acceptodds
Under review as a conference paper at ICLR 2027

VideoSeeker: Empowering Cross-Video Understanding via Multi-Turn Evidence-Seeking Agentic Reasoning

Abstract

Cross-video understanding (CVU) requires models to locate key evidence across videos, establish temporal correspondences, and integrate intermediate findings through explicit comparison. Evidence needs become clearer as reasoning unfolds, yet end-to-end methods that answer directly from fixed frame samples lack mechanisms to supplement and verify evidence based on intermediate judgments. Under limited context budgets, sparse sampling may miss brief yet critical events, while dense encoding of entire videos scales poorly to multiple long videos. To this end, we formulate CVU as multi-turn agentic reasoning driven by active evidence seeking, integrating targeted segment replay, temporally interleaved comparison, and persistent evidence memory into a closed loop of observation, comparison, and reasoning. To learn reliable evidence-seeking policies, we first construct 4,000 multi-turn trajectories through verifiable trajectory distillation for supervised fine-tuning. During reinforcement learning, we assign fine-grained credit to evidence notes using information-gain rewards estimated from changes in the log-likelihood of the GT under the current policy. We further introduce a visual grounding gate that restricts positive credit to notes citing previously observed frames, mitigating reward hacking. Extensive experiments demonstrate the effectiveness of our multi-turn agentic reasoning framework for CVU.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
VLMs, Video Understanding, Agent

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

VideoSeeker introduces a multi-turn agentic framework for cross-video understanding. It uses three tools to replay relevant video segments, compare evidence across videos, and record findings in persistent memory. The authors construct 4,000 verified multi-turn trajectories across nine task types for supervised fine-tuning. They further introduce note-level information gain rewards based on changes in ground-truth answer likelihood. A visual grounding gate restricts positive rewards to notes that reference previously observed frames. Experiments on CrossVid, CVBench, and MVU-Eval show consistent improvements over the corresponding base models.

Critical strengths
  1. Effective cross-video reasoning framework. VideoSeeker combines targeted frame retrieval, explicit cross-video comparison, and persistent evidence memory.

  2. Careful training data construction. The authors construct 4,000 multi-turn trajectories using independent ground-truth generation, answer-leakage checks, and trajectory verification. These procedures improve the reliability of agentic training supervision.

  3. Fine-grained RL design. The proposed information gain reward assigns credit to individual evidence notes, while the visual grounding gate discourages unsupported notes.

Critical weaknesses
  1. Insufficient related works. The paper lacks direct comparisons with recent agentic video reasoning methods, such as Video-o3, Video-MTR, VideoSEAL, Open-o3-Video and VideoTemp-o3 (ICML 2026). This makes it difficult to establish the advantages of VideoSeeker over existing evidence-seeking and grounding approaches.

  2. Unclear necessity of joint cross-video reasoning. The paper lacks video concatenation and independent per-video reasoning baselines. Several tasks may be solved by analyzing each video separately and aggregating the results, leaving the benefit of explicit cross-video comparison unclear.

  3. Insufficient evaluation of long multi-video scalability. The paper motivates its approach through the high cost of processing multiple long videos, but lacks performance analysis across different video durations and video counts.

  4. Limited gains from the proposed RL rewards. The full reward design improves over outcome-only GRPO by only 0.5–1.0 percentage points across the three benchmarks. The paper lacks statistical significance tests, reward sensitivity analysis, and direct evidence that the grounding gate improves note correctness.

Questions for authors
  1. Training data reliability. What fraction of teacher-authored training samples have only partial teacher agreement? Can the authors provide a human annotation auditt?

  2. Role of persistent memory. How does performance change when correct evidence notes are removed, replaced, or reordered? Can the authors quantify whether persistent memory improves reasoning beyond simply retaining the full interaction history?

6 comments
AuthorsOpenReview ↗

Response Part1

Dear Reviewer AzCm,

We greatly appreciate the care and attention you devoted to reviewing our work. Your constructive comments prompted us to conduct additional comparisons and controlled evaluations, which help clarify the contributions of VideoSeeker. We address your concerns point by point below.

W1: On Comparisons with Recent Agentic Video Reasoning Methods

We agree that comparisons with recent agentic video reasoning methods are important for validating our approach. To directly address this concern, we conducted additional experiments with four representative methods, including Video-o3 [1], Video-MTR [2], Open-o3-Video [3], and VideoTemp-o3 [4], across three cross-video understanding benchmarks.

ModelCrossVIDCVBenchMVU-Eval
VideoTemp-o3 [4]28.9259.4654.96
Video-o3 [1]28.9857.3455.13
Video-MTR [2]26.3555.9451.71
Open-o3-Video [3]28.6957.9553.28
VideoSeeker-9B (Ours)50.2774.0466.83
VideoSeeker-35B-A3B (Ours)51.8675.3568.24

Table A: Comparison with recent agentic video reasoning methods (%).

As shown in Table A, VideoSeeker consistently outperforms all four methods across the three benchmarks. Notably, VideoSeeker-9B surpasses the strongest competing results by 21.29, 14.58, and 11.70 percentage points, respectively, achieving an overall average of 63.71%, compared with 47.78% for the strongest competing method.

We attribute these improvements to our dedicated training for active evidence seeking and cross-video evidence integration. Unlike methods primarily designed for single-video reasoning, VideoSeeker jointly learns targeted replay, explicit cross-video comparison, and persistent evidence recording. Furthermore, as shown in Table 1 of the manuscript, simply equipping Qwen3.5-9B with the same tools yields only a marginal improvement from 50.33% to 51.14%, whereas our agentic training achieves 63.71%. These results highlight the importance of learning effective tool-use strategies rather than merely providing tool access. We will incorporate these additional comparisons into the revised manuscript.

W2: On the Necessity of Joint Cross-Video Reasoning

We agree that distinguishing joint cross-video reasoning from independent per-video analysis is important for validating our approach. We first clarify that our Official Direct baseline already processes frames from all input videos jointly within a single context. To further address this concern, we conducted an additional experiment using Independent Per-video Reasoning + Aggregation, where each video is analyzed separately before aggregating the results.

SettingCVBench (%)MVU-Eval (%)CrossVid (%)
Official Direct (Base)62.5855.8032.60
Base + Tools (Zero-shot)62.2758.9932.17
Independent Per-video + Aggregation60.0652.3231.67
VideoSeeker (Ours)74.0466.8350.27

Table B: Comparison of joint-input reasoning, independent per-video reasoning, and VideoSeeker.

The independent baseline covers 9,006 of 9,015 CrossVid samples.

