acceptodds
Under review as a conference paper at ICLR 2027

Videopus: Parallel Multimodal Evidence Seeking for Long-Horizon Video Understanding

Abstract

Long-horizon video understanding requires models to locate sparse evidence across extended timelines while preserving the visual detail needed for accurate reasoning. Existing agentic systems often inspect candidate intervals sequentially, repeatedly processing an interaction history. Moreover, local observations are usually summarized in text, making it difficult for the global reasoner to verify fine-grained visual details or connect evidence across intervals. We introduce Videopus, a multi-agent framework in which an Orchestrator coordinates interval-specific Executors. It has two components. Parallel Task Spawn groups inspections derived from the same evidence state into a concurrently executable batch. Multimodal Evidence Aggregation returns structured local findings with a compact set of supporting visuals selected from each Executor's input. The Orchestrator uses this evidence to resolve ambiguous descriptions and compare objects or event stages across the video. Local perception remains interval-specific, while global reasoning operates over a compact, visually inspectable evidence state. Videopus alternates parallel inspection with evidence-guided refinement. Across six video benchmarks, Videopus consistently outperforms the state-of-the-art methods while reducing token use and sequential rounds and increasing persistent-context reuse across reasoning turns.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Long Video Understanding, Multimodal Agents, Tool-Augmented Reasoning, Agent Harness

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_9dizOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The empirical evidence is broad. Leading results on six benchmarks, gains with two unrelated open backbones, and ablations of parallel spawning, task objectives, inspection modes, and multimodal aggregation all point in the same direction. The same-backbone efficiency comparison with VideoSeek is particularly useful, although it needs stronger protocol reporting.

Critical strengths

The empirical evidence is broad. Leading results on six benchmarks, gains with two unrelated open backbones, and ablations of parallel spawning, task objectives, inspection modes, and multimodal aggregation all point in the same direction. The same-backbone efficiency comparison with VideoSeek is particularly useful, although it needs stronger protocol reporting.

Critical weaknesses

The main comparative evidence still does not fully isolate the framework’s contribution. Videopus uses GPT-5 as both Orchestrator and Executor and can invoke multiple models per query, but most headline tables compare its results with heterogeneous methods using different models, sampling protocols, and numbers of inference calls. A direct GPT-5 end-to-end baseline is absent on most benchmarks, and it is unclear whether VideoSeek and every other baseline were rerun with identical subtitles, visual-unit budgets, prompts, and answer parsing. Because the paper claims consistent superiority over state-of-the-art methods, these controls are essential.

The efficiency and robustness claims also remain under-specified. Token and model-call accounting does not explain whether all parallel Executor calls, image tokens, retries, and failed sessions are included.

Questions for authors
  1. Please provide no-tool GPT-5 results on all six benchmarks under the same global sampling, subtitle policy, prompts, and answer parser. Also specify exactly which benchmarks used subtitles or ASR transcripts.

  2. Were VideoSeek and every rerun baseline given the same base model, visual-unit budget, executor concurrency, inference-round limit, and retry policy? Please report matched-budget results rather than only best reported scores.

  3. Please disaggregate average tokens and model calls into Orchestrator, Executor, image, input, output, and retry costs.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 9diz [1/2]

Thank you for your valuable feedback, which has helped us further improve our work. We have revised the manuscript accordingly and added the requested material to the appendix. -Abl denotes our ablation subset; -Ful or no suffix denotes the full dataset.

W1&Q1. No-tool GPT-5 controls and input policy.

To isolate the contribution of Videopus from the backbone model, we evaluate GPT-5 (No Tools) on all six benchmarks. Both settings use the same GPT-5 version, initial global visual sampling, question prompts, and answer parsing rules. No-tool GPT-5 answers directly from the initial visual observations, whereas Videopus performs targeted inspections and iterative evidence aggregation.

Regarding input policy, none of our experiments use ASR. Only LongVideoBench uses dataset-provided subtitles following its evaluation protocol. LVBench, EgoSchema, MLVU, Video-MME, and Video-Holmes use neither subtitles nor additional video captions. The new no-tool controls follow the same policy, including identical subtitle inputs for LongVideoBench.

MethodLVBench↑ER↑EU↑Rea↑KIR↑Sum↑TG↑EgoSchema Val.↑
GPT-5 (No Tools)63.260.465.164.766.362.160.075.4
Videopus75.673.176.069.281.182.880.977.2
MethodLongVideoBench↑Long↑MLVUVideo-MME Short↑Medium↑Long↑Average↑
GPT-5 (No Tools)71.367.277.080.979.472.677.6
Videopus78.175.582.986.983.076.482.1
MethodSR↑IMC↑TCI↑TA↑MHR↑PAR↑CTI↑Overall↑
GPT-5 (No Tools)57.548.951.652.048.849.040.749.8
Videopus63.454.757.558.054.855.246.755.7

Videopus consistently outperforms no-tool GPT-5 across all six benchmarks by 1.8-12.4 percentage points. On LVBench, Summarization and Temporal Grounding improve by 20.7 and 20.9 points, respectively. These results demonstrate the value of targeted evidence acquisition through Parallel Task Spawn and source-grounded visual aggregation through Multimodal Evidence Aggregation, which help recover information missed by global sampling and support cross-interval reasoning.

W1&Q2. Matched-budget comparisons and rerun protocols.

Like previous works, our main comparisons use published results to preserve each method's original model and inference settings, avoiding potential performance degradation from reimplementation. For example, VideoARM evaluated GPT-5 but achieved better performance with OpenAI o3 + GPT-4o. Thus, the main tables report performance under each method's original configuration rather than controlled comparisons. We have clarified the backbone models, additional input modalities, and result sources in Appendix C of the revised manuscript.

The paper already compares Videopus with VideoSeek using the same GPT-5 backbone, showing a 7.2 pp accuracy improvement and 44.6% fewer tokens on full LVBench. To further control for backbone differences, we rerun ParaVT, VideoSeek, and VideoARM with GPT-5. Their inference frameworks are adapted to GPT-5 where necessary, with ASR disabled for VideoARM to match our visual-only setting.

LVBench-FulVenueBackboneAvg. ACC↑ER↑EU↑Rea↑KIR↑Sum↑TG↑
ParaVTNeurIPS'26GPT-562.258.665.560.760.167.274.5
VideoSeekCVPR'26GPT-565.763.766.863.264.679.380.5
VideoARMCVPR'26GPT-570.468.768.668.777.075.976.8
Videopus—GPT-575.673.176.069.281.182.880.9

These reruns use the same GPT-5 version, evaluation samples, input modalities, answer parsing, and retry policy, while preserving each method's agent instructions, tool sampling, and task planning. Since visual-unit definitions, call structures, concurrency, and round limits differ across frameworks, we retain their original mechanisms and additionally compare VideoSeek and Videopus under matched total-token budgets.

