acceptodds
Under review as a conference paper at ICLR 2027

AURA: Always-On Understanding and Real-Time Assistance via Video Streams

Abstract

Video Large Language Models (VideoLLMs) have achieved strong performance on many video understanding tasks, but most existing systems remain offline and are not well-suited for live video streams that require continuous observation and timely response. Recent streaming VideoLLMs have made progress, yet current approaches often rely on decoupled trigger-response pipelines or are limited to captioning-style narration, reducing their effectiveness for open-ended question answering and long-horizon interaction. We propose AURA (Always-On Understanding and Real-Time Assistance), an end-to-end streaming visual interaction framework that enables a unified VideoLLM to continuously process video streams and support both real-time question answering and proactive responses. AURA integrates context management, data construction, training objectives, and deployment optimization for stable long-horizon streaming interaction. It achieves state-of-the-art performance on streaming benchmarks and supports a real-time demo system with ASR and TTS running at 2 FPS on two 80 GB GPUs.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Streaming video understanding, multimodal large language model, interaction model

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 3 reviews, above 85% of the others.
Reviewer_W75LOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces AURA, a unified streaming VideoLLM framework built on Qwen3-VL-8B. AURA continuously processes video chunks, decides between silence and response, and supports Real-Time QA, Proactive QA, and Multi-Response QA. Its main components are a dual sliding window over recent video and textual QA history, a synthetic streaming-data pipeline, a Silent-Speech Balanced Loss, and a vLLM-based inference system with batched context truncation and KV-cache reuse. The authors report 73.1% on StreamingBench, 65.3% on OVO-Bench, and 25.4% on OmniMMI, plus offline-benchmark preservation and a component-wise latency analysis.

Critical strengths

The paper addresses an important and well-motivated gap between offline VideoLLMs and live, asynchronous visual assistance. Coordinating response timing, silence, memory, data construction, and serving is substantially more useful than treating streaming as ordinary long-video inference, and the unified design gives the work a coherent systems contribution.

Critical weaknesses

The paper’s “long-horizon” and “always-on understanding” claims are not adequately established by the proposed memory mechanism or experiments. The default system retains only the most recent 30 seconds of visual content; older visual evidence disappears unless a previous response already verbalized it. There is no sensitivity analysis over the video and QA window sizes, no accuracy analysis conditioned on evidence age, and no test of questions about events observed but never previously queried. Multi-Response QA, a central claimed capability, is also illustrated but not evaluated with a separate accuracy, timing, or consistency metric. As written, the evidence supports stable operation over an unbounded stream, but not necessarily long-horizon visual understanding.

The experimental protocol does not yet support the strength of the state-of-the-art and real-time claims. The paper mixes official baseline results with newly evaluated baselines, and explicitly disables MiniCPM-o-4.5’s native full-duplex streaming mode while using it as a headline comparison. The 115K streaming training samples are generated and verified entirely by unspecified MLLMs, with no human error-rate estimate or train/evaluation video-overlap audit.

Questions for authors
  1. Please report accuracy as a function of evidence age and ablations over the 30-second video window and QA-history length. Can AURA answer questions about an event from several minutes ago if that event was not previously queried?

  2. Can you provide a quantitative evaluation specifically for Multi-Response QA, including response timing, duplicate responses, missed updates, and answer consistency over time?

  3. Why was MiniCPM-o-4.5’s full-duplex mode disabled, and what are its results under its native streaming mode?

4 comments

Critical Weaknesses

W1. The paper’s “long-horizon” and “always-on understanding” claims are not adequately established by the proposed memory mechanism or experiments. The default system retains only the most recent 30 seconds of visual content; older visual evidence disappears unless a previous response already verbalized it. There is no sensitivity analysis over the video and QA window sizes, no accuracy analysis conditioned on evidence age, and no test of questions about events observed but never previously queried. Multi-Response QA, a central claimed capability, is also illustrated but not evaluated with a separate accuracy, timing, or consistency metric. As written, the evidence supports stable operation over an unbounded stream, but not necessarily long-horizon visual understanding.

We add four evaluations for this concern: a two-hour continuous run, a video-window and QA-history test, an evidence-age test, and Multi-Response QA. In a 7,200-second Inf-Streams-Eval replay through one TCP connection, AURA crosses 446 context evictions without a post-start reset or disconnect. Changing the video window has only a small effect on the completed subset. Keeping earlier QA helps answer delayed questions after the relevant frames leave the recent-video window. The Multi-Response test adds explicit timing, duplicate, recall, and consistency metrics. Q1 and Q2 give the remaining settings and results.

