD using preconstructed histories, actor-generated histories reduce the LiveProBench average from 59.64 to 57.20. The Duplicate Counting (DC) score also decreases from 34.33 to 32.84, indicating a drop in the model’s ability to handle tasks requiring repeated proactive responses. This control includes histories in which earlier turns may contain premature or missed responses, and shows how using such histories affects subsequent performance.

3. Comparison with timing-reward RL (Weakness 2, Question 3)

We compare Live-OPSD with GRPO using the same Qwen3-Omni backbone, initialization, and training data. For GRPO, each annotated response point defines a timing window extending from that point through the following two seconds; responses within this window receive an IoU-based timing reward.

Live-OPSD uses one rollout per example naturally, while GRPO uses eight as it need to estimate group-relative advantages. Both runs use 121 update steps and a learning rate of 5e-6. We report the optimization settings so that the scope of the comparison is clear.

Table R3. Comparison on all LiveProBench metrics (Qwen3-Omni-30B-A3B, %).

MethodES Resp.ES Sil.ES Geo.ORWRAASSDCAvg.
Live-OPSD40.2899.5963.3461.9370.0068.4959.7834.3359.64
GRPO38.4397.9761.3657.8075.0070.5942.3910.4552.93

Table R4. Training settings.

MethodRollouts per exampleUpdate stepsLearning rate
Live-OPSD11215e-6
GRPO81215e-6

This comparison illustrates the different training signals used by the two methods. GRPO relies on a trajectory-level timing reward. Designing this reward requires distinguishing premature responses from responses made within the annotated window and delayed responses outside it. Since the reward is assigned at the trajectory level, it also provides limited credit assignment when a multi-turn trajectory contains several response decisions with different outcomes. In addition, GRPO’s group-based update requires multiple rollouts of a multi-turn video stream, while Live-OPSD only needs one actor rollout per example.

AuthorsOpenReview ↗

Responce to Reviewer qmWh (Part 3/3)

4. How GRPO/RLVR and Live-OPSD differ, and why we consider OPSD may especially suited to proactive streaming video models (Question 3)

The considerations below motivated our choice of Live-OPSD but were not discussed in sufficient detail in the paper due to space constraints. We are grateful that the reviewer recognized this underlying motivation and prompted us to articulate it more clearly.

GRPO/RLVR:

  • Requires a timing-sensitive reward that distinguishes timely, evidence-supported responses from premature responses that may be hallucinated and delayed responses that miss the appropriate interaction window. Designing such a reward is challenging.
  • Standard policy-gradient updates often distribute the same advantage across all tokens in a trajectory. Yet a trajectory may contain multiple proactive responses, some appropriate and others not, so a single outcome reward provides weak credit assignment to the individual decisions.
  • Tasks requiring repeated proactive responses, such as counting push-ups, make this problem harder: the reward must assess multiple response events and their timing. More elaborate rewards may increase the risk of reward hacking, while sparse rewards and coarse credit assignment can still limit learning.
  • RLVR requires multiple sampled rollouts to estimate group-relative advantages. In multimodal streaming video, each rollout must process a multi-turn video stream, making this a particularly time-consuming part of training.

Live-OPSD:

  • Uses hinted token-level teacher targets instead of a manually designed outcome reward, provided the teacher supplies a useful target that improves on the unhinted student’s output.
  • Provides token-level distillation targets, allowing training to focus on tokens relevant to proactive response decisions and content, rather than assigning one trajectory-level advantage to every token.
  • This fine-grained supervision naturally supports examples requiring multiple proactive responses within one video, since it can guide the relevant response tokens at each decision point.
  • Does not require a group of sampled rollouts to estimate relative rewards. In our training setup, it uses one on-policy actor rollout per example, reducing the heavy rollout cost(especially for multimodal examples) compared with multi-sample GRPO.
  • Real-time proactive streaming video models may introduce model-specific control tokens (speak/slience token), making it difficult to find a stronger external teacher with a compatible vocabulary and output structure. Live-OPSD uses the same model as a teacher, avoiding the need for a separate and stronger teacher model.

We thank the reviewer again and we hope these analyses make the source of the gains and the scope of on-policy training explicit. We are looking forward to your response.

