- Rating
- 3(+2)
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper studies Moment-level Information Need Anticipation (MINA): predicting viewers' actual information needs from preceding video context, conditioned on known timestamps of real viewer questions. The authors construct MINA-Data from 872 Bilibili videos and 12,510 real viewer questions and evaluate predictions using an asymmetric information-coverage criterion. To address the limited performance of direct question generation, they propose MINA-GABT, which first predicts a structured intermediate state comprising an observed cue, a gap-formation process, and the resulting information gap, then realizes that state as a natural-language question. A taxonomy induced from training-set viewer behavior guides prediction. The method achieves 21.20% Hit@5 on test_ID, compared with 15.53% for the best Direct Forecasting baseline, and retains its performance advantage on a held-out content category.
Critical strengths
-
The dataset is constructed from real viewer behavior. The authors collect 12,510 timestamped questions from Bilibili viewers to build MINA-Data, which evaluates prediction of actual viewer questions given a known question timestamp and the preceding video context.
-
The authors propose and validate an effective prediction method. MINA-GABT combines structured information-gap prediction with a taxonomy induced from viewer behavior to guide question generation. It outperforms multiple Direct Forecasting baselines on the proposed benchmark, and ablations show that both main components improve performance.
Critical weaknesses
-
The research value of the benchmark has not been sufficiently justified. The paper is motivated by proactive video assistants that anticipate viewers' information needs and prepare answers in advance, yet MINA evaluation assumes that the timestamps of real viewer questions are known. Such timestamps are not available in advance in practice, and the paper does not validate how the current task can be applied to real proactive-assistance scenarios. In addition, MINA uses a single real viewer question as the prediction target, although different viewers may have entirely different information needs given the same video context. The paper does not analyze the distribution or agreement of real viewer questions, making it difficult to determine whether low Hit@K primarily reflects limitations in model prediction or substantial uncertainty in the target itself. The experiments therefore do not adequately establish a clear research value and practical significance for the capability measured by MINA.
-
The distinctive contributions of the proposed method have not been sufficiently validated. MINA-GABT guides question prediction with a structured gap-formation state and a taxonomy induced from viewer behavior. However, the experiments do not rule out the possibility that the gains mainly come from more detailed reasoning guidance rather than the specific proposed gap-formation mechanism. Although the authors compare Free-form Need and Intermediate S, they do not examine whether a general free-form method with sufficiently strong reasoning guidance can achieve similar results. The paper also does not compare the behavior-derived taxonomy with generic or manually constructed question patterns. The results establish that MINA-GABT outperforms the selected baselines, but do not adequately isolate the distinctive contributions of its two core designs relative to general prompt-based reasoning strategies.
Questions for authors
-
The paper is motivated by proactive video assistants that anticipate viewers' information needs and prepare answers in advance, but MINA evaluation depends on known timestamps of real viewer questions. How would a practical system use MINA without knowing these timestamps? If historical viewer questions are used to determine when to predict, would it be more reasonable to use those questions directly? If the system predicts continuously or uses an additional timing detector, what are the resulting prediction performance, computational cost, and practical benefit? Could the authors explain and provide evidence for how the current conditional prediction task effectively supports the claimed application?
-
MINA uses a single real viewer question as the prediction target, although different viewers may have entirely different information needs given the same context. Have the authors analyzed the agreement or distribution of real viewer questions at the same or nearby video timestamps? If these questions are highly diverse, could low Hit@K mainly reflect uncertainty in individual information needs rather than a lack of behavioral prediction capability? Could the authors provide statistics to clarify the learnability and practical value of the predictive capability measured by MINA?
-
The current experiments do not rule out the possibility that the gains mainly arise because structured prompts encourage more thorough reasoning rather than because of the specific gap-formation mechanism. Could the authors provide evidence that distinguishes these explanations? For example, under a comparable reasoning budget, could Free-form Need Forecasting with more explicit reasoning guidance achieve similar performance? Does the taxonomy induced from real viewer behavior provide an additional advantage over generic question patterns?
-
In the human data audit in Appendix A, did reviewers see only the same 10-second video window available to the prediction models? Being able to interpret a known question from video context does not establish that the cues that elicited it fall within that window. Have the authors verified the relationship between real viewer questions and the observation window? If a question depends on earlier video content, could the benchmark attribute failures caused by missing context to the difficulty of predicting viewer behavior?