W2. The experimental protocol does not yet support the strength of the state-of-the-art and real-time claims. The paper mixes official baseline results with newly evaluated baselines, and explicitly disables MiniCPM-o-4.5’s native full-duplex streaming mode while using it as a headline comparison. The 115K streaming training samples are generated and verified entirely by unspecified MLLMs, with no human error-rate estimate or train/evaluation video-overlap audit.

Sec. 6.2 distinguishes published results from our runs. Qwen3-VL-8B-Instruct and MiniCPM-o-4.5 are evaluated under the official benchmark protocol. MiniCPM-o-4.5's native full-duplex mode is a different interface, so it is not mixed into the main benchmark table. Q3 below explains this protocol mismatch. We measure wall-clock service latency separately with 20 recorded spoken queries across 10 videos. ASR and TTS share one H800 GPU, while AURA runs on a second H800 GPU. All 20 queries complete the full pipeline. From complete query-audio delivery to receipt of the first complete output PCM frame, latency is 1.45 s at p50, 1.63 s at p90, and 2.36 s at p99. Thus, 90% of the queries return the first complete PCM frame within 1.63 seconds.

Sec. 4 and Appendix A describe the data engine, which uses models from the Gemini and Doubao-Seed series for generation and verification. The 115K figure denotes training samples, not unique videos. Three annotators reviewed 1,000 samples each and judged about 95% of the answers and 92% of the response windows accurate. We also sampled train and evaluation videos at 1 FPS and matched pHash values at Hamming distance 5. We found no exact duplicate videos or reused segments; 0.2% of evaluation videos were near-duplicates. The manual audit provides direct quality estimates, while the overlap check finds no exact train-evaluation reuse.

Questions for Authors

Q1. Please report accuracy as a function of evidence age and ablations over the 30-second video window and QA-history length. Can AURA answer questions about an event from several minutes ago if that event was not previously queried?

We use the original AURA checkpoint trained with selective supervision and class balancing. Both tests are diagnostics built from OmniMMI Multi-turn Dependency Reasoning annotations rather than official OmniMMI scores. The N/M ablation replays the original question sequence for the same 94 completed episodes while changing the recent-video window and retained QA history. GPT-6.1-Sol scores all open-ended answers with one fixed semantic-match prompt.

Video window N (s)QA history MEpisode-macro accuracy
151044.5%
301047.7%
601048.8%
30047.0%

The N/M ablation measures general sensitivity across all questions in the completed episodes. Accuracy changes only slightly when N varies from 15 to 60 seconds or when M changes from 0 to 10. The evidence-age test below isolates questions about a specific event after its visual evidence becomes old.

To isolate old-event recall, we construct 166 matched pairs from the same OmniMMI source. The target question is delayed by 15 to 180 seconds after its evidence ends. In one condition, the event is not queried earlier. In the other, AURA first answers the question while the evidence is recent, and this generated QA is retained until the delayed query. The reference answer is never added to the context.

Time since the eventNo earlier questionEarlier AURA QA retainedAccuracy gainPairs
15 s22.0%26.0%+4.0 points50
30 s23.9%43.5%+19.6 points46
60 s31.3%53.1%+21.8 points32
120 s7.7%42.3%+34.6 points26
180 s0.0%41.7%+41.7 points12
Overall20.5%39.8%+19.3 points [95% CI: 12.1, 26.5]166

Retaining AURA's earlier QA improves accuracy at every tested age, and the gain increases with delay. Without earlier QA, questions about old events are much harder once their frames leave the 30-second window. The retained QA therefore helps preserve relevant information beyond that video window.

Q2. Can you provide a quantitative evaluation specifically for Multi-Response QA, including response timing, duplicate responses, missed updates, and answer consistency over time?

We construct a targeted evaluation from OmniMMI Dynamic State Grounding with 290 complete episodes and 658 required answer updates that can be scored. Each question is asked once, and each annotated state change defines a valid response window. GPT-6.1-Sol judges semantic matches with one fixed prompt. Correct updates are pooled counts, while recall and consistency are averaged per episode. An update is consistent when it receives at least one response and every response in its window matches the target.

Correct updates (pooled)Recall (episode avg.)Consistency (episode avg.)Missed updatesDuplicatesLatency p50 / p90
93 / 65814.8%14.5%56541 s / 1 s

Correctly matched updates arrive within one second at both p50 and p90, and duplicate matches are rare.

