Paper deep dive

Adaptive Greedy Frame Selection for Long Video Understanding

Yuning Huang, Fengqing Zhu

Year: 2026Venue: arXiv preprintArea: cs.CVType: PreprintEmbeddings: 31

Intelligence

Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 95%

Last extracted: 3/23/2026, 12:13:00 PM

Summary

The paper introduces a question-adaptive greedy frame selection method for long-video question answering, designed to optimize the trade-off between query relevance and semantic representativeness. By utilizing a modular relevance term and a facility-location coverage term, the method selects a subset of frames that avoids redundancy and coverage collapse. A lightweight text-only classifier routes queries to one of four preset strategies, demonstrating consistent accuracy gains on the MLVU benchmark compared to uniform sampling and existing baselines like AKS.

Entities (5)

MLVU · dataset · 100%DINOv2 · embedding-model · 95%Question-type classifier · component · 95%SigLIP · embedding-model · 95%AKS · baseline-method · 90%

Relation Signals (4)

Greedy frame selection method → evaluatedon → MLVU

confidence 100% · Experiments on MLVU show consistent accuracy gains

Question-type classifier → routesto → Preset Strategy

confidence 95% · a lightweight text-only question-type classifier that routes each query to its best-performing preset

Greedy frame selection method → uses → SigLIP

confidence 95% · constructs a 1 FPS candidate pool... embeds candidates in two complementary spaces (SigLIP for question relevance)

Greedy frame selection method → uses → DINOv2

confidence 95% · embeds candidates in two complementary spaces... DINOv2 for semantic similarity

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Large vision--language models (VLMs) are increasingly applied to long-video question answering, yet inference is often bottlenecked by the number of input frames and resulting visual tokens. Naive sparse sampling can miss decisive moments, while purely relevance-driven selection frequently collapses onto near-duplicate frames and sacrifices coverage of temporally distant evidence. We propose a question-adaptive greedy frame selection method that jointly optimizes query relevance and semantic representativeness under a fixed frame budget. Our approach constructs a 1~FPS candidate pool (capped at 1000) with exact timestamp alignment, embeds candidates in two complementary spaces (SigLIP for question relevance and DINOv2 for semantic similarity), and selects frames by greedily maximizing a weighted sum of a modular relevance term and a facility-location coverage term. This objective is normalized, monotone, and submodular, yielding a standard (1-1/e) greedy approximation guarantee. To account for question-dependent trade-offs between relevance and coverage, we introduce four preset strategies and a lightweight text-only question-type classifier that routes each query to its best-performing preset. Experiments on MLVU show consistent accuracy gains over uniform sampling and a strong recent baseline across frame budgets, with the largest improvements under tight budgets.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

30,433 characters extracted from source content.