As shown in Table B, independent per-video reasoning underperforms joint-input direct reasoning across all three benchmarks. More importantly, VideoSeeker surpasses the independent baseline by 13.98, 14.51, and 18.60 percentage points, respectively.

Furthermore, our existing tool ablation (Table 4 of the manuscript) provides complementary evidence for explicit cross-video comparison. Adding the Compare tool alongside Choose Frames and Note improves performance by 0.80, 1.45, and 0.88 points across the three benchmarks, respectively. These results demonstrate that our improvements stem not merely from combining multiple videos, but from actively acquiring, comparing, and integrating cross-video evidence through multi-turn reasoning. We will incorporate these additional results and clarifications into the revised manuscript.

AuthorsOpenReview ↗

Response Part2

W3: On the Scalability of Long Multi-Video Understanding

We agree that evaluating long multi-video scenarios is important for validating the scalability of VideoSeeker. To directly address this concern, we conducted additional experiments across three benchmarks, adopting a strict criterion that every input video must be at least 7 minutes long. As shown in Table C, this yields 2,003 questions.

BenchmarkQuestions
CVBench17
MVU-Eval0
CrossVid1,986
Total2,003

Table C: Distribution of long multi-video questions (each video ≥7 minutes).

We further evaluated four settings, including a Frame-Matched Direct baseline, with results aggregated equally across all 2,003 questions.

MethodScore (%)
Official Direct32.48
Base + Tools31.62
Frame-Matched Direct32.47
VideoSeeker (Ours)47.33

Table D: Performance comparison on long multi-video inputs (each video ≥7 minutes).

As shown in Table D, VideoSeeker achieves 47.33%, outperforming Official Direct, Base + Tools, and Frame-Matched Direct by 14.85, 15.71, and 14.86 percentage points, respectively. These substantial improvements demonstrate the effectiveness of our active evidence-seeking framework on long multi-video inputs, where targeted replay, explicit comparison, and persistent evidence memory enable more effective evidence accumulation than fixed-frame observations.

While further evaluations across varying video counts and computational budgets would provide a more comprehensive scalability analysis, these additional results strengthen the empirical validation of VideoSeeker on long multi-video understanding tasks. We will incorporate these findings into the revised manuscript.

W4: On the Effectiveness and Reliability of the Proposed RL Rewards

To further validate the effectiveness of our reward design, we conducted additional experiments with three independent training seeds, comparing outcome-only GRPO, note-level credit assignment without the grounding gate, and our full method.

MethodCVBenchMVU-EvalCrossVid
GRPO72.84 ± 0.2665.62 ± 0.6749.41 ± 0.40
CA (Note) w/o Gate73.34 ± 0.7064.85 ± 0.7648.79 ± 0.59
VideoSeeker-9B (Ours)74.61 ± 0.5666.96 ± 0.4950.82 ± 0.49
Δ vs. GRPO+1.77+1.34+1.41

Table E: Performance across three independent training seeds (mean ± standard deviation, %).

As shown in Table E, our method achieves consistent improvements over GRPO across three benchmarks, with average gains of 1.77, 1.34, and 1.41 percentage points, respectively. Furthermore, removing the grounding gate reduces performance by 1.27, 2.11, and 2.03 points, demonstrating its contribution to effective note-level credit assignment.

Importantly, multi-turn video reasoning involves progressively acquiring and refining visual evidence, as explored in VideoSearcher [5] and FrameThinker [6]. Our grounding gate is therefore designed to constrain positive information-gain rewards through verifiable visual provenance, encouraging evidence-grounded note generation during iterative reasoning.

To further assess note reliability, we employed Kimi K3 to audit all 2,060 notes generated by VideoSeeker-9B across three benchmarks, evaluating their consistency with referenced visual evidence and contribution to final answers.

Audit CategoryCountPercentage
Fully supported by visual evidence62830.49%
Partially supported by visual evidence1,35065.53%
Contradicted by visual evidence823.98%
Notes supporting final answers1,71783.35%

Table F: Visual consistency and final-answer support analysis of 2,060 generated notes.

As shown in Table F, 96.02% of notes were judged to be fully or partially supported by visual evidence, with only 3.98% classified as contradicted. Moreover, 83.35% were judged to support final answers. Together with the multi-seed experiments and grounding-gate ablation, these findings provide complementary empirical evidence for the effectiveness and reliability of our grounded note-level reward design. We will incorporate these additional analyses into the revised manuscript.

AuthorsOpenReview ↗

Response Part3

Q1: On the Reliability of Training Data

We agree that ensuring training data reliability is essential. To directly address this concern, we analyzed teacher agreement across all 4,000 SFT training samples and conducted an additional human validation study. Specifically, our pipeline employs two teacher models, Kimi K3 and Kimi K2.6, for independent verification. Samples are classified as Full Consensus when both teachers pass verification and Half Consensus when only one succeeds.

GT SourceTotalHalf ConsensusRatio
Constructed1,226746.04%
Label-anchored1,3991339.51%
Teacher-authored1,375966.98%
Total4,0003037.58%

Table G: Teacher agreement statistics across different GT sources.

As shown in Table G, only 6.98% of teacher-authored samples exhibit partial agreement, while 92.43% of the entire training set achieves full consensus. To further assess annotation reliability, we conducted a human audit of all 303 Half Consensus samples, verifying their ground-truth answers against the corresponding video evidence.

Audited SamplesCorrectIncorrectAccuracy
303294997.03%

Table H: Human validation results on Half Consensus samples.

As shown in Table H, 294 out of 303 samples (97.03%) were confirmed correct through human validation, with only 9 problematic samples identified. These results demonstrate that partial teacher agreement does not necessarily indicate incorrect annotations and provide additional evidence for the reliability of our multi-stage verification pipeline. We will incorporate these statistics and human validation results into the revised manuscript.

Q2: On the Role of Persistent Evidence Memory

We agree that evaluating whether persistent evidence memory provides benefits beyond retaining the full interaction history is important. To directly address this concern, we conducted an additional intervention study on all 872 questions from CrossVid Free-form QA (FF/CCQA), which require integrating evidence across multiple videos.

Specifically, we evaluated four settings by retaining, removing, randomly replacing, or shuffling the evidence notes, while keeping the remaining interaction history unchanged. The results are summarized in Table I.

SettingOriginal NotesRemovedReplacedShuffled
Accuracy (%)33.6427.8728.2128.81

Table I: Performance under different note interventions on CrossVid FF (872 questions).