-
The asymmetric information-coverage criterion allows a broader predicted question to cover a viewer's specific information need, but such predictions may require preparing more additional information. Appendix A combines specific questions and legitimately broader questions in the reported 95.5%. Could the authors report their proportions separately and analyze the effect of broad predictions on Hit@K? In addition, the reported agreement between the LLM judge and human judgments is 98.2%, but misses may account for most samples. Could the authors report precision and recall for hits to validate the reliability of the core evaluation results?
9 comments
Response to Reviewer h8LR (2/7)
[Response to Q1 continued]
Cost and Practical Benefit of MINA-GABT
We further estimated the inference cost of MINA-GABT using the official API pricing of Gemini and OpenAI, without applying cached-input discounts (i.e., all input and output tokens are charged at standard non-cached rates). The average cost is approximately \4.50 per video hour under a 10-second prediction interval. Importantly, such anticipation is precomputed once for the video and reused for all viewers.
To put this cost into perspective, we conducted an illustrative platform-scale deployment analysis using publicly available YouTube statistics. YouTube has reported an upload rate of approximately 720,000 hours of video per day [5], and one billion hours of daily viewing on television devices alone [6]. Besides, video popularity typically follows a long-tailed distribution, with a small fraction of videos accounting for most viewing activity. According to [7], the most-watched 1% of YouTube videos account for 91% of total viewing time. Consider an illustrative scenario in which popular videos have approximately the same average duration as the overall video population. Preprocessing only the most-watched 1% would then cost approximately \0.59 per million covered viewing minutes**. For comparison, YouTube's advertising and subscription revenue exceeded \5.9 per million covered viewing minutes, or 0.2% of annual revenue.
We emphasize that this is an illustrative deployment scenario rather than a measurement of actual YouTube deployment costs. Nevertheless, this analysis illustrates the substantial potential for reducing preprocessing costs through selective deployment while retaining broad viewing-time coverage.
Finally, the above estimates are based on a simple uniform temporal sampling strategy and a heuristic popularity-based video selection strategy, without deployment-specific optimizations. There are several additional opportunities to reduce computational costs, including leveraging cached tokens for repeated inputs, deploying self-hosted models instead of relying on commercial APIs, and distilling the prediction capabilities into smaller, more efficient models. Furthermore, prior research on proactive video agents has investigated when assistance should be provided [9,10,11]. Future work could explore more efficient context-aware anticipation scheduling, rather than relying on uniform temporal sampling, to further improve the trade-off between practical benefits and computational cost.
References
[5] UK Parliament. Oral evidence: Children's TV and Video Content, HC 1338. March 10, 2026. https://committees.parliament.uk/oralevidence/17307/html
[6] Neal Mohan. From the YouTube CEO: Our Big Bets for 2025. YouTube Official Blog, 2025. https://blog.youtube/inside-youtube/our-big-bets-for-2025
[7] Omdia. YouTube Reaches 29 Billion Videos as Music and Shorts Dominate Viewing. January 28, 2026. https://omdia.tech.informa.com/pr/2026/jan/youtube-reaches-29-billion-videos-as-music-and-shorts-dominate-viewing
[8] Alphabet Inc. 2025 Q4 Earnings Call. February 4, 2026. https://abc.xyz/investor/events/event-details/2026/2025-Q4-Earnings-Call-2026-Dr_C033hS6/default.aspx
[9] Rui Qian, Shuangrui Ding, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Dahua Lin, Jiaqi Wang: Dispider: Enabling Video LLMs with Active Real-Time Interaction via Disentangled Perception, Decision, and Reaction. CVPR 2025: 24045-24055
[10] Yueqian Wang, Songxiang Liu, Disong Wang, Nuo Xu, Guanglu Wan, Huishuai Zhang, Dongyan Zhao: MMDuet2: Enhancing Proactive Interaction of Video MLLMs with Multi-Turn Reinforcement Learning. ICLR 2026
[11] Gong et al. Vinci2: Providing Proactive Assistance in Continuous Egocentric Videos. ECCV, 2026.
Response to Reviewer h8LR (3/7)
Response to Q2 (Using a single question; Distribution of questions)
We will address your sub-questions one by one.
Distribution and agreement of real viewer questions.
Using a 5-second temporal neighborhood, we find that 47.81% of questions have at least one other real viewer question nearby. Among multi-question windows, 89.32% contain at least two distinct information needs, indicating substantial diversity in real viewer questions under the same or similar video contexts. Meanwhile, 15.35% of nearby question pairs express the same information need, and 40.10% of windows with at least three questions exhibit repeated information needs. These results reveal recurring structures in viewer questions and show that different viewers may express the same information need under the same or similar contexts. Furthermore, 31.30% of windows with at least three questions exhibit both repeated and distinct information needs, indicating that diversity and recurring structure frequently coexist.