Expand or collapse full text

Adaptive Greedy Frame Selection for Long Video Understanding Yuning Huang and Fengqing Zhu Purdue University Abstract. Large vision–language models (VLMs) are increasingly ap- plied to long-video question answering, yet inference is often bottlenecked by the number of input frames and resulting visual tokens. Naive sparse sampling can miss decisive moments, while purely relevance-driven selec- tion frequently collapses onto near-duplicate frames and sacrifices cov- erage of temporally distant evidence. We propose a question-adaptive greedy frame selection method that jointly optimizes query relevance and semantic representativeness under a fixed frame budget. Our ap- proach constructs a 1 FPS candidate pool (capped at 1000) with exact timestamp alignment, embeds candidates in two complementary spaces (SigLIP for question relevance and DINOv2 for semantic similarity), and selects frames by greedily maximizing a weighted sum of a modular rel- evance term and a facility-location coverage term. This objective is nor- malized, monotone, and submodular, yielding a standard (1−1/e) greedy approximation guarantee. To account for question-dependent trade-offs between relevance and coverage, we introduce four preset strategies and a lightweight text-only question-type classifier that routes each query to its best-performing preset. Experiments on MLVU show consistent ac- curacy gains over uniform sampling and a strong recent baseline across frame budgets, with the largest improvements under tight budgets. Keywords: Long video understanding· vision–language models· frame selection· submodular maximization· facility-location coverage· question- adaptive routing 1 Introduction Large vision-language models (VLMs) are increasingly applied to video question answering and multimodal reasoning, but long videos remain a bottleneck: feed- ing more frames quickly exceeds practical compute budgets and context limits, while sparse sampling can miss rare yet decisive moments. Recent work tackles long-form video understanding with memory mechanisms and sparse representa- tions (e.g., compressing dense tokens into structured memories) [13], hierarchical or adaptive video representations [15], retrieval-augmented pipelines that iter- atively search for missing evidence [10], and training-free or lightweight frame selection heuristics that improve downstream VLM accuracy [9]. In parallel, learnable selection policies and optimization-based samplers have been proposed arXiv:2603.20180v1 [cs.CV] 20 Mar 2026 2F. Author et al. to choose informative frames under a strict budget [1,14,17]. Meanwhile, long- video benchmarks such as Video-MME and LongVideoBench highlight that per- formance gains often correlate with a model’s ability to allocate its limited visual tokens to the right temporal evidence [4,16]. However, selecting a small subset of frames is not merely a relevance-ranking problem. Two failure modes are common in practice: (i) redundancy collapse, where selectors over-concentrate on visually similar frames around a salient seg- ment (wasting budget), and (i) coverage collapse, where selectors over-penalize redundancy and spread too thin, missing localized but critical events. These modes motivate two complementary desiderata: diversity to discourage repet- itive selections and coverage to ensure the chosen subset spans distinct parts of the video so that important evidence is not skipped. This perspective is closely related to classic relevance–diversity retrieval and subset selection: Max- imal Marginal Relevance (MMR) trades off relevance and novelty [2], while cov- erage/representativeness objectives are often modeled as (approximately) sub- modular functions that admit efficient greedy maximization with approximation guarantees [5, 8, 12]; diversity can also be promoted via DPP/log-determinant objectives [7, 18]. Existing long-video selectors often emphasize one side of the trade-off and can be brittle when the query requires a different balance. Crucially, the best trade-off is question dependent. For example, summary- style questions benefit from broader temporal and semantic coverage, whereas fine-grained questions about a specific event may favor relevance with mild re- dundancy control. This observation suggests that a single fixed selection strategy is suboptimal across heterogeneous question types. To address this, we propose a query-adaptive greedy frame selector that (1) optimizes a parameterized ob- jective jointly balancing relevance and semantic-space coverage (which implic- itly penalizes redundancy), and (2) selects the trade-off policy conditioned on a lightweight question-type classifier. Our design is complementary to recent adap- tive acquisition frameworks that switch information-gathering behaviors for long videos [21]. We evaluate on MLVU [20], which provides an official question taxonomy (nine categories) enabling systematic, per-type analysis of selection behavior. While using a single benchmark is not ideal, MLVU is uniquely suitable for studying question-type adaptive frame selection because it couples long-video challenges with explicit question type labels. We therefore use MLVU for in- depth ablations with four strategies—relevance-only, coverage-only, relevance- oriented, and coverage-oriented—and report per-question-type results under a fixed strategy assignment for each type; extending to other long-video bench- marks such as Video-MME, LongVideoBench, and EgoSchema is left to future work [4,11,16]. Contributions. Our main contributions are: (1) a parameterized greedy frame se- lection objective that explicitly couples diversity (anti-redundancy) and coverage (anti-collapse) under a fixed frame budget; (2) a query-adaptive mechanism that maps question types to selection strategies via a small classifier, enabling differ- ent relevance–coverage trade-offs for different intents; and (3) extensive per-type Abbreviated paper title3 evaluation and ablations on MLVU demonstrating consistent gains over single- strategy baselines under the same frame budget. 