AuthorsOpenReview ↗

Response to Reviewer 9diz [2/2]

Specifically, we evaluate VideoSeek and Videopus on LVBench-Abl with the same 100K-token budget target and 84K-token pre-call threshold. Token accounting includes text and image inputs and outputs from all model calls, including the Orchestrator and Executors.

LVBench-AblBackboneToken BudgetACC↑Actual Tokens (K)↓
VideoSeekGPT-5100K target / 84K pre-call check62.367.7
VideopusGPT-5100K target / 84K pre-call check74.265.2

Both methods follow the same budget policy, while actual token consumption may differ due to early stopping and evidence acquisition efficiency. We account for actual image tokens and all model calls rather than treating equal visual-unit counts as equivalent computational budgets.

Under the unified GPT-5 setting, Videopus outperforms ParaVT, VideoSeek, and VideoARM by 13.4, 9.9, and 5.2 pp, respectively. Under matched token budgets, it maintains an 11.9 pp advantage over VideoSeek while using 2.5K fewer tokens on average. These results show that Videopus's gains persist under both backbone-controlled and budget-matched comparisons, supporting the effectiveness of its evidence acquisition and aggregation mechanisms.

W2&Q3. Breaking down all calls, tokens, and retries by role.

Section 4.4 reports average tokens and model calls on full LVBench, while Appendix F analyzes repeated context-processing costs. We further provide a request-level breakdown of text input, image input, output tokens, and model calls, covering all parallel Executors in Videopus and the corresponding components in VideoSeek.

LVBench-FulComponentText Input (K)Image Input (K)Output (K)Total Tokens (K)Model Calls / QA ↓
VideoSeekMain Agent65.20.013.578.65.5
Action Parser7.70.02.710.35.5
Answer Tool11.50.00.211.71.0
Overview&Inspection Tools1.640.09.651.24.8
Total86.040.025.9152.016.9
VideopusOrchestrator15.018.24.337.62.8
Executors5.521.819.346.63.9
Total20.540.123.684.26.8

All values are per-question averages. Minor discrepancies in component sums arise from rounding. Videopus reduces model calls by 59.8% and total tokens by 44.6%, including all 46.6K tokens consumed by parallel Executors. While image input remains nearly identical (40.0K vs. 40.1K), text input decreases from 86.0K to 20.5K (76.2%). These results show that the token savings primarily stem from fewer global reasoning rounds and less repeated context processing, rather than reduced visual input.

Regarding retries, the Videopus Harness handles non-fatal intermediate errors and reports them to the Orchestrator without restarting execution. Only execution-blocking failures, such as empty Orchestrator outputs, missing required Evidence fields in Executor reports, or tool execution failures, trigger local task restarts or full trajectory reruns. No such retries occurred in the full LVBench logs, resulting in zero additional model calls or token costs.

Additional Revisions and Experiments

Beyond the concerns addressed above, we have further clarified benchmark protocols in Appendix B, baseline configurations and result sources in Appendix C, and distinctions from related video agents in Appendix D. We have also expanded Appendix E with additional controlled comparisons, full LVBench ablations, inspection-policy analyses, hyperparameter sensitivity studies, and cross-backbone generalization. Appendix I further presents systematic failure and recovery analyses with additional inference trajectories.

We hope these revisions provide further evidence for assessing the effectiveness, efficiency, and generalization of Videopus. We appreciate your thoughtful feedback and consideration.

Thank you for the detailed rebuttal and additional experiments. Please make sure these results and protocol clarifications are clearly integrated into the main text as well as the appendix. I have raised my score to 3.

We sincerely appreciate your careful reassessment and for raising your score to 3. We are encouraged by your recognition of our additional experiments and clarifications. We will ensure that the key results and protocol details are clearly integrated into the main text, with further details provided in the appendix. Thank you again for your constructive feedback, which has helped us improve the manuscript.

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

This paper uses a parallel, iterative multi-agent system to improve long-video understanding. Global sampling provides a coarse overview, followed by parallel tasks in detail or coverage mode. Their results guide the system to answer or inspect further. The method improves accuracy across multiple benchmarks while reducing inference costs.

Critical strengths
  1. Parallel inspection and visual evidence return are well motivated by the challenges of locating sparse evidence across long videos while preserving the visual details needed for accurate reasoning.
  2. The evaluation covers multiple benchmarks, with nearly consistent improvements in overall accuracy.
  3. The proposed method improves accuracy while reducing model calls, token use, and reasoning rounds compared with VideoSeek using the same backbone.
Critical weaknesses
  1. The main contribution is the overall framework design, including parallel task execution and unified iterative reasoning. It is therefore important to evaluate its generalizability across both backbones and datasets, whereas the current cross-backbone evaluation is limited to LVBench.
  2. Additional cases would help illustrate both how the system recovers from missing or misleading evidence across rounds and whether parallel tasks provide complementary evidence without redundant inspection.
  3. It is unclear what changes when Parallel Task Spawn is disabled. Does the system execute the same tasks sequentially, or does it also change how tasks are planned? This distinction matters because the accuracy gain may come from broader evidence coverage rather than parallel execution itself.
Questions for authors

See Weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer ygHS [1/2]

Thank you for your valuable feedback, which has helped us further improve our work. We have revised the manuscript accordingly and added the requested material to the appendix. -Abl denotes our ablation subset; -Ful or no suffix denotes the full dataset.

W1. Cross-backbone and cross-dataset generalization.

The original paper evaluates Videopus across six long-video benchmarks and further examines two open-source backbones, Qwen3.6-27B and Gemma4-31B, in Table 5. Under matched visual token budgets, Videopus improves LVBench accuracy by 11.6 and 9.7 percentage points, respectively, showing that its gains are not limited to a single backbone.

We extend both backbone comparisons to Video-MME Long and LongVideoBench, using the same Videopus configuration and inference budgets without dataset-specific tuning.

BackboneMethodLVBench-Abl ACC↑Video-MME Long ACC↑LongVideoBench ACC↑
Qwen3.6-27BBase56.866.970.7
Qwen3.6-27BVideopus68.470.677.4
Gemma4-31BBase61.369.472.2
Gemma4-31BVideopus71.072.877.8

Both backbones consistently improve across all three benchmarks. These results support the effectiveness of Parallel Task Spawn and Multimodal Evidence Aggregation across different backbone families and task distributions, rather than being specific to LVBench or a particular model.