Reviewer_a52gOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the response-timing problem in proactive streaming multimodal assistants, where a model must not only generate correct responses but also determine when to respond and when to remain silent. To address this problem, the authors propose Live-OPSD, which first lets the current policy generate its own response prefixes and then uses the same model, augmented with training-time privileged hints, as a teacher to distill knowledge on these prefixes. The method further incorporates Decision-Aware Selection (DAS) to prioritize time points where response decisions are incorrect. The main technical contribution is the integration of on-policy prefix learning, privileged self-distillation, and decision-aware sample selection into streaming post-training. Experiments are conducted on Qwen2.5-Omni-7B and Qwen3-Omni-30B-A3B and evaluated on LiveProBench, OmniMMI, OVO-Bench, and StreamingBench. The results show that the 30B model achieves a score of 59.64 on LiveProBench, outperforming the strongest continued-SFT baseline by 5.26 points, indicating that the proposed method can effectively improve response timing while maintaining strong overall video understanding capability.

Critical strengths

Strength 1: The paper identifies an important train–deployment distribution mismatch and proposes a targeted solution

The paper addresses the limitation that conventional SFT cannot directly train on prefixes generated by the current policy in proactive streaming assistants. It proposes Live-OPSD, which combines on-policy self-distillation with privileged hints. This design directly targets the distribution mismatch encountered at deployment and does not introduce additional inference-time modules, giving the method strong motivation and practical value.

Strength 2: The main experimental results are competitive

On Qwen3-Omni-30B-A3B, Live-OPSD improves the average LiveProBench score from 54.38 for the strongest continued-SFT baseline to 59.64. It also performs well across multiple streaming video benchmarks, suggesting that the method is effective and exhibits a certain degree of generalization.

Critical weaknesses

Weakness 1: The core contribution of on-policy prefixes is not independently validated through a key controlled experiment

The paper presents on-policy student-generated prefixes as a core contribution, but does not provide a strictly matched comparison against reference/off-policy prefix distillation. Therefore, the current results cannot disentangle whether the performance gains primarily come from the privileged teacher or from the use of on-policy prefixes themselves. This weakens the causal attribution of the paper’s central mechanism.

Weakness 2: Insufficient reproducibility and result stability

The paper has notable reproducibility limitations. Appendix A explicitly states that complete corpus counts are unavailable in the retained artifacts. More importantly, privileged hints are a core supervision signal, yet the authors state that the identity of the external annotation model used to generate these hints, as well as the hint retention counts, can no longer be recovered from the current artifacts. In addition, the paper does not report means and standard deviations over multiple random seeds for the main results. For relatively small differences, such as reverse KL (59.64) vs. JSD (59.21), and Live-OPSD (48.97) vs. Vanilla-SFT (48.33) on the 7B model, it is therefore difficult to assess the statistical stability of the reported gains.

Questions for authors

Q1. Can the authors provide a strictly controlled experiment to isolate the contribution of on-policy prefixes?

Specifically, I would like to see a comparison between privileged-context distillation on reference/recorded prefixes and the same privileged-context distillation on current-policy-generated prefixes. The two settings should keep the hint, teacher, divergence objective, training data, number of training steps, optimizer, selected-point budget, and computational budget as consistent as possible. This result would substantially affect my assessment of the paper’s central claim.

Q2. Can the authors provide or recover the key information required for reproducibility and report the variance of the main results?

In particular, I would like clarification on the identity and version of the privileged-hint annotation model, hint filtering/retention statistics, complete training-data sources and scale, the actual DAS quota/scoring parameters, the fixed-suffix rule used in the w/o-DAS setting, the values of \beta and \lambda_{\text{text}} for CB-SFT, and whether multi-seed results can be provided for the main 30B experiments.

Q3. How do the authors ensure that privileged hints do not leak future information?

Appendix B states that the external annotator can access the full clip, whereas the student at time t can only access the currently available streaming context. I would therefore like the authors to clarify whether the hints are strictly restricted to evidence observable before the current decision time, whether any manual or automated audit was conducted to detect future-only information, and, if so, what the corresponding error or pass rate was. If future information is present in the hints, part of the teacher’s decision advantage may come from information that the student can never access at deployment time, which would affect the interpretation of the method’s gains.

4 comments
AuthorsOpenReview ↗

Responce to Reviewer a52g (Part 1/2)

We thank the reviewer for constructive suggestions, and we respond to each concern below.

1. Controlled comparison of recorded and self-generated prefixes(Weakness 1, Q1)

To directly test the effect of prefix source, we compared hinted distillation on (1) preconstructed histories and (2) actor-generated preceding histories. Both settings use the same point-specific hints and dynamically updated actor as the teacher, with the remaining training settings matched as closely as possible.