Q3. Why was MiniCPM-o-4.5’s full-duplex mode disabled, and what are its results under its native streaming mode?

AURA targets streaming video QA and does not claim native audio-video full-duplex capability. The QA tasks in our main comparison insert text questions at fixed video timestamps. MiniCPM-o-4.5's native full-duplex mode instead requires spoken input and produces spoken output, so it cannot be enabled without changing the benchmark timeline.

Converting the questions with TTS is not valid because a spoken question may last longer than the interval before the next scheduled question. A spoken answer also takes time to complete, while the benchmark may not leave enough uninterrupted video after the question for the model to finish its answer. MiniCPM-o-4.5 often remained silent or returned unrelated answers on streams longer than two minutes, so we did not use this mode in the main comparison. We therefore do not report a native-mode score on these QA benchmarks because it would not be a like-for-like result.

Dear Reviewer W75L, as the author-reviewer discussion period is coming to an end, we would like to kindly remind you that we have posted our responses and an updated manuscript. We would be very grateful if you could take a look at your convenience, and we remain happy to discuss any point further. We look forward to hearing from you. Thank you for your time and consideration.

Dear Reviewer W75L, as the author-reviewer discussion period is coming to an end, we would like to kindly remind you that we have posted our responses and an updated manuscript. We would be very grateful if you could take a look at your convenience, and we remain happy to discuss any point further. We look forward to hearing from you. Thank you for your time and consideration.

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

AURA adapts Qwen3-VL-8B-Instruct for streaming video assistance with immediate, delayed, and repeated responses. It combines bounded video history and retained QA context, synthetic training data, silence-aware supervision, and efficient cache management. Experiments on three streaming benchmarks show gains over the base model, complemented by a real-time voice-interaction demonstration.

Critical strengths
  1. Useful integrated capabilities with substantial empirical gains.

AURA supports three practical temporal QA modes and substantially improves over the base model across three streaming benchmarks. Gains in HLD and timing-aware proactive assistance support its perception and proactive-interaction capabilities

  1. Informative controlled evidence for practical design choices.

Fig. 5 compares no window, per-step eviction and batched eviction, supporting improved cache efficiency, while the combined training objective helps the model respond when needed instead of remaining silent in Table 5.

Critical weaknesses
  1. The evidence for sustained real-time operation remains limited.

The paper does not specify a limit on the number of responses or tokens within each QA group. And latency is estimated from components rather than measured end-to-end wall-clock time

  1. The training mechanism needs clearer evidence, and the data pipeline is not fully documented.

The loss ablation changes both supervision masks and silent-turn weights in Table 8, so it does not isolate which component reduces hallucinations. Judge model versions, actual prompts and annotation-quality results are also not reported, limiting data reproducibility.

Questions for authors
  1. What limits apply to QA-group responses/tokens and request lifetime?

  2. Does any existing persistent-request trace establish stable context size and task success across eviction cycles?

  3. Is hallucination reduction attributed specifically to selective supervision or observed only for the full system?

4 comments

Critical Weaknesses

W1. The evidence for sustained real-time operation remains limited. The paper does not specify a limit on the number of responses or tokens within each QA group. And latency is estimated from components rather than measured end-to-end wall-clock time.

Each response is capped at 128 generated tokens. AURA normally keeps 30 one-second video rounds and may grow to 45 before removing the oldest 15 together. Outside this window, it retains at most 10 past QA groups. These limits bound the active context and per-turn model workload. They do not impose an additional cap on session duration or cumulative response count, and a longer session does not by itself make inference progressively slower. We verify continuous operation and end-to-end latency directly below.

First, AURA processed a continuous 7,200-second replay of an NBA video from Inf-Streams-Eval through one TCP connection. It crossed 446 context evictions without a post-start reset or disconnect. The connection remained active while old context was repeatedly removed over two hours.

Second, we measured end-to-end service latency using 20 recorded spoken queries across 10 videos. ASR and TTS shared one H800 GPU, while AURA ran on a second H800 GPU. All 20 queries completed the full pipeline and were matched to their audio responses. We measure from complete query-audio delivery to receipt of the first complete output PCM frame. Latency was 1.45 s at p50, 1.63 s at p90, and 2.36 s at p99, with a maximum of 2.49 s. The 1,500-second run used one connection and crossed 93 context evictions without a reset, disconnect, missing turn, duplicate turn, or trace error. Thus, 90% of the queries returned the first complete PCM frame within 1.63 seconds. These are direct wall-clock service measurements rather than sums of component latencies.

