acceptodds
Under review as a conference paper at ICLR 2027

Tempo-Zero: Programmatic Self-Evolution for Video Understanding

Abstract

Video reasoning requires understanding how events unfold over time, yet videos with accurate temporal annotations are expensive to collect. Self-play offers an alternative by letting a model generate and solve its own tasks without manual annotation. However, existing video self-play methods train on fixed sets of recorded videos, so they cannot alter the underlying event dynamics and rely on noisy pseudo-labels derived from the model's own outputs, such as majority voting, that risk reinforcing shared model errors. To overcome these limitations, we introduce Tempo-Zero, a self-improvement framework that closes a loop of executable environments, execution-verified self-play, and environment evolution, where executing a program renders a video with the ground-truth option and interval of every question. A single policy plays both Questioner and Solver, proposing questions and answering them with timestamped evidence lines, and is trained on execution-grounded rewards instead of model consensus. The Solver's performance steers a curriculum over tasks and difficulty levels, and the policy selects rule changes that enrich the event dynamics of later rounds. Despite being trained solely on simple programs, without supervised fine-tuning or external video data which prior methods typically require, Tempo-Zero outperforms the base models and recent self-play baselines across twelve real-world video benchmarks at different model scales, with particularly strong gains on temporal grounding tasks. Code will be made publicly available.

Primary area:
foundation or frontier models, including LLMs
Keywords:
video reasoning, recursive self-improvement, self-play, vision-language models

Reviews

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

This paper introduces Tempo-Zero, a framework that treats executable programs as environments for video-based self-play. As a program executes, it simultaneously renders a video and records the sequence of events, from which questions, correct answers, and the specific temporal intervals serving as evidence are derived. A single vision-language policy functions as both the questioner and the solver; instead of relying on majority voting or an external model-based judge, the framework uses the success of program execution as a reward signal. Furthermore, the environment evolves through a curriculum based on success rates and dynamic adjustments to the rules governing policy selection.

Critical strengths

The paper clearly articulates the limitations, such as issues with supervision or adaptability, inherent in existing video self-play methods and program-generated video approaches. It proposes using program execution to simultaneously control video content and provide verifiable supervision.

Execution traces provide both the correct answers and the corresponding temporal evidence intervals. By re-evaluating sampled frames and checking for temporal dependencies, the framework ensures that labels remain valid for the specific frames visible to the model, thereby reducing noise in consensus-based labeling.

Critical weaknesses

The environment suite, task patterns, and rule sets are predefined by the authors. The range of physical behaviors and task types is limited, making it difficult to capture the open-ended events, common-sense relationships, and perceptual complexity found in natural, real-world videos.

The supervision mechanism relies on deterministic rendering and precise temporal coordinates. In contrast, real-world video events often feature ambiguous boundaries, frequent occlusions, overlapping events, and complex camera motion; thus, the precision of execution-based labels does not necessarily equate to a robust capability for understanding real-world video.

The reliance on fixed syntactic parsing, statement matching, and temporal interval overlap means that only specific phrasings and granularities are accepted, potentially leading to the rejection of valid evidence that is phrased differently. Additionally, the gating mechanism may inadvertently filter out legitimate questions, and the multi-component reward structure requires significant hyperparameter tuning.

Questions for authors

Please refer to the weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer YNoG (1/2)

Thank you very much for your careful review and for recognizing that our frame-recomputed execution traces keep labels valid for the exact frames visible to the model while reducing consensus noise. Below we address your first two concerns on synthetic environment scope and real-video robustness.

W1 (Questions For Authors: Please refer to the weaknesses): The environment suite, task patterns, and rule sets are predefined by the authors. The range of physical behaviors and task types is limited, making it difficult to capture the open-ended events, common-sense relationships, and perceptual complexity found in natural, real-world videos.

Answer: As you note, rendered synthetic scenes do not attempt to replicate natural visual textures or open-world encyclopedic knowledge. Visual recognition of everyday objects, scenes, and common-sense priors comes from the pretrained vision-language backbone, whereas executable programs cultivate transferable temporal skills: tracking multi-entity state changes, resolving event order and duration, maintaining entity continuity across occlusions, counting recurring transitions, and anchoring answers to verified evidence intervals [1, 2]. At zero-shot test time, these temporal composition skills combine naturally with pretrained perceptual priors across all 12 real-world benchmarks and all four backbones in Tables 1 and 2, while we also discuss manual program design scope in our Limitations section. Under the same 3,200-instance, 100-step budget and verification interface, our 20-family suite retains all 12 original families and adds 8 new executable families. Four of them model multi-process occlusion via lane-crossing handoffs, hooded cascades, extended-occluder re-identification, and patrol beacons, and the other four model grid kinematics via atomic motion, 2D waypoints, long-horizon state memory, and terminal tracking.

Environment diversity on Qwen3-VL-4BTemporal GroundingVideo QA & ReasoningOverall
4B Base25.0854.6844.81
12 families37.0057.7150.81
20 families41.5358.2552.68

Under matched compute, broadening the executable suite from 12 to 20 families improves both real-video temporal grounding and video QA & reasoning on 4B.

W2: The supervision mechanism relies on deterministic rendering and precise temporal coordinates. In contrast, real-world video events often feature ambiguous boundaries, frequent occlusions, overlapping events, and complex camera motion; thus, the precision of execution-based labels does not necessarily equate to a robust capability for understanding real-world video.

Answer: Deterministic execution prevents consensus drift during self-play, where Figure 7A shows majority voting errs on over half of collected groups, while our training signals teach the model to track state transitions across frames rather than latch onto isolated single-frame cues. Because visibility is recomputed on sampled frames, Equation 3 scores continuous temporal IoU conditioned on matching the right event first, and dedicated families model full visual occlusion, the trained policy compares start-to-end state changes across the sampled sequence. On natural videos, this helps follow gradual action onsets and completions, bridge brief occlusions via pre- and post-occlusion continuity, separate concurrent interleaved actors by event semantics, and stay anchored to subject actions during camera pans, as illustrated by verified 4B test predictions in Appendix F.