We do not include a no-hint distillation control. Without a stronger external teacher with a compatible response protocol, removing the hint would leave the current policy as both teacher and student, providing no meaningful teacher–student learning signal.

LiveProBench results on Qwen3-Omni-30B-A3B (%).

VariantES Resp.ES Sil.ES Geo.ORWRAASSDCAvg.
Live-OPSD40.2899.5963.3461.9370.0068.4959.7834.3359.64
- Self-generated prefix39.8197.9762.4562.3972.5068.4944.5732.8457.20

The “- Self-generated prefix” control keeps the same point-specific hints and dynamically updated actor as the teacher, but replaces the preconstructed preceding turns with actor-generated turns. The target continuation is generated by the current policy in both settings. The actor-generated histories can include premature or missed responses, testing how earlier decisions affect the context at later turns.

We use hints that describe the currently available evidence and explicitly indicate whether the assistant should respond or remain silent. This gives the teacher guidance on both the stream evidence and the appropriate action. To examine the contribution of each part, we compared the full hint with variants:

LiveProBench hint-component ablations on Qwen3-Omni-30B-A3B (%).

VariantES Resp.ES Sil.ES Geo.ORWRAASSDCAvg.
Live-OPSD(full hint)40.2899.5963.3461.9370.0068.4959.7834.3359.64
- w/o response/silence recommendation39.8197.1562.1960.5563.7560.9253.2620.9053.60
- w/o stream discription31.9497.5655.8348.1752.5050.4260.8710.4546.37

[1] Qiyong Zhong et al., “SOD: Step-wise On-policy Distillation for Small Language Model Agents,” arXiv:2605.07725, 2026. https://arxiv.org/abs/2605.07725.

AuthorsOpenReview ↗

Responce to Reviewer a52g (Part 2/2)

2. Reproducibility and stability (Weakness 2, Q2)

Implementation details. We generate privileged hints offline using Qwen3.5-Omni. In the w/o DAS ablation, we select a fixed suffix of n=4n=4 decision rounds ending at the last ground-truth proactive-response turn. This selection does not depend on the current policy, and earlier history remains in the input context. CB-SFT follows original StreamPro, with β=0.9999\beta=0.9999 for class reweighting and λtext=2\lambda_{\text{text}}=2 for response-text weighting.

Data source. The streaming-SFT stage uses approximately 196K examples generated from the Shot2Story and valor-32k datasets. For Live-OPSD, we randomly sampled 5,000 examples from the SFT data to form the candidate pool. After example-level quality filtering and reserving 50 examples for the test set, 3,712 examples remained for Live-OPSD training and 1,238 were discarded.

We report the Qwen3-Omni results across three seeds. The first table gives the mean and standard deviation for each LiveProBench metric, and the second gives the 95% confidence interval for the overall average score.

LiveProBench results across three seeds (mean ± SD, %).

MethodES Resp.ES Sil.ES Geo.ORWRAASSDCAvg.
Live-SFT-Zero-30B41.67 ± 1.2284.55 ± 0.4159.35 ± 0.8652.29 ± 1.2143.75 ± 2.5073.11 ± 0.7342.39 ± 2.1737.31 ± 2.9951.37 ± 0.42
+ Vanilla-SFT41.67 ± 0.9399.32 ± 0.4764.33 ± 0.5963.30 ± 3.0167.50 ± 2.1757.56 ± 1.5170.65 ± 1.882.99 ± 1.4954.39 ± 0.83
+ CB-SFT68.06 ± 2.1286.18 ± 0.4176.58 ± 1.2672.94 ± 1.6545.00 ± 3.3168.07 ± 1.4631.52 ± 2.8823.88 ± 2.9953.00 ± 0.30
+ SSB-SFT72.69 ± 2.8278.05 ± 0.7075.31 ± 1.3464.68 ± 1.2141.25 ± 2.1765.55 ± 2.2223.91 ± 2.1732.84 ± 1.4950.59 ± 0.52
+ Live-OPSD40.28 ± 2.4598.64 ± 0.2363.01 ± 1.9361.93 ± 2.0070.00 ± 3.3168.49 ± 1.2659.78 ± 3.2634.33 ± 1.4959.59 ± 1.05

Overall average scores and 95% confidence intervals (%).