W2. The training mechanism needs clearer evidence, and the data pipeline is not fully documented. The loss ablation changes both supervision masks and silent-turn weights in Table 8, so it does not isolate which component reduces hallucinations. Judge model versions, actual prompts and annotation-quality results are also not reported, limiting data reproducibility.

Sec. 5.1 defines the two training choices. We supervise every silent message but only the last non-silent answer. We also divide the weight of each silent message by the number of silent messages in that sample. The paper evaluates the loss ablation on OmniMMI by removing both choices together. We therefore keep the same benchmark and test each choice separately.

For an aligned comparison, we reran inference from all three checkpoints and scored their outputs with the same GPT-6.1-Sol judge. We use the official OmniMMI judge prompt without modification. It asks whether a prediction meaningfully matches the reference. It accepts synonyms and paraphrases and returns a yes-or-no decision with a score from 0 to 5. Proactive Alerting is computed directly from the annotated response window and does not use the judge. The original evaluation implementation used GPT-4o-2024-08-06. Because that model is no longer available through our evaluation endpoint, the table below reports aligned GPT-6.1-Sol reruns for all three checkpoints. The first row is the original AURA checkpoint.

Selective supervisionClass balancingState GroundingAction PredictionMulti-turn DependencySpeaker IdentificationProactive AlertingOverall (5-task macro)
✓✓23.327.08.034.538.526.3
—✓22.727.09.735.526.524.3
✓—24.728.510.337.50.020.2

With class balancing fixed, selective supervision improves Overall and Proactive Alerting. With selective supervision fixed, class balancing also raises Overall, mainly through Proactive Alerting. This controlled OmniMMI comparison isolates the two contributions and shows that their combination gives the best Overall result.

Sec. 4 and Fig. 3 present the full data pipeline. Appendix A.1 details QA synthesis, and Appendix A.2 covers refinement. Appendix A.3 gives the quality checks, and Appendix A.4 reports the data distribution. Generation and verification use models from the Gemini and Doubao-Seed series. The 115K figure denotes training samples, not unique videos. Three annotators also reviewed 1,000 samples each. About 95% of the answers and 92% of the response windows were judged accurate. The manual audit therefore finds high answer and timing accuracy in the sampled data.

Questions for Authors

Q1. What limits apply to QA-group responses/tokens and request lifetime?

Each response is capped at 128 generated tokens, and each video round covers one second. The system normally keeps the latest 30 video rounds. It may grow to 45 rounds, then removes the oldest 15 at once. Outside this recent-video window, it keeps at most 10 past QA groups. After an old QA group is compressed, it keeps at most four rounds, including the first question. These limits bound retained memory, not the total number of responses in a session.

Each generation turn is a separate model request with its own request ID. The TCP connection and interaction history remain active across turns, so the application session can continue after one generation request ends. There is no additional application-level cap on session duration or cumulative response count. In our continuous test, the same session lasted 7,200 replay seconds and crossed 446 evictions without a post-start reset or disconnect.

Q2. Does any existing persistent-request trace establish stable context size and task success across eviction cycles?

We use complementary evaluations for system stability and semantic accuracy because a continuous trace does not measure semantic accuracy at every eviction. In the two-hour Inf-Streams-Eval replay, the recent context stayed within 44 rounds and the retained history stayed within 10 QA groups. The session crossed 446 evictions without a post-start reset or disconnect. In the 1,500-second end-to-end trace, all 20 spoken queries completed through one connection across 93 evictions. No turn was missing or duplicated.

To verify semantic task success under bounded-context inference, we conduct two new experiments with the original AURA checkpoint. Both use OmniMMI Multi-turn Dependency Reasoning annotations, but they are diagnostic evaluations rather than official OmniMMI scores. GPT-6.1-Sol scores all open-ended answers with one fixed semantic-match prompt.

First, the N/M ablation replays the original question sequence for the same 94 completed episodes. We change the recent-video window N and the number of retained QA groups M while keeping the questions and their order fixed.

Video window N (s)QA history MEpisode-macro accuracy
151044.5%
301047.7%
601048.8%
30047.0%

The N/M ablation measures general sensitivity across all questions in the completed episodes. Accuracy changes only slightly when N varies from 15 to 60 seconds or when M changes from 0 to 10. The evidence-age test below isolates questions about a specific event after its visual evidence becomes old.

Second, to isolate old-event recall, we construct 166 matched pairs from the same OmniMMI source. The target question is delayed by 15 to 180 seconds after its evidence ends. In one condition, the event is not queried earlier. In the other, AURA first answers the question while the evidence is recent, and this generated QA is retained until the delayed query. The reference answer is never added to the context.