Real-video challengeBenchmark sample and ground truthBase model versus Tempo-Zero
Ambiguous action boundaryCharades-STA NO1GJ.mp4, id 347, GT 15.96 to 25.33 sBase truncates early at 8.0 to 19.7 s with IoU 0.22; Tempo-Zero aligns 16.0 to 24.9 s with IoU 0.95
Frequent visual occlusionNExT-GQA 8277189448_1, GT 8.7 to 11.8 sBoth choose option D; Base predicts 6.0 to 10.0 s with IoU 0.22, while Tempo-Zero locates 8.7 to 12.8 s with IoU 0.76
Overlapping interleaved eventsANet-RTL v_jpWevi1HBYo.mp4, id 191, GT 9.07 to 16.95 sBase clips opening frames 7.0 to 10.0 s with IoU 0.09; Tempo-Zero spans 8.0 to 16.4 s with IoU 0.82
Continuous camera motionActivityNet v_Y7VWbYGI0Oc.mp4, id 3831, GT 3.76 to 24.30 sBase stops during initial pan at 0.0 to 6.0 s with IoU 0.09; Tempo-Zero tracks 5.0 to 24.0 s with IoU 0.93

References

[1] Kexin Yi et al. CLEVRER: Collision Events for Video Representation and Reasoning. In ICLR, 2020.

[2] Songtao Jiang et al. Learning Transferable Temporal Primitives for Video Reasoning via Synthetic Videos. In CVPR, 2026.

AuthorsOpenReview ↗

Response to Reviewer YNoG (2/2)

Thank you again for your thoughtful comments on evidence flexibility, question gating, and hyperparameter sensitivity. Below we address each of these points with direct empirical audits.

W3 (part 1): The reliance on fixed syntactic parsing, statement matching, and temporal interval overlap means that only specific phrasings and granularities are accepted, potentially leading to the rejection of valid evidence that is phrased differently.

Answer: Statement matching prevents unrelated guesses from harvesting interval overlap credit, where Figure 7B shows roughly half of unconditioned IoU credit lands on wrong events, while an unmatched evidence line never reduces the main answer reward. To test sensitivity to alternative phrasings directly, we evaluated five controlled 100-statement rewrite probes built from accepted 4B evidence lines, measuring false-pass rates strictly on statements false against every fact in the video, as detailed in Appendix D.4.

Controlled probe, N = 100Truly false statementsOutcome or false-pass rateAudit detail
Natural paraphrase091.0% recall91 of 100 match the original fact with 0 misassigned
Non-instance number990.0% false pass0 of 99 accepted
Another number in same video460.0% false pass0 of 46 accepted; remaining 54 match another true fact
Another entity in same video620.0% false pass0 of 62 accepted; remaining 38 match that entity's true state
Insert predicate negation "not"1001.0% false pass1 of 100 accepted; word-boundary filter blocks 99% of inversions

Across all 314,619 evidence lines generated in the 4B run, only 41 of 3,276 lines containing negation terms were accepted, or 1.25%, nearly all of them compound sentences whose affirmative clause stated a true observation. Grounded scoring also uses continuous temporal IoU rather than rigid frame equality, and Table 3 confirms that both process terms complement answer supervision on Qwen3-VL-4B, while the revised Appendix G outlines future work pairing execution rubrics with lightweight typed decision models such as Jev [1, 2] for free-form multi-granularity evidence.

Process supervision ablation on 4BTemporal GroundingVideo QA & ReasoningOverall
4B Base25.0854.6844.81
Answer reward only31.3056.2847.96
w/o evidence score35.3457.3350.00
w/o grounded score36.6357.1150.28
Full Tempo-Zero37.0057.7150.81

W3 (part 2): Additionally, the gating mechanism may inadvertently filter out legitimate questions, and the multi-component reward structure requires significant hyperparameter tuning.

Proposal gating. Canonical questions and bank-replay items enter Solver training directly without proposal gating, while online Questioner gate pass rates rise quickly above 90% in Figure 6C. Manual inspection of 100 randomly sampled gated-out proposals in Appendix D.4 found objective supervision-corrupting defects in all 100 cases, with zero valid questions rejected.

Defect category in gated-out proposals, N = 100ShareObserved failure modes
Malformed structure or duplicate options56%Wrong option count, truncated output missing answer block, duplicate or symmetric options, placeholder text
Out-of-range or unresolvable options33%Out-of-domain distractors, ground truth omitted, options closer than sampled frame resolution
Answer leakage or missing qualifiers11%Answer leaked in question stem, missing time window or ordinal qualifier causing multiple solutions

Reward hyperparameters. Equation 4 avoids component-wise weight tuning by taking unweighted arithmetic averages over unit-normalized Solver process terms and Questioner terms, with quadratic learnability peaking naturally at 0.5 pass rate. The remaining constants are standard curriculum pass-rate bounds, clipping ratios, and KL regularization in Table 5, fixed once on Qwen3-VL-4B and transferred unchanged to Qwen3-VL-8B, Qwen3-VL-30B-A3B, and Gemma-4-E4B-it without per-backbone retuning.

Thank you again for your constructive feedback. We hope these analyses clarify the scope and robustness of Tempo-Zero, and we are happy to address any further questions.

References

[1] Diogo Almeida. Introducing System One Models & Jev. TypeSafe AI Blog, September 2026. https://typesafe.ai/blog/introducing-system-one-models-and-jev

[2] Guanxu Yu and Yuhang Yao. Visual Jev: Accurate and Efficient Decisions from Shared Visual Context. arXiv:2609.25845, 2026.

Dear Reviewer YNoG,

We hope you are doing well, and thank you again for your careful review and helpful comments. We understand how busy things can be, so we really appreciate you taking a look at our response when you get a chance. If you feel that our clarifications have addressed your main concerns, we would be very grateful if you could consider revisiting your overall assessment. Please also let us know if any questions remain.

Thank you very much!

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

This paper proposes Tempo-Zero, a video self improvement framework combining executable environments, execution-verified self-play, and environment evolution. Executable programs generate synthetic videos with verifiable answers and temporal intervals. A shared policy acts as both Questioner and Solver, while a curriculum based on pass rates and policy-selected rule changes adapt the training environments. Experiments on 12 real world video benchmarks demonstrate improvements.