MethodAvg. Mean ± SD95% CI
Live-SFT-Zero-30B51.37 ± 0.42[50.32, 52.42]
+ Vanilla-SFT54.39 ± 0.83[52.32, 56.46]
+ CB-SFT53.00 ± 0.30[52.24, 53.75]
+ SSB-SFT50.59 ± 0.52[49.29, 51.89]
+ Live-OPSD59.59 ± 1.05[56.98, 62.20]

All benchmark evaluations and checkpoints use greedy decoding, following the protocols convention for mllm benchmarks like LiveProBench. The variation across seeds reflects differences between different runs.

3. Clarification and audit of privileged hints (Q3)

We found that the current Appendix B does not accurately describe our hint-generation pipeline. Future information is excluded by construction. When generating a hint for decision point (t), the annotator receives only the video stream available up to (t), future frames and the complete clip are not provided to it. Full-clip access is used only in a separate post-generation review to check hint quality and temporal appropriateness. Thus, the hint generator cannot use future video evidence when producing a hint.

Our data filtering is performed at the example level: if any turn in an example contains erroneous or unsuitable data or a hint, we discard the entire example. The filtering counts below report overall data curation, not future-information leakage.

Live-OPSD data filtering and train/test split.

Filtering outcomeNumber of examplesPercentage
Candidate SFT examples5,000100.00%
Retained for OPSD training371274.24%
Reserved for the test set501.00%
Discarded123824.76%

4. Motivation of Live-OPSD

Finally, let us clarify why we chose Live-OPSD for proactive streaming video models. As these models must learn when to respond and when to remain silent, not just what to say. Reward-based RL is a natural option, but designing a reward that jointly captures timing, response content, and silence is challenging. Assigning credit to individual response decisions in multi-turn streams is also difficult.

OPD provides token-level feedback but requires a suitable teacher. We lack a stronger proactive streaming model with a compatible action-token vocabulary and response interface, and differing action tokens complicate external distillation. Live-OPSD therefore uses the same backbone, conditioned on privileged hints, to provide teacher targets for its own continuations.

We appreciate the reviewer’s careful reading and thank the reviewer again for your effort. We are looking forward to your responce.

Although the author replied to some of my questions, I still had some doubts about this batch of work, so I will maintain my score.

Thank you for your feedback and for taking the time to review our work. Your insights are extremely valuable to us.

We understand that you still have some doubts about this batch of work. To help us address these issues, could you kindly specify which particular concerns remain unresolved?

If there are specific experiments, clarifications, or comparisons you feel are still missing, please let us know. We are more than happy to provide further details and improve our revised draft.

Best,

Authors

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

This paper proposes Live-OPSD, an on-policy self-distillation method for proactive streaming video assistants. It uses privileged hints to guide the model's own continuations and Decision-Aware Selection (DAS) to prioritize training samples. Experiments on two Qwen-Omni backbones show improvements over continued-SFT baselines.

Critical strengths
  1. The paper addresses the intertwined challenges of premature responses, missed events, and response quality. Applying OPSD offers an intuitive and well-motivated alternative to constructing more elaborate SFT datasets.

  2. The paper evaluates two backbones and three SFT baselines with shared initialization and training data, covering multiple benchmarks including LiveProBench, OmniMMI, OVO-Bench, and StreamingBench.

  3. The paper is well organized and easy to follow.

Critical weaknesses
  1. Insufficient analysis of privileged information and teacher effectiveness.
  • The effectiveness of OPSD depends on both the choice of privileged information and the model's ability to effectively leverage it. The authors should evaluate the performance of the baseline teacher model on the target tasks before and after introducing privileged information to better understand its actual contribution.
  • Furthermore, given a fixed form of privileged information, it would be valuable to investigate whether models of different scales exhibit consistent patterns in their ability to understand and benefit from such information, and how these differences affect the effectiveness of OPSD.
  1. Privileged information selection analysis. Following the previous concern, the choice of privileged information is a key component of OPSD. The authors should provide a clearer rationale for their choice of privileged hints, discuss possible alternatives, and investigate how different types of privileged information affect the results.

  2. Limited attribution of broader benchmark gains. The OVO-Bench comparison includes both streaming SFT and Live-OPSD, making it difficult to isolate the contribution of Live-OPSD.

  3. Insufficient isolation of core distillation components. The main comparisons simultaneously change supervision, training prefixes, and sample selection. Beyond the overall DAS ablation, controlled experiments are needed to isolate the effects of privileged hints, student-generated versus reference prefixes, and updated versus fixed teacher targets.

Questions for authors

See the weaknesses.