As shown in Table I, removing evidence notes reduces performance from 33.64% to 27.87%, a drop of 5.77 percentage points. Randomly replacing or shuffling notes also decreases performance by 5.43 and 4.83 points, respectively.

Importantly, these performance drops occur even when the remaining interaction history is preserved, demonstrating that evidence notes provide additional benefits beyond simply retaining previous observations. These results highlight the importance of structured evidence memory in preserving and integrating relevant information throughout multi-turn cross-video reasoning. We will incorporate these additional findings into the revised manuscript.

References

[1] Zeng X, Zhang Z, Zhu Y, et al. Video-o3: Native Interleaved Clue Seeking for Long Video Multi-Hop Reasoning[C]//Proceedings of the International Conference on Machine Learning. 2026.

[2] Xie Y, Chen T, Ge Z, et al. Video-MTR: Reinforced Multi-Turn Reasoning for Long Video Understanding[C]//Proceedings of the International Conference on Machine Learning. 2026.

[3] Meng J, Li X, Wang H, et al. Open-o3-Video: Grounded Video Reasoning with Explicit Spatio-Temporal Evidence[C]//Proceedings of the International Conference on Machine Learning. 2026.

[4] Liu W, Wang Y, Ma S, et al. VideoTemp-o3: Harmonizing Temporal Grounding and Video Understanding in Agentic Thinking-with-Videos[C]//Proceedings of the International Conference on Machine Learning. 2026.

[5] Gao Z, Bao Y, Peng J, et al. VideoSearcher: Empowering Video Deep Research with Multi-Tool Agentic Reasoning via Reinforcement Learning[C]//Proceedings of EMNLP. 2026.

[6] He Z, Qu X, Li Y, et al. FrameThinker: Learning to Think with Long Videos via Multi-Turn Frame Spotlighting[C]//International Conference on Learning Representations. 2026.

We deeply value your detailed feedback and the opportunity to substantiate our methodological choices with further analyses. These suggestions have been instrumental in improving the clarity and rigor of our revised manuscript.

Best regards, The Authors of VideoSeeker

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

This paper presents VideoSeeker, a multi-turn agentic framework for cross-video understanding that combines targeted frame retrieval, cross-video comparison, and persistent evidence notes. It constructs 4,000 verified trajectories for supervised fine-tuning and introduces note-level information-gain rewards with a visual grounding gate during reinforcement learning. Experiments on Qwen3.5-9B and Qwen3.5-35B-A3B demonstrate substantial improvements across three cross-video understanding benchmarks.

Critical strengths
  1. This paper investigates cross-video understanding, a challenging task that requires integrating evidence across multiple videos. It explores a multi-turn framework combining targeted frame retrieval, cross-video comparison, and persistent evidence notes.

  2. The training data construction is well designed. The pipeline separates ground-truth generation from trajectory exploration and incorporates multiple verification steps, producing 4,000 multi-turn trajectories across nine task types.

  3. The experiments demonstrate substantial improvements over the base models, with average gains of 13.38 and 10.19 percentage points for the 9B and 35B-A3B models, respectively. The paper also includes tool and training ablations to evaluate the effectiveness of different components.

Critical weaknesses
  1. The effectiveness of the proposed note-level credit assignment is not sufficiently demonstrated. While the overall framework achieves substantial improvements, Table 3 shows only modest gains over outcome-only GRPO (0.5–1.0 percentage points). Without repeated experiments or further analysis of how note-level rewards affect evidence-seeking behavior, it remains unclear whether this mechanism provides consistent and meaningful benefits beyond standard RL training.

  2. The visual grounding mechanism is not adequately validated. The proposed gate checks whether cited frames have been observed, but does not verify whether the note content is actually supported by those frames. Therefore, unsupported conclusions may still receive positive rewards. Although the ablation demonstrates improvements in final accuracy, it does not directly establish that the gate improves evidence faithfulness or mitigates the reward-hacking behavior motivating its design.

Questions for authors
  1. Table 3 shows relatively small improvements from note-level credit assignment over outcome-only GRPO. Are these gains consistent across multiple training runs? Can the authors provide further analysis showing how the proposed rewards improve evidence-seeking behavior or note quality?

  2. The visual grounding gate only checks whether cited frames have been observed. How does this prevent notes containing unsupported conclusions from receiving positive rewards? Can the authors directly evaluate whether the gate improves evidence faithfulness or reduces reward hacking, beyond final answer accuracy?

  3. How much of the overall improvement comes from adaptive evidence seeking rather than task-specific training? Can the authors compare against a fixed-frame sampling baseline under matched training and inference budgets, and report the actual visual token usage and end-to-end inference latency?

5 comments
AuthorsOpenReview ↗

Response Part1

Dear Reviewer krvx,

Thank you for your thoughtful and constructive review of our work. We greatly appreciate your recognition of our verifiable trajectory construction pipeline and the overall performance improvements of VideoSeeker. Your valuable feedback has motivated us to conduct additional experiments and analyses on note-level credit assignment, visual grounding reliability, and adaptive evidence seeking under comparable inference budgets. Below, we provide detailed point-by-point responses.

W1 & Q1: On the Effectiveness and Statistical Reliability of Note-Level Credit Assignment

We appreciate the reviewer's emphasis on the empirical validation of note-level credit assignment. We agree that its effectiveness deserves stronger experimental evidence. To directly address this concern, we conducted additional experiments with three independent training seeds, comparing outcome-only GRPO, note-level credit assignment without the grounding gate, and our full method. The results are summarized in Table A.

MethodCVBenchMVU-EvalCrossVid
GRPO72.84 ± 0.2665.62 ± 0.6749.41 ± 0.40
CA (Note) w/o Gate73.34 ± 0.7064.85 ± 0.7648.79 ± 0.59
VideoSeeker-9B (Ours)74.61 ± 0.5666.96 ± 0.4950.82 ± 0.49
Δ vs. GRPO+1.77+1.34+1.41

Table A: Performance across three independent training seeds (mean ± standard deviation, %).

As shown in Table A, our method consistently achieves higher mean accuracy than standard GRPO across all three benchmarks, with improvements of 1.77, 1.34, and 1.41 percentage points, respectively. These results provide additional evidence for the reliability of our reward design across independent training runs.

Furthermore, removing the visual grounding gate reduces performance by 1.27, 2.11, and 2.03 points, respectively, highlighting the importance of grounding constraints for effective note-level credit assignment.