Critical strengths
  1. Meaningful research problem and well supported motivation. The paper addresses two important limitations of existing video self-play methods: unreliable pseudo labels and fixed event dynamics.
  2. A framework with good results. Combining executable environments with self-play and environment evolution is a promising approach to video self improvement. The gains across real-world benchmarks provide evidence that training on synthetic videos can transfer to real-video understanding, particularly temporal grounding. The ablations also support the effectiveness of the Questioner and evidence based Solver rewards.
Critical weaknesses
  1. Some central contributions could be more directly validated. Figure 7A demonstrates the unreliability of majority-vote labels on the collected training questions, but a controlled downstream comparison using execution-derived versus majority-vote labels is not reported. Similarly, Table 3 and the bank-data fine-tuning experiment in Table 4 support the usefulness of changing environment rules, but both compare policy-selected changes against fixed initial rules. They therefore do not establish whether policy-based rule selection outperforms simpler alternatives, such as randomly selecting valid changes from the same rule menus.
  2. Presentation issues. Some figure captions overlap with the surrounding text or equations. These layout issues affect readability and should be corrected.
Questions for authors
  1. Could the authors compare execution-derived and majority-vote answer labels, and report their impact on downstream performance? It would be helpful to clarify how evidence rewards and curriculum feedback are handled in this comparison.
  2. Do policy-selected rule changes outperform random valid changes drawn from the same rule menus, under matched training budgets and comparable rule-change frequencies? This would help show the value of adaptive rule selection beyond the benefit of introducing more diverse event dynamics.
3 comments
AuthorsOpenReview ↗

Response to Reviewer QmRH

We sincerely thank you for your encouraging review and for recognizing how Tempo-Zero addresses key limitations of prior video self-play methods across 12 real-world benchmarks. Below we address each of your comments and questions.

W1 & Q1: Some central contributions could be more directly validated. Figure 7A demonstrates the unreliability of majority-vote labels on the collected training questions, but a controlled downstream comparison using execution-derived versus majority-vote labels is not reported.

  1. Could the authors compare execution-derived and majority-vote answer labels, and report their impact on downstream performance? It would be helpful to clarify how evidence rewards and curriculum feedback are handled in this comparison.

Answer: Thank you for this insightful suggestion. Under the same training setup on Qwen3-VL-4B and Qwen3-VL-30B-A3B, we replaced execution supervision with 8-rollout majority-vote supervision while leaving everything else unchanged, as detailed in Appendix E.3. In this majority-vote run, each question's target answer uses the 8-rollout majority choice. Evidence and grounding rewards score predicted intervals against consensus intervals from agreeing rollouts rather than executed intervals, while curriculum pass rates track agreement with the consensus label.

Supervision sourceTemporal GroundingVideo QA & ReasoningOverall
4B Base25.0854.6844.81
4B Majority-vote35.9355.4648.95
4B Tempo-Zero37.0057.7150.81
30B-A3B Base34.3861.5052.46
30B-A3B Majority-vote39.7562.6255.00
30B-A3B Tempo-Zero41.1263.8956.30

Most of the grounding jump over Base comes from structured timestamped evidence training on program-rendered videos, with execution intervals adding a further margin, whereas exact execution supervision primarily drives Video QA & Reasoning; we refined Section 4.2 to reflect this attribution. Majority voting loses roughly one-third of the overall gain on both backbones and over half to nearly three-quarters of the QA gain by reinforcing shared multi-step errors across rounds.

W1 & Q2: Similarly, Table 3 and the bank-data fine-tuning experiment in Table 4 support the usefulness of changing environment rules, but both compare policy-selected changes against fixed initial rules. They therefore do not establish whether policy-based rule selection outperforms simpler alternatives, such as randomly selecting valid changes from the same rule menus. 2. Do policy-selected rule changes outperform random valid changes drawn from the same rule menus, under matched training budgets and comparable rule-change frequencies? This would help show the value of adaptive rule selection beyond the benefit of introducing more diverse event dynamics.

Answer: Thank you for suggesting this informative baseline. On both Qwen3-VL-4B and Qwen3-VL-30B-A3B under the same 5-round training budget, edit frequency, and preview validation, we compared policy-selected rule updates against drawing valid edits uniformly at random from the same rule menus, reported in Appendix E.3.

Rule selection comparisonTemporal GroundingVideo QA & ReasoningOverall
4B Base25.0854.6844.81
4B Random menu rules36.6355.8449.44
4B Tempo-Zero37.0057.7150.81
30B-A3B Base34.3861.5052.46
30B-A3B Random menu rules40.0562.9355.30
30B-A3B Tempo-Zero41.1263.8956.30

On 4B, random menu edits roughly match keeping fixed initial rules in Table 3 on Overall score, indicating that unguided rule diversity alone adds little. Relative to random menu selection, policy-selected updates keep 4B grounding comparable within run-to-run noise while clearly lifting Video QA & Reasoning and Overall, and improve grounding, QA, and Overall on 30B-A3B. Conditioning rule updates on pass-rate and evidence bottlenecks thus provides targeted gains beyond rule diversity alone, consistent with adaptive curriculum design [1, 2].

W2: Presentation issues. Some figure captions overlap with the surrounding text or equations. These layout issues affect readability and should be corrected.

Answer: Thank you for catching these layout issues. In Section 3.3 of the revised manuscript, we shortened the caption of Figure 5 and its text wrap so that the caption ends above Equation 6 instead of running beside it, and we rechecked the captions of Figures 3 and 4; the main paper stays within the 10-page limit.

We are happy to address any further questions you may have.

References

[1] Minqi Jiang, Edward Grefenstette, and Tim Rocktäschel. Prioritized Level Replay. In ICML, 2021.

[2] Jack Parker-Holder et al. Evolving Curricula with Regret-Based Environment Design. In ICML, 2022.

Reviewer_QmRHOpenReview ↗

Comment for the rebuttal

Thank you for the response and additional experiments. The response have largely addressed my main concerns. I believe the paper is worthy of a Weak Accept and will maintain my original rating.

Thanks again for maintaining your positive rating, and we are glad our response has addressed your main concerns. Your thoughtful comments have really helped make our work more complete.

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

This paper introduces Tempo-Zero, a self-improvement framework for video reasoning using executable synthetic environments. Program execution generates videos together with verifiable answers and temporal evidence intervals. A shared vision-language policy acts as both Questioner and Solver and is optimized through execution-grounded rewards. Task allocation and difficulty adapt to Solver performance, while the policy selects validated rule changes to evolve the environments. Experiments on twelve real-world video benchmarks and four backbones demonstrate synthetic-to-real transfer, particularly for temporal grounding. Component ablations, training diagnostics, and supervised fine-tuning on collected data provide additional validation.

Critical strengths
  1. A useful integration of executable supervision and adaptive video self-play. Generating videos, answers, and evidence intervals from the same execution trace provides a concrete alternative to model-consensus supervision. The sampled-frame verification and event-conditioned temporal reward address meaningful difficulties in obtaining reliable video reasoning feedback.