3 comments
AuthorsOpenReview ↗

Responce to Reviewer SwdG (Part 1/3)

We thank the reviewer for the positive assessment of the paper’s clarity and for the constructive suggestions. We respond to each concern below.

1. Analysis of privileged information and teacher effectiveness(weakness 1).

Our choice of OPSD is not incidental. It is a carefully considered design choice made after careful consideration of the distinctive properties of proactive streaming video models and we believe this setting is particularly well suited to OPSD. Unlike conventional offline VQA, a proactive streaming model must repeatedly decide, as new evidence arrives, whether to respond or remain silent. The capability we aim to develop is not simply better answer generation, but a new response policy for an evolving stream. To the best of our knowledge, no stronger existing model can serve as a teacher for this capability.

This setting also creates practical challenges for conventional OPD. Proactive streaming models often may modify their architecture and introduce task-specific action tokens to represent streaming decisions. Our model, for example, uses a dedicated silence token, while other models may use different action tokens, such as JoyAI-VL’s response token. Consequently, models may differ in both architecture and vocabulary, making direct token-level distillation from an external model difficult or inapplicable. OPSD is therefore a natural fit: the model learns from its own evolving policy, while privileged information conditions its own teacher during training.

We have now measured this directly by evaluating the same backbone and model checkpoint(Live-SFT-zero-30B/7B) with and without hint conditioning (greedy decoding over our training set, 3712 streaming dialogues, 108486 labeled decision points):

Greedy student–teacher comparison with and without hints (%).

Metric (%)30B Student (w/o hint)30B Teacher (w/ hint)7B Student (w/o hint)7B Teacher (w/ hint)
Fully-correct trajectories14.2496.698.7095.13
Unnecessary responses (per silent point)11.830.0320.160.05
Missed proactive responses (per proactive point)24.971.0720.641.58
On-time response (single-response dialogues)15.7997.499.3597.06
Mean trajectory reward87.9199.8582.0299.78

Note that the mean trajectory reward is high for all runs because silence, the majority behavior, accounts for most decision points; the informative rows are the two error rates above it. Hint conditioning improves the mean trajectory reward by +11.94 points on 30B and +17.76 points on 7B. A comparison across external teacher scales is not applicable because we do not have a stronger, suitable proactive streaming teacher for this setting.

2. Rationale for privileged-hint selection.(Weakness 2)

We use hints that describe the currently available evidence and explicitly indicate whether the assistant should respond or remain silent. This gives the teacher guidance on both the stream evidence and the appropriate action. To examine the contribution of each part, we compared the full hint with variants:

LiveProBench hint-component ablations on Qwen3-Omni-30B-A3B (%).

VariantES Resp.ES Sil.ES Geo.ORWRAASSDCAvg.
Live-OPSD(full hint)40.2899.5963.3461.9370.0068.4959.7834.3359.64
w/o response/silence recommendation39.8197.1562.1960.5563.7560.9253.2620.9053.60
w/o stream discription31.9497.5655.8348.1752.5050.4260.8710.4546.37

This comparison tests whether explicitly indicating the desired response decision helps the hinted teacher provide more useful supervision.

AuthorsOpenReview ↗

Responce to Reviewer SwdG (Part 2/3)

3. Rationale for the privileged hints and their temporal scope.

Our hints summarize evidence relevant to the response decision at time t, providing a concise training-time signal while keeping the deployed policy hint-free. We would like to clarify a possible ambiguity in the data-pipeline description: during hint generation, the generator only has access to the video stream available up to the current decision time t, and it does not access future frames or the complete clip. Full-clip access is used only in the subsequent review of generated hints, not to generate them. We acknowledge that the paper may not clearly distinguish these two stages. We will revise the description to state explicitly what information is available during hint generation and during hint review, and report the review procedure and statistics: [details, TBD]. We will also clarify why this hint representation is used and discuss alternative forms of privileged information. The review statistics are as follows:

Live-OPSD data filtering and train/test split.

Filtering outcomeNumber of examplesPercentage
Candidate SFT examples5000100.00%
Retained for OPSD training371274.24%
Reserved for the test set501.00%
Discarded123824.76%

4. Attribution of results on broader benchmarks(Weakness 3).

As OVO-Bench and StreamingBench only evaluate streaming understanding, but not directly evaluate proactive response behavior, which is the capability targeted by Live-OPSD. So these benchmarks mainly measure passive, question-driven video understanding in an offline QA format for our model. They do not test whether the model should proactively respond or remain silent as a stream unfolds. Benchmarks for proactive streaming response are currently scarce. We therefore include OVO-Bench and StreamingBench as complementary evaluations to examine whether optimizing proactive response with Live-OPSD comes at the cost of passive video-question-answering ability in streaming settings. Their full results are consequently reported in the appendix and discussed only briefly in the main text. These results are not intended to establish Live-OPSD’s proactive-response gains. We evaluate proactive response on LiveProBench, which covers six sub-tasks, and the PA subset of OmniMMI.