Regarding the possibility that there may be different information needs given the same video context.
The diversity of information needs exists indeed. As 47.81% of questions have other real questions within a 5-second neighborhood, and 89.32% of multi-question windows contain at least two distinct information needs, it confirms that question diversity within a certain video context is an important and common characteristic of the MINA task. However, this does not imply that using a single observed viewer question as the ground truth is an inappropriate task design.
- A real viewer question event can be viewed as a sample from the distribution of viewer information needs. Our dataset may not be able to recover the complete question distribution under a specific video context, as only limited questions are sampled under the context. Nevertheless, by collecting a large number of real events across diverse video contexts, the dataset provides samples from the overall distribution of real viewer questions. Metrics such as Hit@K can therefore estimate a method's expected coverage under this distribution and evaluate the relative predictive capabilities of different methods.
- Similar principles have been widely adopted in research and practical applications. For example, in session-based recommendation or sequential recommendation, datasets typically record only the next item actually selected by a user, without observing the other items that the user might also have selected under the same state [1,2,3,6]. Moreover, metrics such as Hit Rate and MRR are well suited to settings where only a single behavioral event is observed, even though multiple alternative behaviors may be possible under the same context. These metrics are also widely used in sequential recommendation.
[Response to Q2 continued on next comment]
References
[1] Wang-Cheng Kang, Julian McAuley (2018). Self-Attentive Sequential Recommendation. In Proceedings of IEEE International Conference on Data Mining (ICDM'18)
[2] Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, Domonkos Tikk: Session-based Recommendations with Recurrent Neural Networks. ICLR (Poster) 2016
[3] Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, Peng Jiang: BERT4Rec: Sequential Recommendation with Bidirectional Encoder Representations from Transformer. CIKM 2019: 1441-1450
[6] Malte Ludewig, Noemi Mauro, Sara Latifi, Dietmar Jannach: Empirical analysis of session-based recommendation algorithms. User Model. User Adapt. Interact. 31(1): 149-181 (2021)
Response to Reviewer h8LR (1/7)
We would like to thank the reviewer for the detailed and constructive feedback. We hope the response below addresses your concerns.
Response to Q1 (Prediction without Timestamps)
We thank you for raising this insightful question. We agree that the reliance on video timestamps is an important factor affecting the practical applicability of MINA-GABT.
Anticipation without Oracle Timestamps
To investigate this issue, we conducted additional experiments under two types of temporal perturbations. For N∈3,5,10, we either randomly shifted the input video window by an offset uniformly sampled from [−N,+N] seconds (Jitter), or shifted the window backward such that its ending timestamp was rounded down to the nearest multiple of N seconds (Floor). The latter setting evaluates the feasibility of anticipating information needs at fixed temporal intervals without access to the exact timestamps of viewer questions.
The results demonstrate that MINA-GABT is robust to moderate variations in the input timestamp. We hypothesize two possible reasons. First, video semantics often exhibit temporal continuity and redundancy, properties that have been extensively exploited in video understanding and representation learning [1,2,3]. Second, the recorded timestamps may themselves be imprecise and may not always align exactly with the video moments that give rise to the information needs [4]. We leave the systematic investigation of these mechanisms to future work.
More importantly, MINA-GABT also exhibits limited performance degradation under simulated fixed-interval anticipation. Even with a 10-second interval, Hit@5 decreases by only 1.32 percentage points. This supports a straightforward deployment strategy in which information needs are anticipated periodically rather than at known question timestamps. For prerecorded videos, a 10-second interval requires only six predictions per minute of video. Crucially, predictions can be precomputed once and reused across subsequent viewing sessions, rather than requiring separate inference for each viewer.
[Response to Q1 continued on next comment]
References
[1] Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, Huisheng Wang, Serge J. Belongie, Yin Cui: Spatiotemporal Contrastive Video Representation Learning. CVPR 2021: 6964-6974
[2] Zhan Tong, Yibing Song, Jue Wang, Limin Wang: VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training. NeurIPS 2022
[3] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, Kaiming He: SlowFast Networks for Video Recognition. ICCV 2019: 6201-6210
[4] Bo Yang, Oskar Lindwall, Christian Licoppe, Coherence through sedimentation: Temporal organization and emergent interaction in Danmaku, Discourse, Context & Media, Volume 72, 2026.