  2. Convincing transfer from simple synthetic environments to real videos. The improvements span multiple backbones and twelve benchmarks, with particularly substantial temporal grounding gains: the average score increases from 25.08 to 37.00 at 4B and from 30.28 to 37.97 at 8B. The canonical-question-only and fixed-rule ablations also support the aggregate contributions of proposed questions and environment changes.

Critical weaknesses
  1. Evidence for sustained scaling is limited. Training uses twelve hand-designed environment families, fixed task schemas and rule menus, and 100 updates. The intermediate evaluations track synthetic held-out performance, but do not establish how real-video performance scales with environment diversity or larger training budgets. Consequently, it remains unclear whether the framework supports sustained improvement as training proceeds or becomes constrained by its predefined task and rule space. This limits the evidence for its broader significance as a continuing self-improvement framework.

  2. The constrained reasoning and verification interface leaves important questions about adaptability. Self-play training uses multiple-choice questions and timestamped evidence statements evaluated through fixed grammatical and event-matching rules. Precise execution labels do not automatically establish the semantic reliability of matching generated language to those labels. The paper does not quantify sensitivity to valid paraphrases or incorrect statements containing relevant terms. Although downstream transfer is demonstrated, the extent to which the interface accommodates more diverse reasoning statements and non-multiple-choice answer formats remains insufficiently characterized.

Questions for authors
  1. Scaling — additional validation. How does real-video performance change with training budget and environment diversity? Results across the existing checkpoints would be informative. A modest extension beyond 100 updates or an environment-subset comparison under matched compute would further clarify whether additional training and environmental diversity yield continued gains. Please explain how the difficulty, playback-duration, and rule-evolution schedules would be extended, and report total rollout costs, including proposals evaluated but not selected for Solver training.

  2. Evidence and grounding — clarification and diagnostic analysis. Table 3 shows that removing the evidence score hurts temporal grounding more than removing the grounded score, while the QA ordering is reversed. The QA difference is small; is this pattern reproducible? The existing full-model curves report evidence F1 and joint grounded scores. Comparing event-matching precision/recall and IoU conditioned on successful matching across these ablations would help distinguish event identification from localization.

  3. Questioner training — controlled ablation. The canonical-question-only ablation already demonstrates the aggregate benefit of the Questioner component. Could you retain Questioner-generated questions while removing the Questioner-role training objective, with matched Solver training volume? Since the roles share parameters, the generator would still change through Solver updates; this control would specifically test the additional value of explicitly optimizing question generation. A few paired canonical and generated questions would also help explain what the Questioner contributes.

  4. Format dependence and generality — scope clarification and future directions. Given the temporal-localization results in Table 1 and the typed answers listed in Table 6, which components specifically depend on multiple-choice formatting during self-play training? Would training on direct numerical, sequence, or interval outputs require only changes to parsing and answer rewards, or also changes to the criteria and verification machinery? For more flexible natural-language evidence, what extensions do the authors consider feasible while preserving reliable verification? Please distinguish current capabilities from future directions.

6 comments
AuthorsOpenReview ↗

Response to Reviewer fjNk (1/3)

Thank you for your thoughtful review and for recognizing our executable supervision and synthetic-to-real transfer. Below we address W1 and Q1.

W1: Evidence for sustained scaling is limited. Training uses twelve hand-designed environment families, fixed task schemas and rule menus, and 100 updates. The intermediate evaluations track synthetic held-out performance, but do not establish how real-video performance scales with environment diversity or larger training budgets. Consequently, it remains unclear whether the framework supports sustained improvement as training proceeds or becomes constrained by its predefined task and rule space. This limits the evidence for its broader significance as a continuing self-improvement framework. Q1 (part 1): Scaling: additional validation. How does real-video performance change with training budget and environment diversity? Results across the existing checkpoints would be informative. A modest extension beyond 100 updates or an environment-subset comparison under matched compute would further clarify whether additional training and environmental diversity yield continued gains.

Training budget and checkpoints. We evaluated Rounds 1 to 5 on Qwen3-VL-4B and extended training under the same setup to Round 10.

4B roundR1R2R3R4R5R6R7R8R9R10
ANet-RTL mIoU22.0624.5329.3730.2331.3430.4729.0130.1731.1233.70
NExT-GQA33.5334.7735.0436.4139.1740.4840.9341.2840.5938.92
LongVideoReason71.9071.1073.6072.1072.4072.6070.9071.2070.8071.00
Holdout %27.932.732.433.231.132.231.332.235.237.0

Grounding climbs steadily through Round 5 while long-video reasoning stays stable. Extending to Round 10 lifts multi-event localization and holdout accuracy further, with a mild late trade-off on open QA once the fixed 12-family suite and capped playback scale lean toward fine localization.

Environment diversity. Under the same 3,200-instance, 100-step budget and verification interface, our 20-family suite retains all 12 original families and adds 8 executable families. Four of them model multi-process occlusion via lane handoffs, hooded cascades, extended-occluder re-identification, and patrol beacons, and the other four model grid kinematics via atomic motion, 2D waypoints, long-horizon memory, and terminal tracking.

Environment suite on 4BTemporal GroundingVideo QA & ReasoningOverall
4B Base25.0854.6844.81
12 families37.0057.7150.81
20 families41.5358.2552.68

Under the same compute budget, expanding from 12 to 20 families improves both Temporal Grounding and Video QA & Reasoning on 4B.

Q1 (part 2): Please explain how the difficulty, playback-duration, and rule-evolution schedules would be extended, and report total rollout costs, including proposals evaluated but not selected for Solver training.

Schedule extension. Extending past Round 5 adds no hyperparameters: difficulty steps at the 0.80 and 0.20 pass-rate bounds in Section 3.3 while cycling option counts, rule menus prioritize unvisited combinations, and playback scale stays capped at 2.0 so sampled frames preserve brief actions.

Rollout accounting. Truncated and gate-rejected proposals are filtered on CPU without video inference, while Figure 6C plots gate pass rates over parsed non-truncated proposals. Every unique gated proposal receives 8 Solver rollouts for Questioner pass-rate scoring; at most one preselected proposal per instance enters the Solver pool alongside canonical and replay items, where non-zero-variance groups update the policy directly without re-sampling, per Appendix A.3.