W2. Recovery and complementary evidence acquisition.

Videopus uses structured findings, reliability information, and source visuals to help the Orchestrator identify missing or conflicting evidence and plan targeted follow-up inspections. In Figure 8, [Coverage] locates the throw-in event but cannot identify the jersey number, as noted in the Executor's reliability report. The Orchestrator then narrows the interval and switches to [Detail] for verification. Figure 3 similarly shows how parallel [Coverage] inspections identify competing candidate labels, followed by targeted [Detail] inspection to verify the decisive evidence rather than accepting local findings directly.

Complementarity among parallel tasks comes from task-specific objectives rather than redundant sampling. In Figure 7, three Executors independently inspect identity, matchup information, and clothing cues. No single finding is sufficient to answer the question, but their aggregation establishes cross-interval associations. Tasks operate independently from the same pre-round evidence state, providing complementary evidence, while subsequent inspections address newly identified gaps by adjusting intervals or modes rather than repeating earlier observations.

We have added four illustrative inference trajectories in Appendix I.3 (Figures 9-12), showing the detection and correction of misleading local findings and the acquisition of non-redundant complementary evidence. These trajectories preserve inspection intervals, objectives, modes, reliability reports, source visuals, and subsequent decisions. The cases demonstrate that evidence-guided refinement enables targeted recovery, while complementary inspections and cross-interval visual associations support reasoning that individual Executors cannot resolve independently.

W3. Separating batch planning from concurrent execution.

Parallel Task Spawn (PTS) combines batch task planning with concurrent execution. The w/o PTS ablation in Table 4 does not sequentially execute the same pre-planned batch. Instead, it restricts the Orchestrator to planning one inspection per round, updating the evidence state and replanning only after receiving its result. In contrast, PTS plans multiple inspections with distinct objectives and intervals from the same pre-round evidence state, aggregating their results only after the entire batch completes. Thus, the ablation changes both task planning and execution concurrency, rather than merely serializing parallel tasks.

We additionally measure end-to-end latency with and without PTS under the same load of 32 concurrent QA requests.

LVBench-AblPlanningACC↑Tokens (K)↓Rounds↓Latency (s)↓
VideopusBatch77.777.32.7476.4
w/o Parallel Task SpawnSequential73.987.53.3775.2

Removing PTS reduces accuracy by 3.8 percentage points while increasing token consumption, reasoning rounds, and end-to-end latency. This accuracy difference should not be attributed to execution concurrency alone, since the ablation also changes how evidence acquisition is organized. Batch planning allows the Orchestrator to collect and jointly assess complementary evidence across different intervals and objectives within a single round, reducing repeated global decisions and context processing. Concurrent execution further shortens the time required to complete each batch. These results indicate that the benefits of PTS are not limited to execution concurrency but are closely related to how evidence acquisition and global reasoning are organized.

AuthorsOpenReview ↗

Response to Reviewer ygHS [2/2]

Additional Revisions and Experiments

Beyond the concerns addressed above, we have further clarified settings and protocols in Appendix B, baseline, auxiliary modalities, and result sources in Appendix C, and the distinctions between Videopus and related video agents in Appendix D.

We have also expanded the experimental analysis in Appendix E, including controlled GPT-5 comparisons, extensions to existing agents, full LVBench ablations with statistical analysis, inspection-policy and hyperparameter studies, and detailed accuracy-cost-latency evaluations. Appendix I further provides systematic failure and recovery analysis, together with additional inference trajectories illustrating successful recovery, complementary evidence acquisition, and remaining failure cases.

We hope these revisions provide further evidence for assessing the contributions and effectiveness of Videopus. We sincerely appreciate your thoughtful feedback and consideration.

Thanks for the added experiments and explanations. I still think the PTS ablation leaves open how much comes from batch planning versus concurrent execution, but the other concerns have been addressed. I’ve raised my score and lowered my confidence, as I’m less certain about this part of the evaluation.

We sincerely thank the reviewer for the thoughtful reassessment and for raising the score to 3, which is very encouraging to us! We greatly appreciate the constructive feedback and would like to further clarify the distinct roles of batch planning and concurrent execution.

Our Harness independently limits the maximum tasks per batch (12 by default) and concurrent Executors (8 by default) through runtime controls and prompt instructions, preventing excessive task spawning and execution. The former controls planning capacity, while the latter determines execution parallelism and is primarily constrained by the API provider and gateway capacity.

Batch Planning and Complementary Evidence Acquisition.

Batch planning allows the Orchestrator to assign inspections with complementary objectives, intervals, and modes from the same evidence state, then jointly assess their findings within one reasoning round. In contrast, sequential planning inspects one task at a time and replans after each result, requiring repeated global decisions and context processing.

We further evaluate different batch planning capacities:

LVBench-AblMax Tasks / BatchACC↑Tokens (K)↓Rounds↓
Fewer tasks474.283.32.83
Fewer tasks875.888.02.83
Videopus1277.777.32.71
More tasks1675.583.82.81

Smaller batches restrict complementary evidence acquisition across intervals and objectives, potentially requiring additional rounds and repeated context processing. Notably, increasing planning capacity from 8 to 12 improves accuracy by 1.9 pp and reduces token consumption by 10.7K, while keeping execution concurrency capped at 8. Conversely, excessive capacity may introduce redundant or less informative inspections, increasing aggregation overhead without improving accuracy. These results indicate that the default capacity balances evidence coverage and planning efficiency.

Concurrent Execution and Inference Efficiency.

Concurrent execution processes independently planned inspections simultaneously. Since the Orchestrator updates its evidence state only after batch completion, changing execution concurrency does not alter the planned inspections or evidence aggregation. We measure latency under the same load of 32 concurrent QA requests:

LVBench-AblMax ConcurrencyRounds↓Latency (s)↓
Videopus42.7542.2
Videopus82.7476.4

Increasing concurrency from 4 to 8 reduces latency by 65.8s without changing the average reasoning rounds. This supports its primary role in execution efficiency rather than accuracy-oriented planning.

Overall, the w/o PTS ablation reduces accuracy from 77.7% to 73.9%, increases token consumption from 77.3K to 87.5K, and increases latency from 476.4s to 775.2s. Together, these results show that batch planning organizes complementary evidence acquisition and reduces repeated global reasoning, while concurrent execution primarily reduces latency. Their combination improves long-video reasoning under bounded task and concurrency budgets.