AuthorsOpenReview ↗

Responce to Reviewer SwdG (Part 3/3)

5. Disentangling the core distillation components(Weakness 4).

We will isolate the effects of prefix source and teacher-target updates. To clarify, our current method uses prefixes from preconstructed streaming dialogues: for each dialogue, responses for the turns are specified in advance, and each training context is formed by progressively splitting the dialogue. The teacher distribution, in contrast, is produced by our own actor and dynamically updated during training. Using the current Live-OPSD setting as the reference, we add two controlled comparisons on LiveProBench (Qwen3-Omni):

LiveProBench teacher and prefix ablations on Qwen3-Omni-30B-A3B (%).

VariantES Resp.ES Sil.ES Geo.ORWRAASSDCAvg.
Live-OPSD (original)40.2899.5963.3461.9370.0068.4959.7834.3359.64
- Fixed teacher42.1398.3764.3860.5567.5067.2370.6525.3759.28
- Self-generated prefix39.8197.9762.4562.3972.5068.4944.5732.8457.20

We did not compare vanilla OPD because we could not find a better teacher model with same vocabulary, and that is why we use OPSD. The fixed-teacher control achieves slightly lower performance but requires additional memory to maintain a separate teacher model; using the actor policy directly as the teacher is therefore a simpler and more memory-efficient choice. The self-generated-prefix control will test the effect of replacing preconstructed prefixes with prefixes generated by the actor, while keeping the teacher-target update rule and other training settings fixed. We will report the results of this comparison in the revision.

6. Why we think Live-OPSD naturally suits proactive streaming model(Weakness 1).

Proactive streaming requires optimizing when to speak, not just generating answers. We first considered RL, but jointly rewarding timing, content, and silence is difficult, especially when multiple responses require individual credit assignment.

OPD offers token-level feedback, yet we lack a stronger teacher suited to this model, and our added action tokens complicate external token-level distillation. OPSD therefore uses the same backbone with privileged hints, preserving vocabulary compatibility while improving the teacher's decision signal. Point 1 directly measures this effect.

We generate and review hints offline with an external annotator because online judging is too slow for multi-turn video training and the actor's own hint-generation capability is insufficient. Verified preceding turns limit error propagation and keep these hints aligned with the history. The target continuation still comes from the actor, and teacher targets remain dynamically updated. These choices motivate the hint pipeline in Point 2 and the prefix and teacher controls in Point 4.

We thank the reviewer again for raising these important questions. We will revise the paper and we are looking forward to your responce.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely thank all reviewers for their constructive feedback. During the rebuttal, we conducted additional experiments and provided clarifications to address their concerns:

  1. Privileged hints (SwdG W1/W2; a52g Q1; qmWh W1/Q1)

We demonstrated the effectiveness of hint-conditioned teachers (fully-correct trajectories: 14.24% → 96.69%) and conducted hint-component ablations, confirming the importance of both stream descriptions and response/silence recommendations.

  1. Core component analysis (SwdG W4; a52g W1/Q1; qmWh W1/Q1/Q2)

We added prefix-source and teacher-update ablations. Self-generated preceding histories reduce the average score from 59.64 to 57.20, while fixed teacher targets achieve 59.28.

  1. Temporal validity of hints (a52g Q3; qmWh W1/Q1)

We clarified that hint generation only accesses video frames available up to the current decision time, preventing future-information leakage by construction.

  1. Additional comparisons and clarifications (qmWh W2/Q3; SwdG W3)

We added a GRPO comparison (59.64 vs. 52.93) and clarified the complementary roles of OVO-Bench/StreamingBench and proactive-response benchmarks.

  1. Reproducibility and stability (a52g W2/Q2)

We reported three-seed results (59.59 ± 1.05), clarified the annotation model, training data sources, filtering statistics, and implementation details.

We believe these additional results and clarifications substantially strengthen the empirical validation, methodological analysis, and reproducibility of Live-OPSD. We sincerely thank all reviewers for helping us improve the paper.

Best Regards,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.