4B rollout stageCountRollout role and reuse
Questioner proposals12,7922,209 CPU-rejected incl. truncated outputs, 0 rollouts; 10,583 deduplicated to 5,052
Gated proposal rollouts40,4165,052 unique proposals at 8 rollouts: 2,435 preselected for Solver, 2,617 Questioner-only
Canonical & replay rollouts31,8243,186 canonical and 792 replay questions, 8 rollouts each in the Solver pool
Total Solver rollouts72,24046,408 in 5,801 Solver groups updated; 4,896 in 612 zero-variance and 20,936 in 2,617 unselected groups
AuthorsOpenReview ↗

Response to Reviewer fjNk (2/3)

Thank you again for these diagnostic questions. Below we address Q2 and Q3 on reward synergy, matching diagnostics, and Questioner optimization.

Q2 (part 1): Evidence and grounding: clarification and diagnostic analysis. Table 3 shows that removing the evidence score hurts temporal grounding more than removing the grounded score, while the QA ordering is reversed. The QA difference is small; is this pattern reproducible?

Answer: On Video QA, dropping either evidence or grounded score alone gives nearly identical accuracy within run-to-run variation across three 4B runs, so Section 4.3 removes that QA ordering claim. On Temporal Grounding, Equation 3 gates interval overlap on matching a valid statement first: dropping statement matching detaches intervals from verified events and hurts localization more than dropping overlap alone, while removing both causes a compounded drop.

Q2 (part 2): The existing full-model curves report evidence F1 and joint grounded scores. Comparing event-matching precision/recall and IoU conditioned on successful matching across these ablations would help distinguish event identification from localization.

Answer: Thank you for this helpful suggestion. On the held-out set behind Figure 6B and Table 10, we decompose event-statement precision, recall, and conditional IoU across the 4B ablations in Table 3, added to Appendix D.2.

4B held-out diagnosticsPrecisionRecallCond. IoU
4B Base25.4%23.3%13.2%
w/o novelty reward45.2%45.8%21.9%
w/o learnability reward42.7%42.1%20.4%
w/o Questioner38.4%35.2%18.1%
w/o evidence score34.6%38.9%21.1%
w/o grounded score45.4%46.5%17.6%
Answer reward only28.5%26.8%15.4%
w/o rule selection42.9%41.3%20.1%
Full Tempo-Zero46.3%47.8%22.6%

This decomposition separates event identification from boundary localization. Dropping the evidence score mainly lowers precision and recall while conditional IoU on matched events stays high, whereas dropping the grounded score keeps statement matching high but lowers conditional IoU, and the Answer-reward-only variant degrades both sharply. Removing the Questioner causes a large drop in event recall alongside weaker conditional IoU, while removing learnability or rule selection moderately reduces both and removing novelty has only a minor effect.

Q3: Questioner training: controlled ablation. The canonical-question-only ablation already demonstrates the aggregate benefit of the Questioner component. Could you retain Questioner-generated questions while removing the Questioner-role training objective, with matched Solver training volume? Since the roles share parameters, the generator would still change through Solver updates; this control would specifically test the additional value of explicitly optimizing question generation. A few paired canonical and generated questions would also help explain what the Questioner contributes.

Controlled ablation without Questioner updates. Thank you for this clean control. Under the same shared-weight 4B setup and matched Solver volume, we kept online Questioner proposals while disabling Questioner-role updates so the shared policy updates only on Solver trajectories.

4B Questioner controlTemporal GroundingVideo QA & ReasoningOverall
4B Base25.0854.6844.81
w/o Questioner, matched volume32.5156.8848.76
Solver-only updates34.8257.0649.65
Full Tempo-Zero37.0057.7150.81

Online proposals under Solver-only updates already lift grounding and QA over the matched-volume canonical control via shared Solver updates, while updating the Questioner role further improves both by steering proposals toward novel criteria at intermediate pass rates.

Paired canonical and generated questions. Three real same-video pairs from logs illustrate multi-event and distractor composition, per Appendix D.4.

FamilyCanonical questionQuestioner-generated proposal
DominoesIn 20.46 to 48.87 s, what is the net propagation direction of the falling frontier?What is the temporal relation between the top chain's 10th to 16th and bottom chain's 11th to 12th dominoes?
Checkout queueIn 66.89 to 120.40 s, which customer shape departed from a lane most frequently?In what order did the brown square, cyan circle, cyan square, and red square customers depart?
Block stackingOver what interval does the center pad's block count stay unchanged between placements?How does the span from placement 1 to 3 relate to the span from placement 2 to 4?
AuthorsOpenReview ↗

Response to Reviewer fjNk (3/3)

Thank you again for raising these constructive questions on verification generality and format adaptability. Below we address semantic reliability of statement matching, answer-format dependence, and extensions toward open-ended natural-language evidence.

W2: The constrained reasoning and verification interface leaves important questions about adaptability. Self-play training uses multiple-choice questions and timestamped evidence statements evaluated through fixed grammatical and event-matching rules. Precise execution labels do not automatically establish the semantic reliability of matching generated language to those labels. The paper does not quantify sensitivity to valid paraphrases or incorrect statements containing relevant terms. Although downstream transfer is demonstrated, the extent to which the interface accommodates more diverse reasoning statements and non-multiple-choice answer formats remains insufficiently characterized.

Answer: To evaluate semantic reliability directly, we audited five controlled rewrite probes of 100 statements each built from accepted 4B evidence lines, as reported in Appendix D.4. When substituting a number or entity happens to yield another true event in the same video, matching that true event is counted as valid, and false-pass rates are measured strictly on statements that are false against every fact in the video.

Controlled probe, N = 100Truly false statementsOutcome or false-pass rateAudit detail
Natural paraphrase091.0% recall91 of 100 match the original fact with 0 misassigned
Non-instance number990.0% false pass0 of 99 accepted
Another number in same video460.0% false pass0 of 46 accepted; remaining 54 match another true fact
Another entity in same video620.0% false pass0 of 62 accepted; remaining 38 match that entity's true state
Insert predicate negation "not"1001.0% false pass1 of 100 accepted; word-boundary filter blocks 99% of inversions

Across all 314,619 evidence lines generated during the full 4B training run, only 41 of the 3,276 lines containing negation terms were accepted, or 1.25%, and nearly all were compound sentences whose affirmative clause stated a true observation.

Q4: Format dependence and generality: scope clarification and future directions. Given the temporal-localization results in Table 1 and the typed answers listed in Table 6, which components specifically depend on multiple-choice formatting during self-play training? Would training on direct numerical, sequence, or interval outputs require only changes to parsing and answer rewards, or also changes to the criteria and verification machinery? For more flexible natural-language evidence, what extensions do the authors consider feasible while preserving reliable verification? Please distinguish current capabilities from future directions.