Time since the eventNo earlier questionEarlier AURA QA retainedAccuracy gainPairs
15 s22.0%26.0%+4.0 points50
30 s23.9%43.5%+19.6 points46
60 s31.3%53.1%+21.8 points32
120 s7.7%42.3%+34.6 points26
180 s0.0%41.7%+41.7 points12
Overall20.5%39.8%+19.3 points [95% CI: 12.1, 26.5]166

Retaining AURA's earlier QA improves accuracy at every tested age, and the gain increases with delay. Without earlier QA, questions about old events are much harder once their frames leave the 30-second window. The retained QA therefore helps preserve relevant information beyond that video window.

Q3. Is hallucination reduction attributed specifically to selective supervision or observed only for the full system?

The new OmniMMI ablation separates selective supervision from class balancing. With class balancing fixed, selective supervision improves Overall and Proactive Alerting. With selective supervision fixed, class balancing also raises Overall, mainly through Proactive Alerting. This shows that selective supervision contributes independently rather than only through the full system.

Dear Reviewer JrsA, as the author-reviewer discussion period is coming to an end, we would like to kindly remind you that we have posted our responses and an updated manuscript. We would be very grateful if you could take a look at your convenience, and we remain happy to discuss any point further. We look forward to hearing from you. Thank you for your time and consideration.

Dear Reviewer JrsA, as the author-reviewer discussion period is coming to an end, we would like to kindly remind you that we have posted our responses and an updated manuscript. We would be very grateful if you could take a look at your convenience, and we remain happy to discuss any point further. We look forward to hearing from you. Thank you for your time and consideration.

Reviewer_dqm1OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a new streaming video understanding framework. It utilizes chunk-wise conversational format, dual sliding-window strategy, QA categorization, and a new data engine to boost the real-time understanding ability of VideoLLMs, achieving good results on multiple benchmarks.

Critical strengths
  1. This work indeed compensates for the shortcomings of existing VideoLLM framework, offering a good example for applications in this field.

  2. Compared to existing methods, the author provides stronger test results on several streaming video understanding benchmarks.

  3. The framework proposed by the authors indeed has a relatively complete design, and the processing methods such as Quality Verification introduced in the data engine are also comprehensive and reasonable.

Critical weaknesses
  1. I think that overall it is a complete system, but it seems that the authors’ contribution lies more in engineering system coordination and assembly, and the proposed methods seem to bring limited insight. For example, chunkwise dialogue processing and the Sliding-Window strategy have already had similar processing in works such as Streamo [1] and StreamingVLM [2]. It seems that there are more incremental combination instead of essential changes.

  2. Although the authors point out the necessity of Class-Balancing Loss with detailed motivation, there seems to be no experiment proving that it is better than the focal loss solution applied by Streamo.

  3. I think it could be more convincing if there were test results on longer continuous streams (e.g., Inf-Streams-Eval).

  4. It seems that the comparison with Qwen2.5-VL-based Streamo seems not fair enough. It is recommended to compare under the same base model.

[1] Streaming Video Instruction Tuning

[2] StreamingVLM: Real-Time Understanding for Infinite Video Streams

Questions for authors

Please refer to the weaknesses above. While it’s a good work overall, I think it still needs more detailed refinement.

5 comments

Critical Weaknesses

W1. I think that overall it is a complete system, but it seems that the authors’ contribution lies more in engineering system coordination and assembly, and the proposed methods seem to bring limited insight. For example, chunkwise dialogue processing and the Sliding-Window strategy have already had similar processing in works such as Streamo [1] and StreamingVLM [2]. It seems that there are more incremental combination instead of essential changes.

Streamo reformulates video as chunk-wise dialogue. Its public inference path keeps one FIFO dialogue window and can add one global question back after truncation, but it does not maintain a separate history of multiple QA groups. StreamingVLM keeps attention-sink tokens, a short recent-vision window, and a longer recent-text window in its KV cache. Its streaming SFT uses overlapped video chunks with per-second narration targets. All four VQA benchmarks reported for StreamingVLM constrain the answer space to multiple-choice or fixed-format outputs, rather than testing persistent open-ended multi-turn streaming QA.