We sincerely appreciate the reviewer's careful consideration and constructive suggestions, which have helped us improve the clarity of our work!

Reviewer_tP9HOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper presents a contribution to long-horizon video understanding, where the goal is to locate sparse evidence across extended timelines while preserving the visual detail needed for accurate reasoning.

To this end, the authors present Videopus, a multi-agent framework with two main components:

  • Parallel Task Spawn, which groups inspections derived from the same evidence state into a concurrently executable batch.
  • Multimodal Evidence Aggregation, which returns a structured set of local findings supported by visuals selected from each Executor’s input.

Videopus alternates parallel inspection with evidence-guided refinement. I see the main contribution to the literature as a less expensive (token-wise) approach than the state of the art, without the trade-off of reduced accuracy, as Videopus outperforms state-of-the-art methods across six video benchmarks.

Critical strengths

Good experimental results: Videopus outperforms state-of-the-art methods across six video benchmarks, which positions the contribution well in the literature and relies on various dataset comparisons.

Important topic: Less expensive long-horizon video understanding is fundamental in the literature. Under a finite inference budget, a model must be robust enough to search across an extended timeline with sufficient visual fidelity to resolve decisive local cues, while some cues might be a brief interaction occupying only a few seconds of an hour-long video. To me, this is what makes authors contribution to long-horizon video understanding interesting, and I liked seeing improvement in that regard. The main topic of this contribution is well stated (existing methods, bottlenecks, and a contribution based on these principles) and important to the field.

Critical weaknesses

Ablation is not strong enough. All ablations use the same fixed 20% stratified subset of LVBench, which makes one 310-question subset. This is very important, as some scores, and therefore some conclusions the authors make, rest on differences of a few points. For exemple, with only 310 questions, the two ablated variants (75.5 and 75.8) in Table 4: differ by 0.3 points, which is one question.

The closest related work is not compared. The related work section says that Yang et al. (2026b), ParaVT, already proposes parallel tool use in agentic video understanding, with a main agent that dispatches concurrent inspections and reasons over the returned evidence. It seems to be one of the closest related works, but the authors never compare their contribution with this paper. Stating that it returns only text, instead of the multimodal text and visual cues they use, is not sufficient. This is especially true because the ablation shows results with text-only cues but never compares with ParaVT, even in that section. This makes the improvement more difficult to read and analyse, and it is unclear whether some strategies are indeed better.

Orchestrator modes are not sufficiently explored, while being, to me, one major improvement. The Orchestrator chooses one mode for each inspection: coverage or detail. Figure 7 shows 3 inspections using detail mode. Figure 8 shows 2 coverage inspections first, then a detail one to read the player’s number. To me, the paper gives a general justification (coverage to localize events, detail for small objects and scene text), and Table 4 shows the contribution of each mode. However, it does not state explicit criteria for how the Orchestrator picks one mode over another, and this selection is never experimentally tested (for instance, with mode usage statistics or a forced or random mode choice). It is left to the reader's own interpretation. The paper needs more clarity on that part.

The configuration of each mode is also predefined (Table 7). The Orchestrator chooses the mode but does not change the resolution or sampling density. My guess is that these have an influence on how to guide the next inspection, which is something that is missing. Furthermore, the experiments don’t show the importance of the confidence (the reliability information U_i), or how it is used as a factor in choosing which mode to pick next. An ablation without the confidence section might help in that regard.

Missing experiments and unclear baselines. Overall, I feel that some ablation studies and experiments are missing to make the contribution more appealing. The authors also never state which backbones DVD, MR.Video, VCA, VideoLucy and the others use. This makes the evidence supporting the claim that the paper “consistently outperforms state-of-the-art methods” less convincing.

Novelty is incremental, as the prompt-level orchestration design combines known elements from the literature (VideoSeek, MR.Video, ParaVT).

Questions for authors
  1. Can you report confidence intervals, results across repeated runs, or paired significance tests? Additionally, can you perform the ablation studies on the full LVBench dataset to assess the robustness and generalizability of the results?
  2. What criteria does the Orchestrator use to choose between coverage and detail modes, and are these criteria explicitly specified in the prompt? Can you report the frequency of each mode and compare the current selection policy against forced and random mode selection?
  3. How does the reliability information (U_i) influence the selection of the next interval or mode? Could you include an ablation study that removes (U_i), or analyze whether lower confidence levels are associated with switches to detail mode? This would help clarify the contribution
3 comments
AuthorsOpenReview ↗

Response to Reviewer tP9H [1/3]

Thank you for your valuable feedback, which has helped us further improve our work. We have revised the manuscript accordingly and added the requested material to the appendix. -Abl denotes our ablation subset; -Ful or no suffix denotes the full dataset.

W1&Q1. Full LVBench ablations and statistical reliability.

We extend four key ablations to full LVBench under matched evaluation settings. Using video-clustered paired bootstrap, we report accuracy (ACC), its 95% confidence interval (CI), and paired accuracy differences (ΔACC\Delta\mathrm{ACC}) against full Videopus with their CIs. This accounts for correlations among questions from the same video while preserving paired comparisons.

LVBench-FulACC↑95% CIΔACC\Delta\mathrm{ACC} (pp)95% CI of ΔACC\Delta\mathrm{ACC}Tokens (K)↓Rounds↓
Videopus75.6[73.0, 78.2]0.0—84.22.8
w/o Parallel Spawn72.2[69.3, 75.1]−3.4[−5.6, −1.3]95.93.2
w/o Textual Report (Ei+UiE_i+U_i)73.3[70.8, 75.9]−2.3[−4.0, −0.5]84.32.9
w/o Returned Visuals72.9[70.3, 75.4]−2.7[−4.3, −1.1]81.62.9
w/o [Detail] ([Coverage]-only)73.1[70.5, 75.7]−2.5[−4.5, −0.6]82.52.7

Each ablation changes only the corresponding mechanism. Without Parallel Task Spawn, the Orchestrator plans one inspection per round and replans after each result. Removing textual reports preserves task metadata and source visuals; removing returned visuals preserves textual findings; removing [Detail] enforces [Coverage]-only inspection.

All four ablations reduce accuracy, with paired 95% CIs for ΔACC\Delta\mathrm{ACC} excluding zero, consistent with Table 4. Removing Parallel Task Spawn also increases tokens and reasoning rounds. These results confirm that the key components contribute to accuracy on LVBench.

W2. Separating ParaVT from its same backbone adaptation.