Format dependence in current training. In our current loop, evidence lines already emit direct start-to-end timestamp intervals scored by continuous IoU, while final answers remain multiple-choice. Only three surface steps depend on multiple-choice formatting: option-domain and spacing checks in proposal gating, option-letter parsing, and binary option matching in the answer reward. Training on direct numerical, sequence, or interval answers therefore requires updating only surface answer parsing, typed answer scoring, and proposal-gate format checks, while program execution, frame-recomputed criteria, and evidence verification stay unchanged.

Flexible natural-language evidence. Currently, our deterministic slot-normalized matcher covers natural paraphrases with zero judge-model memory or latency overhead. As future work noted in the revised Appendix G, unconstrained free-form evidence could pair structured event rubrics from exact program traces with lightweight typed decision models such as Jev [1, 2], encoding the execution rubric once to return calibrated entailment decisions over atomic event claims without open-ended judge generation.

We are happy to answer any further questions.

References

[1] Diogo Almeida. Introducing System One Models & Jev. TypeSafe AI Blog, September 2026. https://typesafe.ai/blog/introducing-system-one-models-and-jev

[2] Guanxu Yu and Yuhang Yao. Visual Jev: Accurate and Efficient Decisions from Shared Visual Context. arXiv:2609.25845, 2026.

Reviewer_9TuxOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Tempo-Zero, a video self-play framework that trains on videos rendered from executable programs rather than recorded videos. Program execution gives exact answers and event intervals, which are checked against the sampled frames. One policy plays both Questioner and Solver: the Questioner rewrites verified facts into new questions, and the Solver answers with timestamped evidence, rewarded for both the answer and correctly located events. A pass-rate curriculum and policy-chosen rule changes adjust the environments over training. Trained only on synthetic videos, Tempo-Zero improves four backbones on twelve real-video benchmarks and beats VisPlay, V-Zero, and Video-Zero on most metrics, with the largest gains in temporal grounding (+7.69 average at 8B). However, the experiments do not isolate which part of the recipe drives these gains, and the advantage over majority voting is argued rather than tested.

Critical strengths
  • Training only on simple rendered videos transfers well to real-video temporal grounding. At 8B, Charades-STA mIoU goes from 36.70 to 46.10 without any real interval annotation (Table 1), and the gains hold on the other benchmarks and across four backbones. This is the main reason I lean toward acceptance.
  • The supervision is built carefully. Labels are recomputed on the sampled frames, order-independent questions are filtered out, and interval credit is given only on lines that describe the right event. Figure 7B shows this matters, since about half of the naive IoU credit falls on lines about other events.
Critical weaknesses
  • My main concern is that the paper does not show where the gains come from. Compared with Video-Zero, Tempo-Zero changes the video source, the label source, the output format, and the rewards all at once, yet the paper credits the grounding gains to exact execution labels. The majority-voting analysis does not test this either. It only shows that consensus labels would be poor on synthetic tasks where the base model is right about a quarter of the time, not that they would hurt training. Table 4 adds to my doubt: one epoch of SFT on the bank data already gets most of the way (48.65 vs. 50.81 Overall). I would like to see the same pipeline trained with majority-vote labels, and a simple fixed-program synthetic baseline.
  • I am not convinced the ablations support the claimed mechanisms. Each variant is a single 4B run, and several gaps are under one point. Removing the Questioner also removes part of the Solver’s training questions, so its large drop may partly reflect less data. Rule selection is only compared with keeping the initial rules. Since every menu option is valid, random selection from the same menu is the obvious baseline, and without it the gain may come from rule diversity rather than the policy’s choices.

Other concerns:

  • It is not clear whether the baselines were re-trained or copied, or how many frames they use. VisPlay and V-Zero are image methods and barely move from the base model, so they add little.
  • Some of the grounding gain may simply come from learning the [[start, end]] output format. Parse-failure rates before and after training would help.
  • The Questioner is given the criteria and answers and mostly rewrites questions, so “constructing tasks” overstates its role.
  • Compute is reported only as “eight GPUs”.
  • At 4B, held-out accuracy drops after round 4 (Table 10), but the round-5 checkpoint is always used.
Questions for authors

Questions that could change my assessment:

1.Can the authors train the same pipeline, on the same programs and budget, with majority-vote labels and intervals in place of execution labels? How much of the gain in Tables 1 and 2 disappears?

2.How does Tempo-Zero compare with a fixed-program synthetic-video GRPO baseline: canonical questions only, initial rules, no curriculum, same reward and budget? How does it compare with SynRL or SIMS-V data under the same backbone?

3.Can the authors report a w/o Questioner variant whose Solver sees the same number of questions per step, and a random-selection baseline over the same rule menus?

4.Can the main ablations in Table 3 be repeated over at least three seeds, with standard deviations? Do the sub-one-point differences remain clear?

Clarification requests:

5.Were VisPlay, V-Zero, and Video-Zero re-trained on the same backbones, and how many frames do all methods use at evaluation?

6.What are the answer and interval parse-failure rates of the base models and Tempo-Zero on the four grounding benchmarks?

7.In Table 4, are the SFT targets all recorded Solver responses or only correct ones, and which checkpoint generated them?

8.What GPU type and wall-clock time does one training run take, including rendering?

4 comments
AuthorsOpenReview ↗

Response to Reviewer 9Tux (1/3)

Thank you for your thoughtful review and for recognizing our synthetic-to-real transfer and supervision. Below we address W1, Q1, Q2, and Q7.

W1: My main concern is that the paper does not show where the gains come from. Compared with Video-Zero, Tempo-Zero changes the video source, the label source, the output format, and the rewards all at once, yet the paper credits the grounding gains to exact execution labels. [...] Q1: 1. Can the authors train the same pipeline, on the same programs and budget, with majority-vote labels and intervals in place of execution labels? How much of the gain in Tables 1 and 2 disappears?

Answer: Thank you for suggesting this test. Under the same setup, programs, and budget on 4B and 30B-A3B, we replaced only execution labels with 8-rollout majority-vote labels, leaving all else unchanged; see Appendix E.3.

Supervision sourceTemporal GroundingVideo QA & ReasoningOverall
4B Base25.0854.6844.81
4B Majority-vote35.9355.4648.95
4B Tempo-Zero37.0057.7150.81
30B-A3B Base34.3861.5052.46
30B-A3B Majority-vote39.7562.6255.00
30B-A3B Tempo-Zero41.1263.8956.30

We updated Section 4.2 accordingly: most of the grounding gain over Base comes from program-rendered videos with timestamped evidence, with execution intervals adding a further margin, while execution labels mainly drive Video QA & Reasoning. Majority voting loses about one-third of the overall gain on both backbones, and nearly three-quarters of the QA gain on 4B and over half on 30B-A3B.