We acknowledge that SFT and outcome-based RL contribute substantially to the overall improvements. Nevertheless, our grounded note-level rewards provide complementary fine-grained supervision for intermediate evidence accumulation, beyond final-answer correctness alone. These additional experiments further support the effectiveness of our reward design. We will incorporate the multi-seed evaluation into the revised manuscript.

W2 & Q2: On the Effectiveness and Faithfulness of the Visual Grounding Gate

We appreciate the reviewer's careful examination of the visual grounding mechanism. We would like to clarify that intermediate notes in multi-turn agentic reasoning may represent evolving observations that are progressively refined through subsequent interactions, as explored in VideoSearcher [1] and FrameThinker [2]. Our grounding gate therefore aims to constrain positive information-gain credit through verifiable visual provenance, ensuring that rewarded notes reference previously observed frames.

To further evaluate evidence faithfulness, we employed Kimi K3 to audit all 2,060 notes generated by VideoSeeker-9B across CVBench, MVU-Eval, and CrossVid, examining their consistency with referenced visual evidence and contributions to final answers.

Audit CategoryCountPercentage
Fully supported by visual evidence62830.49%
Partially supported by visual evidence1,35065.53%
Contradicted by visual evidence823.98%
Notes supporting final answers1,71783.35%

Table B: Note–image consistency and final-answer support analysis of 2,060 generated notes.

As shown in Table B, 96.02% of notes were judged to be fully or partially supported by visual evidence, while only 3.98% directly contradicted the observed frames. Moreover, 83.35% were judged to provide useful support for final answers. These findings indicate strong overall consistency between generated notes and their referenced visual evidence.

Furthermore, the ablation results in Table A show that removing the grounding gate reduces accuracy by 1.27, 2.11, and 2.03 points across three benchmarks, respectively. Together, these results provide complementary empirical support for the practical effectiveness of our grounding-constrained reward design in promoting reliable evidence accumulation.

We will incorporate these additional analyses and further clarify the role of visual provenance constraints in the revised manuscript.

AuthorsOpenReview ↗

Response Part2

Q3: On Adaptive Evidence Seeking and Computational Efficiency

We are grateful for the reviewer's suggestion to further investigate the contribution and computational cost of adaptive evidence seeking. We agree that distinguishing its benefits from task-specific training is important. To address this concern, we conducted additional experiments using the same Qwen3.5-9B backbone.

1. Contribution of Adaptive Evidence Seeking.

We additionally trained two direct-answer baselines, Direct SFT and Direct GRPO, using task-specific answer supervision without evidence-seeking tools or multi-turn interactions.

ModelCVBenchMVU-EvalCrossVidAvg.
Qwen3.5-9B62.5855.8032.6050.33
Direct SFT68.6163.0845.9559.21
Direct GRPO69.4263.3646.1359.64
VideoSeeker (Ours)74.0466.8350.2763.71

Table C: Comparison with task-specific direct-answer training baselines (accuracy, %).

As shown in Table C, VideoSeeker outperforms Direct GRPO by 4.62, 3.47, and 4.14 points across the three benchmarks, respectively. These improvements indicate that our gains cannot be attributed solely to task-specific answer supervision, supporting the additional benefits of learned multi-turn evidence seeking.

2. Computational Efficiency under Comparable Frame Budgets.

We further compared official direct inference, tool-enabled inference without agentic training (Base + Tools), uniformly sampled frames matched to VideoSeeker's observation budget (Frame-Matched Direct), and our method.

BenchmarkSettingAccuracy (%)Avg. FramesLatency (s)
CVBenchOfficial Direct62.5822.3216.88
Base + Tools62.2722.3220.78
Frame-Matched Direct68.91112.4943.04
VideoSeeker (Ours)74.04117.2652.05
MVU-EvalOfficial Direct55.80151.6314.30
Base + Tools58.99154.3427.29
Frame-Matched Direct59.1052.9817.29
VideoSeeker (Ours)66.8351.9127.89
CrossVidOfficial Direct32.60127.6211.94
Base + Tools32.17138.5843.58
Frame-Matched Direct32.29132.0012.30
VideoSeeker (Ours)50.27132.9565.72

Table D: Comparison of accuracy, observed frames, and end-to-end inference latency.

As shown in Table D, VideoSeeker outperforms Frame-Matched Direct by 5.13, 7.73, and 17.98 points, respectively, while observing comparable numbers of frames. Notably, on MVU-Eval, our method achieves higher accuracy with substantially fewer frames than Official Direct (51.91 vs. 151.63). Although multi-turn interactions introduce additional latency, these results highlight the benefits of adaptive evidence localization and cross-video comparison, rather than simply increasing frame consumption.

We will incorporate these additional experiments and the accuracy–latency analysis into the revised manuscript.

References

[1] Gao Z, Bao Y, Peng J, et al. VideoSearcher: Empowering Video Deep Research with Multi-Tool Agentic Reasoning via Reinforcement Learning. Proceedings of EMNLP, 2026.

[2] He Z, Qu X, Li Y, et al. FrameThinker: Learning to Think with Long Videos via Multi-Turn Frame Spotlighting. Proceedings of ICLR, 2026.

We would like to thank the reviewer once again for the thoughtful feedback and constructive suggestions. We believe that the additional experiments and analyses provide stronger empirical support for our method and help address the concerns raised. We greatly appreciate the opportunity to clarify and strengthen our work.

Best regards,

The Authors of VideoSeeker

Dear Reviewer krvx,

Thank you again for your time and thoughtful review of our paper. Your comments and suggestions have been very helpful in improving our work. We hope our responses have addressed your concerns. If you have any further questions or would like us to clarify anything, please feel free to let us know. We would be happy to provide more details.

Best regards, The Authors of VideoSeeker

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

This paper proposes VideoSeeker, a multi-turn agent for cross-video understanding. It uses frame selection, cross-video comparison, and persistent evidence notes to gather information during reasoning. The model is trained with verified trajectories followed by reinforcement learning with note-level information-gain rewards and a visual grounding gate. Experiments show substantial improvements across several video understanding benchmarks.

Critical strengths

The paper addresses an important problem in cross-video understanding. Active evidence seeking is well motivated for tasks requiring information from multiple videos.

The framework is clearly presented, and the training pipeline provides structured supervision for multi-turn reasoning.

The experiments cover multiple benchmarks, model sizes, and ablations. The improvements over the base models are substantial, particularly on temporal reasoning tasks.

Critical weaknesses

The visual grounding mechanism is insufficient. The grounding gate only checks whether a note references previously observed frames. It does not verify whether the note's content is actually supported by those frames. A note containing an incorrect or unsupported claim could still receive positive information-gain credit if it cites valid frame identifiers. This weakens the claim that the method prevents reward hacking.