ParaVT supports parallel inspection but relies on generic summarization objectives, fixed sampling configurations, and text-only returns. In contrast, Videopus uses evidence-specific objectives, complementary sampling modes, and source-grounded multimodal aggregation for direct verification and cross-interval reasoning. Thus, the w/o Visual Return ablation in Table 4 is not equivalent to ParaVT, as it retains Videopus's task planning and inspection mechanisms. We have clarified these distinctions in Appendix D of the revised manuscript.

We further adapt ParaVT's agent inference framework to GPT-5 for a direct comparison under the same backbone and LVBench evaluation settings.

MethodBackboneLVBench ACC↑ER↑EU↑Rea↑KIR↑Sum↑TG↑
ParaVTGPT-562.258.665.560.760.167.274.5
VideopusGPT-575.673.176.069.281.182.880.9

Videopus outperforms ParaVT across all task categories, particularly in key-information retrieval and summarization. Together with the component ablations in Table 4, these results show that evidence-specific inspection and verifiable multimodal aggregation support more effective long-video reasoning than parallel inspection with text-only summaries.

W3&Q2. Mode-selection criteria and policy comparisons.

Videopus specifies mode selection through both prompt instructions and sampling configurations. Appendix J.1 directs [Coverage] toward event localization and temporal tracking, and [Detail] toward fine-grained verification within short intervals. Both share a 16-visual-unit cap: [Coverage] packs up to 64 frames into 2×22\times2 grids, while [Detail] uses up to 16 individual high-resolution frames. The 16-second [Detail] limit aligns with its frame budget, with longer requests automatically falling back to [Coverage] to avoid sparse sampling. This design supports evidence-driven mode selection rather than a fixed coarse-to-fine sequence, as illustrated in Figures 7-8.

Beyond Table 4, we compare adaptive selection against Forced [Coverage], Forced [Detail], and Random Mode, reporting executed mode frequencies and inference costs. Unlike Table 4, these controls override only the inspection mode after task planning, preserving the objectives and intervals.

LVBench-AblACC↑Tokens (K)↓Rounds↓Executed [Coverage] (%)Executed [Detail] (%)
Videopus77.777.32.7148.751.3
Forced [Coverage]74.281.42.72100.00.0
Forced [Detail] (Falling back disabled)74.588.22.850.0100.0
Random Mode74.591.32.8768.431.6

Adaptive selection achieves higher accuracy and lower token consumption than fixed or random policies. Forced [Coverage] limits fine-detail verification, while Forced [Detail] weakens temporal coverage over long intervals. Random selection also fails to match sampling to evidence needs. These results demonstrate that evidence-driven mode selection effectively balances temporal coverage and spatial detail under constrained visual budgets.

AuthorsOpenReview ↗

Response to Reviewer tP9H [2/3]

W4.1 Mode configurations and their effects.

Videopus uses two predefined inspection modes rather than requiring the Orchestrator to jointly optimize sampling density, resolution, and input count, reducing planning complexity. The Orchestrator still adapts inspection intervals, objectives, and modes based on evidence needs. A fixed frame cap does not imply fixed sampling density: shorter intervals allow denser temporal sampling, while [Detail] preserves finer visual information. Executor-reported sampling and visibility limitations guide subsequent inspections. In Figure 8, the Orchestrator narrows the interval and switches to [Detail] after [Coverage] locates the event but fails to resolve the jersey number.

We further vary global grid count, local visual budgets, resolution, and [Detail] duration, changing one parameter at a time while updating the corresponding prompt and tool constraints.

LVBench-AblGlobal Grids[Coverage] Grids / Range[Detail] Frames / RangeLocal Resolution (Long / Short)Max [Detail] Duration (s)Max Tasks / BatchACC↑Tokens (K)↓Rounds↓
Default6416161024 / 512161277.777.32.71
Fewer global grids3216161024 / 512161273.589.13.04
More global grids9616161024 / 512161278.487.02.64
Fewer local units64881024 / 512161274.575.22.89
More local units6424241024 / 512161278.490.32.75
Lower local resolution641616768 / 384161275.882.82.73
Higher local resolution6416161280 / 640161277.7119.82.71
Shorter [Detail] interval6416161024 / 51281276.188.92.87
Longer [Detail] interval6416161024 / 512321277.181.92.79
Fewer tasks per batch6416161024 / 51216474.283.32.83
Fewer tasks per batch6416161024 / 51216875.888.02.83
More tasks per batch6416161024 / 512161675.583.82.81

Reducing global sampling or local visual budgets lowers accuracy, while increasing them provides limited gains at higher token costs. Higher resolution yields no further improvement, and [Detail] duration shows non-monotonic effects. These results show that the predefined configurations balance accuracy and cost, while adaptive interval and mode selection enables effective evidence-guided refinement without continuous parameter optimization.

W4.2&Q3. Reliability information and subsequent inspections.

Reliability information (UiU_i) describes sampling and visibility limitations rather than providing a numerical confidence score or triggering fixed mode-switching rules. The Orchestrator uses (UiU_i), textual findings (EiE_i), and source visuals (ViV_i) to guide subsequent inspections: insufficient temporal sampling may trigger further [Coverage] inspection, while unclear visual details may lead to a narrower interval and [Detail] inspection. Figure 8 illustrates the latter, where the Orchestrator switches to [Detail] to verify an unreadable jersey number after locating the throw-in event.

We further ablate (UiU_i) while retaining (EiE_i) and (ViV_i), comparing accuracy, inference costs, and executed mode frequencies under identical evaluation settings.

LVBench-AblReliability Report UiU_iACC↑Tokens (K)↓Rounds↓Executed [Coverage] (%)Executed [Detail] (%)
Videopus✓77.777.32.7148.751.3
w/o Reliability Report UiU_i✗75.289.22.7846.453.6

Removing (UiU_i) reduces accuracy and increases token consumption despite a slightly higher proportion of [Detail] inspections. This shows that additional [Detail] inspection cannot replace reliability assessment, and that (UiU_i) helps the Orchestrator plan more effective follow-up inspections based on the specific causes of missing or uncertain evidence.

AuthorsOpenReview ↗

Response to Reviewer tP9H [3/3]

W5. Separating backbone effects from framework gains.

Like most previous works, our main comparisons primarily use published results to preserve each method's selected backbone and inference settings, avoiding potential degradation from reproduction differences. For example, VideoARM reports better performance with o3 + GPT-4o than with GPT-5. Thus, these results reflect each method's original configuration rather than a fully matched comparison. We have also clarified the backbones, auxiliary models, and additional modalities used by DVD, MR.Video, VCA, VideoLucy, and other baselines in Appendix C of the revised manuscript.