Q2: 2. How does Tempo-Zero compare with a fixed-program synthetic-video GRPO baseline: canonical questions only, initial rules, no curriculum, same reward and budget? How does it compare with SynRL or SIMS-V data under the same backbone?

Answer: Thank you for this suggestion. On 4B under the same reward and budget, we ran fixed-program GRPO with canonical questions, initial rules, and no curriculum, plus static GRPO using SynRL [1] public programs. SIMS-V [2] targets spatial video QA and provides no event intervals for our evidence reward, so we did not train on it.

4B synthetic setupTemporal GroundingVideo QA & ReasoningOverall
4B Base25.0854.6844.81
Fixed-program GRPO30.0656.4247.63
SynRL static GRPO32.9456.5048.64
Tempo-Zero w/o Questioner31.8655.6347.71
Tempo-Zero37.0057.7150.81

Static baselines improve over Base, confirming synthetic transfer, yet saturate into zero-variance groups late in training without online questioning, curriculum, and rule evolution. Comparison with published SynRL [1] on all shared benchmarks:

Shared benchmarkSynRL [1]Tempo-Zero
Setup7.7K 235B-CoT SFT, 1K real videos + 7K RLNo SFT, no real videos, 3.2K synthetic
4B Charades-STA mIoU41.9 → 47.0, +5.137.52 → 42.53, +5.01
4B Charades-STA [email protected]20.5 → 23.0, +2.516.23 → 19.49, +3.26
4B Video-MME-Long51.1 → 54.5, +3.451.60 → 53.30, +1.70
4B VideoMMMU54.1 → 54.5, +0.457.56 → 63.11, +5.55
8B VideoMMMU60.0 → 61.0, +1.063.78 → 64.44, +0.66

SynRL needs 235B-distilled SFT plus real videos and gets sub-point RL gains without SFT; on under a quarter of the samples, Tempo-Zero matches its Charades-STA mIoU gain and leads on [email protected] and 4B VideoMMMU, while SynRL gains more on Video-MME-Long and slightly more on 8B VideoMMMU.

W1: Table 4 adds to my doubt: one epoch of SFT on the bank data already gets most of the way (48.65 vs. 50.81 Overall). Q7: 7. In Table 4, are the SFT targets all recorded Solver responses or only correct ones, and which checkpoint generated them?

Answer: Thank you for catching this, clarified in Section 4.4 and Appendix E.2. Table 4 uses all bank questions across five rounds paired with all execution-verified correct Solver trajectories sampled online at admission, not outputs from one checkpoint.

4B SFT vs. online RLTemporal GroundingVideo QA & ReasoningOverall
4B Base25.0854.6844.81
Fixed-program SFT28.9555.3146.52
Initial-rule bank SFT30.7356.7148.05
Evolved-rule bank SFT31.7857.0948.65
Tempo-Zero37.0057.7150.81

Bank SFT in Table 4 already benefits from trajectories curated online via questioning, curriculum, and evolved rules, whereas fixed-program SFT is weaker. The within-group contrast of online GRPO yields much stronger grounding than positive-only SFT.

References

[1] Songtao Jiang et al. Learning Transferable Temporal Primitives for Video Reasoning via Synthetic Videos. In CVPR, 2026.

[2] Ellis Brown, Arijit Ray, Ranjay Krishna, Ross Girshick, Rob Fergus, and Saining Xie. SIMS-V: Simulated Instruction-Tuning for Spatial Video Understanding. arXiv:2511.04668, 2025.

AuthorsOpenReview ↗

Response to Reviewer 9Tux (2/3)

Thank you again for your constructive feedback on experimental controls and ablation design. Below we respectfully address W2, Q3, and Q4.

W2 & Q3: Removing the Questioner also removes part of the Solver’s training questions, so its large drop may partly reflect less data. 3. Can the authors report a w/o Questioner variant whose Solver sees the same number of questions per step, and a random-selection baseline over the same rule menus?

Answer: Thank you for suggesting this controlled comparison. Under the same Qwen3-VL-4B training setup, we ran a matched-volume control by filling each step's Solver batch with compiled canonical questions so the Solver sees the same number of questions per step.

Questioner control on 4BTemporal GroundingVideo QA & ReasoningOverall
4B Base25.0854.6844.81
Tempo-Zero w/o Questioner31.8655.6347.71
Tempo-Zero w/o Questioner, matched volume32.5156.8848.76
Tempo-Zero37.0057.7150.81

Even at identical per-step question volume, Tempo-Zero clearly outperforms the canonical-only variant on both Temporal Grounding and Video QA & Reasoning. Beyond supplying more training questions, the online Questioner uses pass-rate feedback to target weak criteria and composes multi-event stems with fine-grained distractors on the same verified traces, improving both multi-step reasoning and boundary localization.

W2: Rule selection is only compared with keeping the initial rules. Since every menu option is valid, random selection from the same menu is the obvious baseline, and without it the gain may come from rule diversity rather than the policy’s choices.

Answer: Thank you for suggesting this baseline. On both Qwen3-VL-4B and Qwen3-VL-30B-A3B, we ran a controlled baseline under the same 5-round protocol that draws valid edits uniformly at random from the same rule menus at matched edit frequency, reported in Appendix E.3.

Rule selection comparisonTemporal GroundingVideo QA & ReasoningOverall
4B Base25.0854.6844.81
4B Random menu rules36.6355.8449.44
4B Tempo-Zero37.0057.7150.81
30B-A3B Base34.3861.5052.46
30B-A3B Random menu rules40.0562.9355.30
30B-A3B Tempo-Zero41.1263.8956.30

On 4B, random menu selection roughly matches keeping fixed initial rules in Table 3 on Overall score, showing that unguided rule diversity alone adds little. Compared with random menu rules, policy-selected rules leave 4B grounding essentially unchanged within run-to-run noise while clearly lifting Video QA & Reasoning and Overall, and improve all aggregates on 30B-A3B by targeting pass-rate and evidence bottlenecks.

W2 & Q4: I am not convinced the ablations support the claimed mechanisms. Each variant is a single 4B run, and several gaps are under one point. 4. Can the main ablations in Table 3 be repeated over at least three seeds, with standard deviations? Do the sub-one-point differences remain clear?