The contribution of the proposed reward mechanism is relatively small. In Table 3, the full method improves over outcome-only GRPO by only 0.51, 1.00, and 0.50 points across the three benchmarks. These gains are much smaller than the overall improvements from agentic training. Results across multiple training seeds would help establish the reliability of the proposed credit assignment.

Computational efficiency is not evaluated. Active evidence seeking introduces multiple model calls, repeated visual encoding, and additional reasoning tokens. The paper does not report inference latency, token consumption, or the number of observed frames at evaluation time. Comparisons under matched computational or visual-token budgets are needed to establish the practical benefits over direct video understanding.

The evaluation protocol limits comparison with prior work. The main cross-video experiments compare against direct-answer models and zero-shot agents, but do not include trained video agents adapted to the same task. In addition, the authors modify benchmark annotations, correcting 45 CVBench and 155 MVU-Eval answers. Reporting results on the original benchmark versions would improve comparability and reproducibility.

Questions for authors

How does the grounding gate prevent a note from making unsupported claims while citing valid frame identifiers?

Are the improvements from note-level credit assignment consistent across different training seeds?

How does VideoSeeker compare with direct-answer models under matched inference budgets, including visual tokens and latency?

Could the authors include comparisons with strong proprietary multimodal models, such as GPT and Claude, under both direct-answer and agentic settings? Evaluating these models with comparable visual inputs and inference budgets would help better assess VideoSeeker's performance.

6 comments
AuthorsOpenReview ↗

Response Part1

Dear Reviewer kGZC,

We sincerely thank you for your thoughtful and constructive review of our work. We greatly appreciate your valuable comments regarding visual grounding, note-level credit assignment, computational efficiency, and evaluation comparability. To address these concerns, we conducted additional experiments and analyses to further strengthen our work. Below, we provide detailed point-by-point responses.

W1 & Q1: On the Reliability and Effectiveness of the Visual Grounding Gate

We agree that our grounding gate verifies whether cited frames have been observed, rather than directly assessing the semantic correctness of each note. However, intermediate notes in multi-turn agentic reasoning may represent evolving observations that are progressively refined through subsequent interactions, as explored in VideoSearcher [1] and FrameThinker [2]. Our gate therefore aims to constrain positive information-gain credit through verifiable visual provenance, rather than guarantee the correctness of every intermediate note.

To further validate its reliability, we employed Kimi K3 to audit all 2,060 notes generated by VideoSeeker-9B across CVBench, MVU-Eval, and CrossVid, assessing their consistency with referenced visual evidence and contribution to final answers.

Audit CategoryCountPercentage
Fully supported by visual evidence62830.49%
Partially supported by visual evidence1,35065.53%
Contradicted by visual evidence823.98%
Notes supporting final answers1,71783.35%

Table A: Visual consistency and final-answer support analysis of 2,060 generated notes.

As shown in Table A, 96.02% of notes were judged to be at least partially supported by visual evidence, with only 3.98% contradicting observed frames. Moreover, 83.35% were judged to provide useful support for final answers. These results suggest that the agent generally records visually relevant evidence rather than arbitrary unsupported claims.

Furthermore, the ablation results in Table 3 show that removing the grounding gate consistently reduces performance across all three benchmarks, supporting its practical effectiveness.

Overall, while the gate does not guarantee semantic correctness, it provides a verifiable visual provenance constraint that helps mitigate potential reward hacking. We will incorporate these analyses and clarify this distinction in the revised manuscript.

W2 & Q2: On the Effectiveness and Robustness of Note-Level Credit Assignment

We agree that SFT and outcome-based RL contribute substantially to the overall improvements, and that the reliability of note-level credit assignment requires further validation. To address this concern, we conducted additional experiments with three independent training seeds, comparing outcome-only GRPO, note-level credit assignment without the grounding gate, and our full method.

MethodCVBenchMVU-EvalCrossVid
GRPO72.84 ± 0.2665.62 ± 0.6749.41 ± 0.40
CA (Note) w/o Gate73.34 ± 0.7064.85 ± 0.7648.79 ± 0.59
VideoSeeker-9B (Ours)74.61 ± 0.5666.96 ± 0.4950.82 ± 0.49
Δ vs. GRPO+1.77+1.34+1.41

Table B: Performance across three independent training seeds (mean ± standard deviation, %).

As shown in Table B, our method consistently achieves higher mean accuracy than GRPO across all three benchmarks, with improvements of 1.77, 1.34, and 1.41 points, respectively. Moreover, removing the grounding gate reduces performance by 1.27, 2.11, and 2.03 points, highlighting its importance for effective note-level credit assignment.

While outcome-based RL provides substantial gains, our grounded note-level rewards offer complementary supervision for intermediate evidence accumulation beyond final-answer correctness. These multi-seed results further support the reliability and effectiveness of our reward design. We will incorporate them into the revised manuscript.

References

[1] Gao Z, Bao Y, Peng J, et al. VideoSearcher: Empowering Video Deep Research with Multi-Tool Agentic Reasoning via Reinforcement Learning[C]//Proceedings of EMNLP. 2026.

[2] He Z, Qu X, Li Y, et al. FrameThinker: Learning to Think with Long Videos via Multi-Turn Frame Spotlighting[C]//Proceedings of ICLR. 2026.

AuthorsOpenReview ↗

Response Part2

W3 & Q3: On Computational Efficiency and Matched Inference Budgets

We agree that computational efficiency is important for evaluating the practical benefits of active evidence seeking. To address this concern, we conducted additional experiments across three benchmarks, measuring average observed frames and end-to-end inference latency.

For a fair comparison, we evaluate four settings using the same Qwen3.5-9B backbone: (1) Official Direct, following the official benchmark protocols; (2) Base + Tools, using our tool environment without agentic training; (3) Frame-Matched Direct, using uniformly sampled frames with budgets matched to VideoSeeker; and (4) VideoSeeker (Ours), employing learned multi-turn evidence seeking.

BenchmarkSettingAccuracy (%)Avg. FramesLatency (s)
CVBenchOfficial Direct62.5822.3216.88
Base + Tools62.2722.3220.78
Frame-Matched Direct68.91112.4943.04
VideoSeeker (Ours)74.04117.2652.05
MVU-EvalOfficial Direct55.80151.6314.30
Base + Tools58.99154.3427.29
Frame-Matched Direct59.1052.9817.29
VideoSeeker (Ours)66.8351.9127.89
CrossVidOfficial Direct32.60127.6211.94
Base + Tools32.17138.5843.58
Frame-Matched Direct32.29132.0012.30
VideoSeeker (Ours)50.27132.9565.72

Table C: Accuracy, average observed frames, and inference latency across different evaluation settings.