The original paper already compares Videopus with VideoSeek under the same GPT-5 backbone, showing higher accuracy and lower token consumption. We further rerun ParaVT, VideoSeek, and VideoARM on full LVBench using GPT-5 and a consistent evaluation protocol while preserving their agent mechanisms. ASR is disabled for VideoARM to maintain a consistent input modality.

LVBench-FulVenueBackboneAvg. ACC↑ER↑EU↑Rea↑KIR↑Sum↑TG↑
ParaVTNeurIPS'26GPT-562.258.665.560.760.167.274.5
VideoSeekCVPR'26GPT-565.763.766.863.264.679.380.5
VideoARMCVPR'26GPT-570.468.768.668.777.075.976.8
Videopus—GPT-575.673.176.069.281.182.880.9

Videopus maintains higher accuracy across all LVBench categories under the same backbone. Together with the component ablations, these results support the contribution of task-specific evidence acquisition and multimodal aggregation beyond backbone selection.

W6. Novelty beyond existing video agents.

VideoSeek uses a text-based Overview that may lose visual cues. Its shared frame cap makes long-range Skim sampling sparse, while serial inspections repeatedly process growing histories. MR.Video parallelizes video processing but requires full-video preprocessing and summarization. ParaVT supports parallel inspection but uses generic summarization objectives and fixed sampling configurations, returning only text and limiting cross-interval visual verification.

Videopus instead retains a global visual overview. Parallel Task Spawn plans objective-specific inspections with distinct intervals and modes from a shared evidence state. Under a bounded visual budget, densely sampled temporal grids locate brief events, while individual high-resolution frames verify details. Executors run concurrently in isolated contexts, with global evidence updated only after batch completion, reducing sequential reasoning rounds and repeated context processing. Multimodal Evidence Aggregation returns structured findings, reliability information, and selected source visuals for direct verification and cross-interval reasoning. These observations guide subsequent inspections, connecting budget-aware task planning with visually grounded refinement.

Reviewer_cUbeOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Videopus, a multi-agent framework for long-video understanding that coordinates targeted evidence acquisition through an Orchestrator and interval-specific Executors. Starting from a sparsely sampled visual overview, the Orchestrator launches parallel inspections in coverage or detail mode, and Executors return structured findings alongside selected supporting visuals. The Orchestrator integrates this evidence to answer the question or initiate further inspection. Across six video-understanding benchmarks, the paper reports improved accuracy over prior methods, together with reduced token consumption.

Critical strengths

The paper addresses a valuable problem in long-video understanding, where effective agentic evidence acquisition can substantially improve upon standalone VLMs, particularly under constrained inference budgets. The proposed framework is intuitive and well motivated: parallel interval inspection enables broader evidence gathering within each reasoning round, while selective visual evidence return supports verification and cross-interval reasoning. Evaluation across six benchmarks demonstrates consistently strong performance, and the reported reductions in token consumption and model calls suggest practical value, although the sources of these gains warrant stronger controlled analysis. Overall, the paper is clearly written and well presented, with helpful illustrations and detailed descriptions of the system.

Critical weaknesses
  1. The methodological novelty appears incremental relative to prior agentic video reasoning systems. Question-dependent temporal search, iterative evidence acquisition, and coarse-to-fine inspection are already central to VideoSeek, while ParaVT already dispatches multiple temporal inspections within a single turn. Videopus primarily combines these established ideas with task-specific objectives and the return of selected source visuals. This is a reasonable systems design, but the paper does not sufficiently establish a distinct algorithmic contribution beyond that combination. Although Table 4 provides component ablations, the paper lacks a direct comparison with ParaVT and a controlled baseline that augments an existing agent with batching and visual evidence return. Such comparisons would clarify whether the gains require the proposed framework or follow from straightforward extensions to existing methods.

  2. The experiments do not establish the comparative accuracy–cost frontier, particularly at larger inference budgets. More precisely, The experiments do not establish whether Videopus remains preferable when competing agents receive substantially more computation. The authors should sweep inference budgets and report accuracy against actual total tokens, monetary cost, and wall-clock latency. This would reveal whether Videopus improves the Pareto frontier or primarily benefits the evaluated budget regime. The decline from 77.7% to 76.5% when increasing the round cap from eight to twelve is also worth investigating.

  3. Hyperparameter selection and robustness are insufficiently established. The framework has many manually specified settings: global sampling density and grid count, grid layouts, local image budgets and resolution, the 16-second detail-mode threshold, eight concurrent Executors, the visual-return cap, and the reasoning-round cap. Table 6 studies only the last two on the same 310-question LVBench subset used for the component ablations. The paper does not explain whether these values were selected using independent development data or whether the ablation subset influenced their selection.

Questions for authors
  1. Do the agentic baselines use the same VLM backbone and inference settings as Videopus? Beyond the matched backbone comparison with VideoSeek, could the authors clarify which results come from controlled reruns versus prior publications?

  2. How were the hyperparameters selected? Was tuning performed on a separate development set, and were the resulting settings fixed across all benchmarks?

  3. Could the authors report end-to-end latency under comparable hardware and API conditions, including the concurrency requirements? Token counts and reasoning rounds alone do not establish the practical speedup.

  4. Could the authors provide a systematic failure analysis, particularly for missed events during sparse sampling, incorrect or conflicting Executor findings, and omitted visual evidence? How effectively can the Orchestrator detect and recover from these errors?

3 comments
AuthorsOpenReview ↗

Response to Reviewer cUbe [1/3]

Thank you for your valuable feedback, which has helped us further improve our work. We have revised the manuscript accordingly and added the requested material to the appendix. In each table, -Abl denotes our ablation subset; -Ful or no suffix denotes the full dataset.

W1.1 Novelty and distinguishing from previous works.

Videopus goes beyond combining coarse-to-fine search with parallel tool calls. Its contribution is a budget-constrained evidence loop that connects complementary task planning with verifiable multimodal evidence updates.

VideoSeek compresses its global overview into text, potentially losing visual cues. Its [Skim] and [Focus] share the same frame cap, making long-range [Skim] sampling sparse, while serial calls repeatedly process growing histories. ParaVT supports parallel inspection but uses generic objectives and fixed sampling configurations, returning only textual summaries.