Answer: Thank you for raising this important point on cross-run variance. Given the compute and time budget during the discussion period, we prioritized running the new controlled ablations and baselines requested across reviewers on 4B and 30B-A3B, and completing three independent full 4B Tempo-Zero training runs under the same setup to establish the run-to-run variation scale, reported in Section 4.3 and Appendix E.3.

4B Tempo-Zero independent runsTemporal GroundingVideo QA & ReasoningOverall
Run 137.0057.7150.81
Run 236.1857.7950.59
Run 337.3256.6650.21
Mean ± SD36.83 ± 0.5957.39 ± 0.6350.54 ± 0.30

Two conclusions emerge from this comparison. First, major ablations in Table 3 fall clearly below full Tempo-Zero well beyond run-to-run variation, whereas sub-one-point gaps between individual ablations lie within single-run noise, so we have removed all sub-point ranking claims in Section 4.3. Second, the modules and reward terms in Table 3 work in synergy rather than in competition. Online questioning and rule selection keep supplying informative multi-event cases at intermediate difficulty. On those cases, statement matching and interval overlap jointly anchor temporal evidence. Removing a single sub-component therefore leaves partial support from its counterparts, whereas removing the Questioner, all process rewards, or rule selection causes a pronounced drop relative to the full method.

AuthorsOpenReview ↗

Response to Reviewer 9Tux (3/3)

Thank you for your questions on evaluation protocols, format parsing, Questioner wording, compute, and checkpoint selection.

Other concern 1 & Q5: It is not clear whether the baselines were re-trained or copied, or how many frames they use. VisPlay and V-Zero are image methods and barely move from the base model, so they add little. 5. Were VisPlay, V-Zero, and Video-Zero re-trained on the same backbones, and how many frames do all methods use at evaluation?

Answer: Thank you for asking us to clarify. In Tables 1 and 2, prior 4B and 8B rows are cited from Video-Zero [1] on the same backbones, with VisPlay and V-Zero kept for table parity; we updated Appendix A.4 to state this explicitly alongside our local evaluations. Under the same protocol used for 30B-A3B and Gemma-4-E4B-it, 2 fps up to 32 frames on nine benchmarks and 64 frames on LongVideoBench, MLVU, and Video-MME-Long, local and cited 4B and 8B Base scores align closely.

Base model alignmentTemporal GroundingVideo QA & ReasoningOverall
4B Base, cited25.0854.6844.81
4B Base, local25.1555.2345.20
8B Base, cited30.2857.7548.59
8B Base, local29.8957.6948.42

Other concern 2 & Q6: Some of the grounding gain may simply come from learning the [[start, end]] output format. Parse-failure rates before and after training would help. 6. What are the answer and interval parse-failure rates of the base models and Tempo-Zero on the four grounding benchmarks?

Answer: Thank you for this diagnostic suggestion. On Qwen3-VL-8B, we measured local parse-failure rates before and after training alongside scores on the jointly parseable subset; see Appendix E.3.

8B grounding benchmarkJointly parseableAnswer unparseableInterval unparseableFull score, local Base → TZJoint subset, Base → TZ
ANet-RTL201 / 229N/A11.8% → 1.3%20.56 → 31.2223.42 → 32.84
ActivityNet15,163 / 17,031N/A10.7% → 0.9%26.81 → 34.1630.04 → 35.59
Charades-STA3,358 / 3,720N/A9.6% → 1.2%36.48 → 46.1040.36 → 47.73
NExT-GQA5,546 / 5,5530.0% → 0.0%0.1% → 0.0%35.71 → 40.3935.76 → 40.43

Format compliance helps modestly on pure interval benchmarks as unparseable outputs fall near one percent, yet clear gains remain on all four jointly parseable subsets, confirming genuine localization progress.

Other concern 3: The Questioner is given the criteria and answers and mostly rewrites questions, so “constructing tasks” overstates its role.

Answer: Thank you for pointing this out. We agree and revised the Abstract and Section 1 to state that the Questioner proposes questions and distractors from verified criteria, matching Section 3.2. In our response to Reviewer fjNk Q3 and Appendix D.4, we also isolated Questioner role optimization and showed paired examples on the same video, illustrating that proposed questions genuinely differ from compiled canonical ones by combining multiple events and finer distractors on weak criteria.

Other concern 4 & Q8: Compute is reported only as “eight GPUs”. 8. What GPU type and wall-clock time does one training run take, including rendering?

Answer: Thank you for asking for these details, now in Appendix A.3. Each run uses 8 × NVIDIA H200 GPUs, plus 1.8 to 2.9 hours for 12-benchmark evaluation. To keep runtime low, we overlap next-step asynchronous CPU rendering across 20 workers with GPU training, filter invalid proposals on CPU before video inference, and reuse the 8 Solver rollouts from pass-rate filtering directly for policy updates. Gemma-4-E4B-it matches 8B generation speed, while its update step takes longer because variable-length sequence packing is not yet enabled for its integration in our backend.

8 × NVIDIA H200 timingQwen3-VL-4BQwen3-VL-8BQwen3-VL-30B-A3BGemma-4-E4B-it
Mean step time, incl. 9 to 10 s render wait98 s114 s242 s278 s
100-step training2.7 h3.2 h6.7 h7.7 h
Init, holdout, rule selection & save1.0 h1.0 h1.1 h1.3 h
Total training wall-clock3.7 h4.2 h7.8 h9.0 h

Other concern 5: At 4B, held-out accuracy drops after round 4 (Table 10), but the round-5 checkpoint is always used.

Answer: Thank you for noting this. We fixed the 5-round budget upfront across all runs to match Video-Zero [1] and evaluate the final checkpoint uniformly without post-hoc selection. Small shifts on the 128-instance holdout set are expected while evidence and grounded scores rise in Round 5, and our 10-round study in the response to Reviewer fjNk and Appendix D.2 shows continued localization and holdout gains.

We are happy to address any further questions.

References

[1] Ruixu Zhang et al. Video-Zero: Self-Evolution Video Understanding. arXiv:2605.14733, 2026.

Dear Reviewer 9Tux,

Thank you again for your thoughtful review and valuable suggestions. Since the discussion period is ending soon, we wanted to check in and see if our response has helped address your concerns. Please feel free to let us know if anything remains unclear.

We truly appreciate your time and feedback!

Comments on the paper

1 comment
Area_Chair_zr7NOpenReview ↗

Reviewers, please respond to the authors while the discussion is still active

Dear reviewers,

The authors have posted responses to your reviews. Please check them and reply to the authors while the discussion period is still open.

Thanks,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.