As shown in Table C, VideoSeeker outperforms the frame-matched direct baseline by 5.13, 7.73, and 17.98 points on CVBench, MVU-Eval, and CrossVid, respectively, while observing comparable numbers of frames. Notably, on MVU-Eval, VideoSeeker achieves 66.83% accuracy with only 51.91 frames, compared to 55.80% with 151.63 frames under the official protocol. Moreover, simply providing tools to the base model yields limited gains, highlighting the importance of learned evidence-seeking policies rather than tool access alone.

We acknowledge that multi-turn interactions introduce additional inference latency. Nevertheless, these results demonstrate substantial accuracy improvements under comparable frame budgets, while revealing the associated latency trade-off. We will incorporate this analysis into the revised manuscript.

W4: On Comparisons with Trained Video Agents and Benchmark Annotation Corrections

To address the reviewer's concerns regarding evaluation fairness and comparability, we conducted additional experiments comparing VideoSeeker with recent trained video reasoning methods and evaluating its performance on the original, unmodified benchmarks.

First, regarding trained video reasoning baselines, we evaluated four recent methods, including Video-o3 [3], Video-MTR [4], Open-o3-Video [5], and VideoTemp-o3 [6].

ModelCrossVidCVBenchMVU-Eval
VideoTemp-o3 [6]28.9259.4654.96
Video-o3 [3]28.9857.3455.13
Video-MTR [4]26.3555.9451.71
Open-o3-Video [5]28.6957.9553.28
VideoSeeker-9B (Ours)50.2774.0466.83
VideoSeeker-35B-A3B (Ours)51.8675.3568.24

Table D: Comparison with recent video reasoning methods (%).

As shown in Table D, VideoSeeker consistently outperforms all four methods, with VideoSeeker-9B exceeding the strongest competing results by 21.29, 14.58, and 11.70 points on CrossVid, CVBench, and MVU-Eval, respectively. These results further demonstrate the effectiveness of our multi-turn evidence-seeking framework for cross-video understanding.

Second, regarding benchmark annotation corrections, we additionally evaluated our method on the original, unmodified benchmarks.

ModelCVBench OriginalCleanedMVU-Eval OriginalCleaned
Qwen3.5-9B60.0062.5848.5755.80
Qwen3.5-9B (Agent)59.0062.2752.6358.99
VideoSeeker-9B72.9074.0460.2566.83

Table E: Performance on original and cleaned benchmarks (%).

As shown in Table E, VideoSeeker maintains substantial gains on the original benchmarks, outperforming Qwen3.5-9B by 12.90 and 11.68 points, compared with 11.46 and 11.03 points on the cleaned versions. This confirms that our conclusions do not depend on annotation corrections. As detailed in Appendix B.3, the 45 CVBench and 155 MVU-Eval corrections were verified against visual evidence. We will release the corrected annotations, representative cases, and evaluation scripts to facilitate reproducibility.

Together, these additional experiments further support the effectiveness of VideoSeeker and the robustness of our evaluation conclusions.

AuthorsOpenReview ↗

Response Part3

Q4: On Comparisons with Proprietary Multimodal Models

We agree that comparisons with proprietary multimodal models would strengthen our evaluation. However, full-benchmark evaluation incurs substantial API costs, particularly for multi-turn multimodal interactions. Given our limited budget, we randomly sampled 101 questions from each benchmark (303 in total) and evaluated Gemini 3.8 Flash under both direct-answer and agentic settings.

For a fair comparison, all models were evaluated on the same sampled questions, with the agentic setting using the same tool interface and tool-call limits as VideoSeeker.

ModelSettingCrossVidCVBenchMVU-Eval
VideoSeeker (Ours)Agentic50.5275.2567.33
Gemini 3.8 FlashDirect54.1469.3170.30
Gemini 3.8 FlashAgentic54.4572.2872.28

Table F: Comparison with Gemini 3.8 Flash on 101 randomly sampled questions per benchmark (%).

As shown in Table F, VideoSeeker achieves competitive performance against Gemini 3.8 Flash. Notably, on CVBench, our model achieves 75.25%, outperforming Gemini's direct and agentic settings by 5.94 and 2.97 points, respectively. Although Gemini performs better on CrossVid and MVU-Eval, these results demonstrate the competitiveness of our relatively small open-weight model against a strong proprietary model.

We acknowledge that broader comparisons with GPT and Claude would further strengthen our evaluation. Nevertheless, these additional results provide preliminary evidence of VideoSeeker's competitiveness. We will incorporate them into the revised manuscript.

References

[3] Zeng X, Zhang Z, Zhu Y, et al. Video-o3: Native Interleaved Clue Seeking for Long Video Multi-Hop Reasoning. Proceedings of ICML, 2026.

[4] Xie Y, Chen T, Ge Z, et al. Video-MTR: Reinforced Multi-Turn Reasoning for Long Video Understanding. Proceedings of ICML, 2026.

[5] Meng J, Li X, Wang H, et al. Open-o3-Video: Grounded Video Reasoning with Explicit Spatio-Temporal Evidence. Proceedings of ICML, 2026.

[6] Liu W, Wang Y, Ma S, et al. VideoTemp-o3: Harmonizing Temporal Grounding and Video Understanding in Agentic Thinking-with-Videos. Proceedings of ICML, 2026.

We sincerely thank the reviewer again for the thoughtful feedback and constructive suggestions. We believe that the additional experiments and analyses provide stronger empirical support for our method and help address the concerns raised. We greatly appreciate the opportunity to clarify and strengthen our work.

Best regards,

The Authors of VideoSeeker

Reviewer_rzk9OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes VideoSeeker, a multi-turn agent for cross-video understanding. The model can selectively replay video segments, compare frames across different videos, and save intermediate findings as evidence notes.

The authors construct 4K verified trajectories for SFT, followed by RL with note-level information gain rewards and a visual grounding gate. Experiments on CrossVid, CVBench and MVU-Eval show improvements over the base Qwen3.5 models, with particularly large gains on temporal understanding tasks.

Critical strengths
  1. The problem is interesting. Cross-video reasoning requires finding and comparing evidence across different videos, which is not always handled well by fixed frame sampling. The combination of targeted replay, comparison and evidence notes is a reasonable design.