Videopus instead retains a global visual overview. Parallel Task Spawn plans objective-specific inspections from a shared evidence state, using densely sampled temporal grids to locate brief events or individual high-resolution frames to verify details under a bounded visual budget. Executors run concurrently in isolated contexts, with global evidence updated only after each batch, reducing sequential reasoning rounds and repeated processing of accumulated context. Multimodal Evidence Aggregation returns structured findings, reliability information, and selected source visuals for direct verification and cross-interval reasoning. These observations guide subsequent inspections, connecting budget-aware task planning with visually grounded refinement rather than simply adding parallel execution and visual returns.

W1.2 Controlled comparison with existing Agents

To further distinguish the contribution of Videopus from straightforward extensions to existing agents, beyond Table 4, we conduct two controlled comparisons. We extend VideoSeek with Batching and Visual Return, individually and jointly. All variants retain VideoSeek's original tools and sampling configurations and use the same LVBench ablation subset.

LVBench-AblBatchingVisual ReturnACC↑Tokens (K)↓Rounds ↓
VideoSeek✗✗65.8155.65.6
VideoSeek✓✗71.6127.54.6
VideoSeek✗✓69.7157.65.4
VideoSeek✓✓73.5129.34.4
Videopus✓✓77.777.32.7

Batching improves accuracy and reduces inference costs, while Visual Return improves accuracy at the cost of additional visual context. However, their combination still falls behind Videopus in both accuracy and token efficiency, demonstrating that the complete framework provides further gains beyond these two extensions.

We further adapt ParaVT to GPT-5 for a direct comparison on full LVBench under the same backbone and input policy.

LVBench-FulBackboneACC↑ER↑EU↑Rea↑KIR↑Sum↑TG↑
ParaVTGPT-562.258.665.560.760.167.274.5
VideopusGPT-575.673.176.069.281.182.880.9

Videopus outperforms ParaVT across all categories, particularly in key-information retrieval and summarization. Together with Table 4, these results show that task-specific planning, complementary visual inspection under constrained budgets, and source-grounded evidence aggregation provide gains beyond simply adding batching and visual returns.

W2&Q3. Accuracy-cost trade-offs and practical efficiency.

Videopus improves efficiency and performance through targeted evidence acquisition and reduced context processing, rather than larger inference budgets. Beyond the 2/4/8/12-round ablation in Table 6, we increase VideoSeek's round cap from 20 to 32 on the same LVBench subset. Since real-time billing is unavailable on our gateway, API costs are estimated at $1.25/10.00 per million input/output tokens.

LVBench-AblMax. RoundsACC↑Tokens (K)↓Est. API Cost ($)↓Forced Ans. (%)↓
VideoSeekDefault (20)65.8155.60.4312.9
VideoSeekExtended (32)65.8186.40.7171.3
VideopusDefault (8)77.777.30.2881.9

Increasing VideoSeek's budget reduces forced answers but leaves accuracy unchanged while increasing costs. Videopus remains more accurate and less costly. Table 6 similarly shows diminishing returns: increasing Videopus from 4 to 8 rounds improves accuracy, but only 1.9% of samples require forced answers at 8 rounds. Increasing the round cap from 8 to 12 raises token consumption from 77.3K to 88.3K but reduces accuracy from 77.7% to 76.5%. This indicates that additional inspections on difficult samples do not consistently improve evidence quality or answer correctness.

AuthorsOpenReview ↗

Response to Reviewer cUbe [2/3]

We further measure end-to-end wall-clock latency under identical GPT-5 API settings and runtime conditions, with 32 concurrent QA tasks to account for network and service contention. Max. Concurrency denotes parallel Executors per batch, rather than concurrent QA tasks.

LVBench-AblMax. ConcurrencyRounds↓Avg. Latency (s)↓
VideoSeek15.61,074.45
Videopus42.7542.2
Videopus82.7476.4

Executor concurrency is an execution setting constrained by the API gateway. Since global evidence updates occur only after batch completion, increasing concurrency reduces scheduling delays without changing the planned inspections. The results show that parallel execution and fewer sequential global rounds reduce actual latency. Together with the budget analysis, Videopus maintains a favorable accuracy-cost-latency trade-off across the evaluated settings, even when VideoSeek receives substantially more inference rounds.

W3&Q2. Hyperparameter selection and robustness.

Videopus hyperparameters are designed to balance temporal coverage and spatial detail under constrained visual budgets. Global sampling uses low-resolution 3×23\times2 grids for broad temporal coverage at low cost. Local inspections share a 16-visual-unit cap: [Coverage] packs up to 64 frames into 2×22\times2 grids to locate brief events, while [Detail] uses up to 16 individual high-resolution frames within 16 seconds to preserve fine details without overly sparse sampling. These complementary configurations support targeted transitions from event localization to detail verification.

Default settings were first determined by these design principles, followed by sensitivity analysis on the fixed LVBench ablation subset. We did not perform systematic hyperparameter search on a separate development set or tune settings for individual benchmarks. Beyond the return-budget and round-cap ablations in Table 6, we examine global sampling, local visual budgets, resolution, [Detail] duration, and batch capacity. Max Tasks / Batch controls planning capacity, distinct from the concurrency limited by the API gateway.

LVBench-AblGlobal Grids[Coverage] Grids / Range[Detail] Frames / RangeLocal Resolution (Long / Short)Max [Detail] Duration (s)Max Tasks / BatchACC↑Tokens (K)↓Rounds↓
Default6416161024 / 512161277.777.32.71
Fewer global grids3216161024 / 512161273.589.13.04
More global grids9616161024 / 512161278.487.02.64
Fewer local units64881024 / 512161274.575.22.89
More local units6424241024 / 512161278.490.32.75
Lower local resolution641616768 / 384161275.882.82.73
Higher local resolution6416161280 / 640161277.7119.82.71
Shorter [Detail] interval6416161024 / 51281276.188.92.87
Longer [Detail] interval6416161024 / 512321277.181.92.79
Fewer tasks per batch6416161024 / 51216474.283.32.83
Fewer tasks per batch6416161024 / 51216875.888.02.83
More tasks per batch6416161024 / 512161675.583.82.81

The results demonstrate reasonable robustness to hyperparameter variations, with noticeable accuracy drops mainly when global sampling, local visual budgets, or batch capacity are substantially reduced. Increasing global grids or local units yields only modest gains at higher token costs, while higher resolution provides no further improvement. [Detail] duration and batch capacity also exhibit non-monotonic effects, reflecting the trade-off between evidence coverage and redundant inspections. Overall, the default configuration balances accuracy and token consumption, while larger visual budgets provide limited additional accuracy gains at higher costs.

AuthorsOpenReview ↗

Response to Reviewer cUbe [3/3]

Q1. Separating published results from controlled reruns.