AURA uses chunk-wise input and sliding windows as parts of a context design for persistent open-ended QA. It keeps recent video and earlier QA groups in separate windows. Training samples are built to match the retained video and QA context used at inference. The loss supervises silent turns and the final non-silent answer supported by the retained context, while earlier non-silent answers remain as history rather than training targets. This aligns training with the context available at inference. AURA also removes old video chunks in batches so the prefix cache can be reused with less KV cache recomputation.

Beyond context management, AURA also contributes a streaming QA data engine, a balanced training objective, and an end-to-end real-time system. Together, they enable AURA to learn what to answer and when to answer, then deliver those responses with low latency.

W2. Although the authors point out the necessity of Class-Balancing Loss with detailed motivation, there seems to be no experiment proving that it is better than the focal loss solution applied by Streamo.

Applying Streamo's focal loss to AURA requires adaptation because the two systems use different label spaces. Streamo applies focal loss to three explicit state tokens named Silence, Standby, and Response. AURA uses one silent token, while a response begins with an ordinary word token. The available focal checkpoint therefore applies focal loss with γ = 2 to all tokens that AURA trains on. It is a focal alternative within AURA, rather than an exact reproduction of Streamo's three-state objective. We evaluate it at 1 FPS with the official OVO-Bench code, matching the original AURA setting.

Training lossBackwardReal-timeForwardOverall
AURA loss (paper)60.479.855.865.3
Focal-loss variant, γ = 260.377.957.165.1

The focal-loss variant slightly improves Forward, but lowers Real-time and does not improve Overall. Under the released OVO-Bench evaluation code used here, Forward Active Responding is evaluated through fixed-timestamp count and yes-or-no probes rather than free-running response triggering, so the Forward difference does not directly measure free-running silent/speak decisions.

W3. I think it could be more convincing if there were test results on longer continuous streams (e.g., Inf-Streams-Eval).

A longer test is useful. Inf-Streams-Eval was introduced in the StreamingVLM paper and contains videos averaging 2.1 hours. StreamingVLM's streaming mode is trained for dense sports commentary, and the benchmark evaluates continuous commentary rather than arbitrary-time open-ended QA. AURA instead focuses on open-ended questions and selective responses. PhoStream [ICML 2026] excludes StreamingVLM because it does not support open-ended question answering. We therefore use Inf-Streams-Eval to test long-run system stability and PhoStream to evaluate long-video QA.

First, we replay 7,200 seconds of NBA video input from Inf-Streams-Eval in one continuous session. AURA crossed 446 context evictions without a post-start reset or disconnect. The connection remained active throughout the two-hour stream. Server latency to first token was 252 ms at p50, 278 ms at p90, and 1.33 s at p99.

Second, we follow PhoStream's official online inference and scoring protocol to evaluate long-video QA. It contains 5,572 open-ended QA pairs from 578 videos, and its average video length is 13.3 minutes. Overall is the official sample-weighted score, not the unweighted mean of the three task scores. The published baselines and our AURA results are shown below.

ModelInstantBackwardForwardOverall
Gemini 3 Pro80.882.216.448.7
Doubao-Seed-1.880.577.333.456.2
Qwen3-VL-8B75.271.57.240.3
VideoLLM-online-8B24.924.70.012.3
AURA75.270.553.163.2

AURA achieves the best Overall score in the table. Its main advantage is Forward, while its Instant and Backward results remain competitive.

W4. It seems that the comparison with Qwen2.5-VL-based Streamo seems not fair enough. It is recommended to compare under the same base model.

We trained AURA with the same Qwen2.5-VL-7B backbone family and evaluated it on OVO-Bench at 1 FPS.

ModelBackboneSourceBackwardReal-timeForwardOverall
AURAQwen2.5-VL-7BNew evaluation63.276.357.765.8
StreamoQwen2.5-VL-7BPublished, 1 FPS46.166.054.855.6

With the same backbone family, AURA is higher on Backward, Real-time, Forward, and Overall. We use Streamo's published 1 FPS result because its weights are not public.

Questions for Authors

Q1. Please refer to the weaknesses above. While it’s a good work overall, I think it still needs more detailed refinement.

Thank you for the helpful feedback and for recognizing the value of the overall system.

Dear Reviewer dqm1, as the author-reviewer discussion period is coming to an end, we would like to kindly remind you that we have posted our responses and an updated manuscript. We would be very grateful if you could take a look at your convenience, and we remain happy to discuss any point further. We look forward to hearing from you. Thank you for your time and consideration.

Comments on the paper

1 comment

We thank the Area Chair and reviewers for their careful evaluation and constructive feedback. We have submitted our responses and an updated manuscript that addresses the raised questions and concerns. We appreciate your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.