  2. The overall results are quite strong. VideoSeeker-9B improves the average score from 50.33 to 63.71 over the base model in Table 1. I also appreciate the tool ablations in Table 4, which show that combining the three tools works better than using them separately.

Critical weaknesses
  1. I'm less convinced by the note-level RL contribution. In Table 3, the improvement over standard GRPO is only 0.51, 1.00 and 0.50 points on the three benchmarks. Most of the gain seems to come from SFT and outcome-based RL. Since note-level credit assignment is a main contribution, I would expect stronger evidence for its effectiveness.

  2. The visual grounding gate seems a bit weak. It only checks whether the cited frames have been observed, not whether those frames actually support the note. A model could still write an incorrect note with valid frame references and receive positive information-gain reward. Also, I couldn't find a comparison of inference latency or total visual tokens processed. This matters because the method uses multiple rounds of frame retrieval and comparison.

Questions for authors

The gain over standard GRPO in Table 3 is rather small. Are these results averaged over multiple runs? Is the improvement statistically significant?

Does the grounding gate verify the actual content of a note, or only its frame references? Have the authors measured how often notes contain unsupported claims despite passing the gate?

Could the authors report average tool calls, visual tokens and inference latency? It would be useful to compare VideoSeeker with a direct model given a similar visual-token budget.

Appendix B reports quite a few corrections to CVBench and MVU-Eval labels. Would the main conclusions change on the original, unmodified test sets?

3 comments
AuthorsOpenReview ↗

Response Part1

Dear Reviewer rzk9,

We sincerely thank you for your thoughtful review and constructive suggestions. We greatly appreciate your recognition of our work and the opportunity to further strengthen its empirical validation. To address your concerns, we conducted additional experiments and analyses on note-level credit assignment, visual grounding reliability, computational efficiency, and benchmark evaluation fairness. We provide our point-by-point responses below.

W1 & Q1: On the Effectiveness and Reliability of Note-Level Credit Assignment

To further validate the effectiveness of note-level credit assignment, we conducted additional experiments with three independent training seeds, comparing outcome-only GRPO, note-level credit assignment without the grounding gate, and our full method.

MethodCVBenchMVU-EvalCrossVid
GRPO72.84 ± 0.2665.62 ± 0.6749.41 ± 0.40
CA (Note) w/o Gate73.34 ± 0.7064.85 ± 0.7648.79 ± 0.59
VideoSeeker-9B (Ours)74.61 ± 0.5666.96 ± 0.4950.82 ± 0.49
Δ vs. GRPO+1.77+1.34+1.41

Table A: Performance across three independent training seeds (mean ± standard deviation, %).

As shown in Table A, our method consistently outperforms GRPO by 1.77, 1.34, and 1.41 points across three benchmarks. Notably, even its lowest run exceeds the highest GRPO run on each benchmark. These results indicate consistent improvements. Furthermore, removing the grounding gate reduces accuracy by 1.27, 2.11, and 2.03 points, supporting its effectiveness. While SFT and outcome-based RL contribute substantially to overall performance, grounded note-level rewards provide complementary supervision for intermediate evidence accumulation. We will incorporate these results into the revised manuscript.

W2 & Q2: On the Reliability and Effectiveness of the Visual Grounding Gate

We agree that our grounding gate verifies visual provenance rather than the semantic correctness of each note. However, multi-turn video reasoning involves progressively acquiring and refining visual evidence, as explored in VideoSearcher [1] and FrameThinker [2]. Our gate therefore aims to constrain positive information-gain credit through verifiable visual provenance, rather than guarantee the correctness of every intermediate note.

To directly assess note reliability, we employed Kimi K3 to audit all 2,060 notes generated by VideoSeeker-9B across three benchmarks, evaluating their consistency with referenced visual evidence and contribution to final answers.

Audit CategoryCountPercentage
Fully supported by visual evidence62830.49%
Partially supported by visual evidence1,35065.53%
Contradicted by visual evidence823.98%
Notes supporting final answers1,71783.35%

Table B: Visual consistency and final-answer support analysis of 2,060 generated notes.

As shown in Table B, 96.02% of notes were judged to be at least partially supported by visual evidence, with only 3.98% classified as contradicted. Moreover, 83.35% were judged to support final answers. Together with the grounding-gate ablation in Table 3, these findings support the practical effectiveness of our lightweight provenance constraint, while acknowledging that unsupported claims may still pass the gate. We will clarify this distinction in the revised manuscript.

AuthorsOpenReview ↗

Response Part2

W2 & Q3: On Computational Efficiency and Matched Inference Budgets

To evaluate computational efficiency, we conducted additional experiments comparing Official Direct, Base + Tools, Frame-Matched Direct, and VideoSeeker, using the same Qwen3.5-9B backbone.

BenchmarkSettingAccuracy (%)Avg. FramesLatency (s)
CVBenchOfficial Direct62.5822.3216.88
Base + Tools62.2722.3220.78
Frame-Matched Direct68.91112.4943.04
VideoSeeker (Ours)74.04117.2652.05
MVU-EvalOfficial Direct55.80151.6314.30
Base + Tools58.99154.3427.29
Frame-Matched Direct59.1052.9817.29
VideoSeeker (Ours)66.8351.9127.89
CrossVidOfficial Direct32.60127.6211.94
Base + Tools32.17138.5843.58
Frame-Matched Direct32.29132.0012.30
VideoSeeker (Ours)50.27132.9565.72

Table C: Accuracy, average observed frames, and end-to-end inference latency across different evaluation settings.

As shown in Table C, VideoSeeker outperforms Frame-Matched Direct by 5.13, 7.73, and 17.98 points across three benchmarks under comparable frame budgets. Furthermore, Figure 4 in the revised manuscript provides detailed tool invocation statistics, including the distribution and total number of Choose Frames, Compare, and Note calls, revealing a shift toward more frequent targeted evidence acquisition after agentic training.

We acknowledge that multi-turn interactions introduce additional latency and that observed-frame counts do not fully capture cumulative visual-token usage. Nevertheless, these results demonstrate substantial accuracy improvements under comparable frame budgets while revealing the associated computational trade-offs. We will incorporate this analysis into the revised manuscript.

Q4: On Benchmark Corrections and Evaluation Reliability

Finally, to verify that our conclusions do not depend on annotation corrections, we conducted additional evaluations on the original, unmodified benchmarks, with results summarized in Table D.

ModelCVBench OriginalCleanedMVU-Eval OriginalCleaned
Qwen3.5-9B60.0062.5848.5755.80
Qwen3.5-9B (Agent)59.0062.2752.6358.99
VideoSeeker-9B72.9074.0460.2566.83

Table D: Performance on original and cleaned benchmarks (%).