Following common evaluation practice, our main tables primarily use published results to preserve each method's selected backbone and inference settings, avoiding potential degradation from differences in reproduction environments. For example, the model-combination ablation section in the VideoARM's paper shows that its chosen o3 + GPT-4o configuration outperforms GPT-5. Thus, the main comparisons reflect each method's original configuration rather than fully matched backbone and inference settings.

We further conduct controlled comparisons of ParaVT, VideoSeek, and VideoARM on full LVBench using the same GPT-5 backbone and evaluation protocol while retaining their original agent mechanisms. ASR is disabled for VideoARM to ensure consistent input modalities. All results in this table are from controlled GPT-5 reruns.

LVBench-FulVenueBackboneAvg. ACC↑ER↑EU↑Rea↑KIR↑Sum↑TG↑
ParaVTNeurIPS'26GPT-562.258.665.560.760.167.274.5
VideoSeekCVPR'26GPT-565.763.766.863.264.679.380.5
VideoARMCVPR'26GPT-570.468.768.668.777.075.976.8
Videopus—GPT-575.673.176.069.281.182.880.9

The results show that Videopus maintains higher accuracy under a shared backbone and evaluation protocol, indicating that its gains are not solely due to backbone selection but also reflect the contribution of its agent framework.

Q4. Failure detection and recovery analysis.

Videopus does not treat Executor findings as definitive. Multimodal Evidence Aggregation returns structured findings, reliability information, and selected source visuals, enabling the Orchestrator to identify missing or conflicting evidence and plan targeted reinspection.

For sparse-sampling misses, reported evidence gaps can trigger denser temporal inspection through [Coverage], followed by [Detail] for fine-grained verification. Incorrect or conflicting findings can be checked against source visuals or resolved through cross-interval comparison. When returned visuals are insufficient, the Orchestrator can revisit the relevant interval. Figures 6-7 illustrate cross-interval verification, while Figure 8 demonstrates targeted recovery from insufficient visual detail. However, completely unobserved events or undetected errors may remain unresolved.

Table 4 shows that removing source visuals reduces accuracy, while Table 6 shows that insufficient visual return increases reinspection and lowers accuracy. We further use GPT-5 to analyze full LVBench inference trajectories, identifying issues detected by the Orchestrator during inference and whether subsequent inspections successfully resolved them.

LVBench-FulDetectedRecoveredRecovered Rate↑
Missed events in sampling1358462.2%
Incorrect / conflicting findings282175.0%
Omitted visual evidence201575.0%
Overall18312065.6%

Overall, Videopus recovers 65.6% of detected issues. Conflicting findings and insufficient returned visuals are more readily resolved through source verification and targeted reinspection, while sampling-related gaps remain harder to recover without reliable temporal localization. These results demonstrate the effectiveness of evidence-guided recovery while revealing its limitations for missed temporal evidence. We provide a dedicated failure analysis in Appendix I.2 and four illustrative cases in Appendix I.3 (Figures 9-12).

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

We sincerely thank all reviewers (cUbe, tP9H, ygHS, and 9diz) for their thoughtful and constructive feedback, which has helped us strengthen our work. We are encouraged that the reviewers recognize the strong performance across six video benchmarks (cUbe, tP9H, ygHS, 9diz), the improvements in accuracy and inference efficiency (cUbe, tP9H, ygHS, 9diz), the motivation for parallel inspection and multimodal evidence aggregation (cUbe, ygHS), and the clear presentation of our framework (cUbe, tP9H).

In response to the reviewers' valuable comments, we have carefully addressed each concern in our individual responses and summarize the main revisions below:

  • We added controlled comparisons with ParaVT, VideoSeek, and VideoARM under the same GPT-5 backbone, no-tool GPT-5 evaluations across six benchmarks, matched-token-budget comparisons, and extensions to VideoSeek with batching and visual return. We also clarified baseline configurations and evaluation protocols to establish the contributions of Videopus beyond backbone differences and existing agent designs.
  • We expanded component ablations and statistical analyses to full LVBench with paired confidence intervals, and evaluated adaptive inspection modes against forced and random policies. Additional reliability-report ablations and analyses of batch planning versus concurrent execution further verify the contributions of Parallel Task Spawn and Multimodal Evidence Aggregation.
  • We extended cross-backbone evaluations to additional benchmarks and conducted sensitivity analyses of global sampling, local visual budgets, resolution, detail duration, and batch capacity. These results support the generalization and robustness of Videopus across different models, datasets, and inference configurations.
  • We provided detailed accuracy-cost-latency analyses, including larger inference budgets, end-to-end latency, and breakdowns of model calls, text/image tokens, and retries. We also added systematic failure recovery statistics and qualitative inference trajectories, illustrating how Videopus gathers complementary evidence, resolves conflicting findings, and handles missing observations.

We thank the reviewers again for their valuable suggestions. We have revised the manuscript accordingly, updating the main text, experimental results, and appendix to incorporate these improvements and provide clearer explanations of our contributions.

We appreciate the reviewers' time and consideration in reviewing our responses. We hope that the additional experiments and clarifications have addressed the concerns raised, and we welcome any further questions or discussion.

Best Regards,

All Authors of Submission #5335

Videopus introduces an Orchestrator-Executor architecture for long-horizon video understanding that decouples concurrent, interval-specific perception from visually grounded global reasoning. By coupling Parallel Task Spawn with Multimodal Evidence Aggregation, it mitigates the sequential decision overhead and visual information loss common in serial, text-mediated video agents.

Parallel Task Spawn enables the Orchestrator to dispatch batches of complementary, objective-specific inspections within a single reasoning round. Conditioned on a low-resolution grid overview and the pre-round evidence state, each subtask independently specifies its temporal interval, objective, and mode. Coverage mode packs densely sampled frames into grids to localize events, while detail mode supplies individual high-resolution frames to inspect fine visual cues, with both modes combinable within one batch. Executors run concurrently in isolated temporary contexts prior to joint assessment, reducing sequential global decisions and repeated context processing.

Multimodal Evidence Aggregation allows the Orchestrator to directly verify and synthesize evidence across independently executed intervals. Each Executor returns structured findings, visibility and sampling reliability reports, and a small set of selected source frames or grids. Retaining task objectives and intervals enables the Orchestrator to verify local claims against source visuals, establish entity correspondence across distant segments, and detect missing or conflicting evidence while keeping the persistent context compact. This visually inspectable state then guides targeted refinement in subsequent rounds. Together, these mechanisms establish an evidence-grounded workflow that unifies parallel local exploration with visually verifiable cross-interval reasoning.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.