As shown in Table D, VideoSeeker maintains substantial gains of +12.90 and +11.68 points over Qwen3.5-9B on the original CVBench and MVU-Eval, respectively, slightly exceeding the improvements on the cleaned versions (+11.46 and +11.03). This confirms that our main conclusions remain unchanged without annotation corrections.

We further verified the benchmark modifications, with detailed statistics presented in Table E.

Modification TypeCVBench (1,000)MVU-Eval (1,824)
Answer corrected45 (4.50%)155 (8.50%)
Samples removed6 (0.60%)39 (2.14%)
Format repaired1 (0.10%)4 (0.22%)
Total affected52 (5.20%)198 (10.86%)

Table E: Statistics of benchmark modifications.

As detailed in Appendix B.3, these corrections underwent multiple rounds of verification against visual evidence, including dense-frame inspection when necessary. We will publicly release all revised benchmark annotations and evaluation data to ensure transparency and reproducibility. These results demonstrate that our improvements are robust to benchmark annotation changes.

References

[1] Gao Z, Bao Y, Peng J, et al. VideoSearcher: Empowering Video Deep Research with Multi-Tool Agentic Reasoning via Reinforcement Learning. EMNLP, 2026.

[2] He Z, Qu X, Li Y, et al. FrameThinker: Learning to Think with Long Videos via Multi-Turn Frame Spotlighting. ICLR, 2026.

We sincerely thank the reviewer again for the valuable feedback, which has helped us further strengthen the rigor and clarity of our work. We hope these additional experiments and clarifications adequately address the concerns and provide stronger evidence for the effectiveness and reliability of VideoSeeker.

Best regards,

The Authors of VideoSeeker

Dear Reviewer rzk9,

Thank you again for your time and thoughtful review of our paper. Your comments and suggestions have been very helpful in improving our work. We hope our responses have addressed your concerns. If you have any further questions or would like us to clarify anything, please feel free to let us know. We would be happy to provide more details.

Best regards, The Authors of VideoSeeker

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

Dear Area Chair and Reviewers,

We extend our deepest gratitude to all reviewers for their thoughtful feedback and careful evaluation. We are encouraged that reviewers recognize the importance of cross-video understanding, our multi-turn evidence-seeking framework, and the substantial performance gains of VideoSeeker. We have carefully addressed their feedback through extensive new experiments and clarifications.

Contributions:

We are pleased that the reviewers highlighted the following strengths:

  • Problem Motivation & Framework. Reviewers acknowledge the importance and challenges of cross-video understanding [rzk9, kGZC, krvx, AzCm]. The integration of targeted replay, cross-video comparison, and persistent evidence memory is recognized as a well-motivated design [rzk9, kGZC, AzCm].

  • Training Data Quality. Our verifiable trajectory pipeline is appreciated for separating ground-truth generation from agent exploration and incorporating multiple verification steps [krvx, AzCm]. The 4,000 trajectories across nine task types provide structured supervision [kGZC, krvx].

  • Methodology & Performance. The information-gain reward and visual grounding mechanism offer an interesting approach to evidence-grounded reasoning [AzCm]. Reviewers recognize the substantial average gains of +13.38 and +10.19 points over the corresponding 9B and 35B-A3B base models [rzk9, kGZC, krvx].

  • Experimental Validation. Reviewers appreciate the comprehensive tool and training ablations demonstrating the effectiveness of our framework [rzk9, krvx].

New Experiments & Clarifications

We thank all reviewers for their insightful suggestions, which prompted us to conduct substantial new analyses. Beyond our point-by-point responses, we summarize the key experiments and clarifications.

New Analysis and Experiments:

To address concerns regarding baselines, reward reliability, efficiency, and data quality, we have added:

  • Comparison with SOTA Video Agents: VideoSeeker-9B outperforms four recent trained video reasoning methods, surpassing the strongest competitors by +21.29, +14.58, and +11.70 points on CrossVid, CVBench, and MVU-Eval, respectively. [AzCm, kGZC]

  • Training Paradigm Ablation: We introduced Direct SFT and Direct GRPO baselines. VideoSeeker achieves 63.71% average accuracy versus 59.64% for Direct GRPO, confirming gains beyond task-specific training. [krvx]

  • Multi-Seed RL Validation: Three independent runs confirm consistent improvements over GRPO by +1.77, +1.34, and +1.41 points. Removing the grounding gate further reduces performance, validating our reward design. [rzk9, kGZC, krvx, AzCm]

  • Matched-Budget Efficiency: VideoSeeker outperforms frame-matched direct inference by +5.13, +7.73, and +17.98 points. We additionally reported inference latency and tool invocation statistics. [rzk9, kGZC, krvx]

  • Long-Video Evaluation: On 2,003 questions involving videos of at least seven minutes each, VideoSeeker achieves 47.33% accuracy versus 32.47% for frame-matched direct inference. [AzCm]

  • Human Data Validation: Among 4,000 trajectories, 92.43% achieve full teacher consensus. Human verification of the remaining 303 samples confirms 97.03% answer correctness. [AzCm]

New Clarifications:

  • Joint Cross-Video Reasoning: We clarified that the direct baseline already processes videos jointly. VideoSeeker further surpasses independent per-video reasoning by +13.98, +14.51, and +18.60 points, validating explicit cross-video evidence integration. [AzCm, krvx]

  • Visual Grounding & Evidence Faithfulness: We clarified the gate's visual provenance objective and audited 2,060 notes. Results show 96.02% are fully or partially visually supported, while 83.35% support final answers. [rzk9, kGZC, krvx, AzCm]

  • Persistent Evidence Memory: Removing, replacing, or shuffling evidence notes reduces accuracy by 5.77, 5.43, and 4.83 points, confirming the value of structured evidence memory beyond interaction history. [AzCm]

  • Evaluation Fairness: VideoSeeker maintains gains of +12.90 and +11.68 points on the original CVBench and MVU-Eval benchmarks. Additional Gemini 3.8 Flash comparisons further demonstrate its competitiveness. [rzk9, kGZC]

We appreciate the encouraging follow-up feedback: reviewers [krvx, AzCm] raised their ratings from Weak Reject to Weak Accept, while [kGZC] confirmed that our responses addressed their concerns. We believe the detailed responses and substantial new evidence effectively resolve the major concerns raised during review. We sincerely thank the Area Chair and all reviewers for their time and effort dedicated to our submission.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.