2 Related Work 2.1 Long Video Understanding Benchmarks and Question Taxonomies Long video understanding (LVU) benchmarks have rapidly evolved to stress- test vision–language models (VLMs) under long temporal horizons and tight context budgets. Among them, MLVU provides a particularly suitable testbed for studying question-adaptive frame selection because it comes with an official question/task taxonomy (nine categories) and supports per-category evaluation at scale [20]. This structured labeling directly enables our setting: learning a lightweight question-type classifier and selecting frames with different coverage– diversity trade-offs per type. We primarily evaluate on MLVU [20] for two pragmatic and scientific rea- sons. First, MLVU’s standardized categorization allows a controlled analysis of which selection strategy helps which question type, and makes it possible to re- port consistent per-type ablations without introducing confounding differences in annotation formats across benchmarks. Second, our goal in this work is to es- tablish a mechanistic understanding of frame selection for token reduction (cov- erage vs. diversity, and question-adaptive strategy switching); MLVU provides the cleanest lens for such analysis under limited experimental bandwidth. While multi-benchmark evaluation is desirable, we view it as complementary future work once the core method and its trade-offs are thoroughly characterized. We note that other LVU datasets (e.g., egocentric long-form video QA) have also been used to probe temporal reasoning, such as EgoSchema [11]. However, these benchmarks typically do not offer the same official, multi-category ques- tion taxonomy as MLVU, making them less ideal for studying type-conditioned selection policies in a unified protocol. 2.2 Token Reduction via Frame Selection for VLMs A prevalent approach to LVU is token reduction by selecting a small set of frames (or visual tokens) before passing them to a VLM. Recent methods explore different selection objectives and implementations. Adaptive Keyframe Sam- pling (AKS) formulates selection through prompt relevance and keyframe cov- erage, yielding an efficient plug-and-play module [14]. Flexible Frame Selection (FFS) learns a policy with a flexible selection operation to focus on informative frames under context constraints [1]. Orthogonally, memory-based systems (e.g., MovieChat) trade dense frame ingestion for persistent memories that support long-range reasoning [13], while hierarchical representations (e.g., VideoTree) compress long videos into structured summaries for downstream reasoning [15]. More recently, adaptive/agentic sampling paradigms (e.g., VideoBrain) learn 4F. Author et al. when and how to acquire more visual evidence, often combining retrieval-style proposals with temporal coverage mechanisms [22]. In contrast to methods that focus on a single notion of redundancy con- trol (e.g., enforcing temporal spacing) or rely on a fixed selection rule across all queries, our work emphasizes simultaneously controlling diversity and coverage with an explicit trade-off, and switching the trade-off based on question type. This design is particularly aligned with LVU question het- erogeneity: summary-style questions benefit from broad coverage, while detail- oriented questions often prefer relevance-focused selection with redundancy avoid- ance. 2.3 Coverage–Diversity Objectives and Relevance–Diversity Keyframe Sampling Our selection objective is closely related to classic subset selection formulations that balance relevance with redundancy control. Maximal Marginal Relevance (MMR) explicitly trades off query relevance with novelty to reduce redundancy in retrieved results [3]. More broadly, many diversity/coverage criteria can be ex- pressed with (approximately) submodular objectives that admit efficient greedy maximization under cardinality (token/frame budget) constraints [6]. In long- video settings, these perspectives motivate selecting frames that are both query- relevant and representative of the full candidate pool, avoiding wasted tokens on near-duplicates while still covering temporally and semantically distinct events. Recent training-free keyframe sampling methods instantiate these ideas in different ways. AKS emphasizes prompt relevance together with coverage under a fixed budget and provides a plug-and-play selector for long-video reasoning [14]. Other approaches combine relevance with explicit diversity objectives, such as log-determinant/max-volume criteria (DPP-style) to encourage non-redundant selections (e.g., AdaRD-Key) [19]. These methods highlight the importance of balancing query alignment and redundancy control, but they typically apply a single selection rule uniformly across heterogeneous question intents. In contrast, we (i) explicitly disentangle coverage from redundancy reduc- tion to mitigate the “local-collapse” failure mode (selecting many near-duplicates and missing salient events elsewhere), and (i) treat the relevance–coverage trade- off as question-dependent. Leveraging MLVU’s official question taxonomy [20], we define four interpretable strategies—relevance-only, coverage-only, relevance- oriented, and coverage-oriented—and assign a fixed strategy per question type. We validate this design with per-type results across all nine categories and sys- tematic strategy ablations, providing a controlled analysis of which selection behavior is best suited for which long-video question intent. 3 Method 3.1 Overview and problem setup Given a long video V and a question q, VLM inference is often bottlenecked by the number of input frames (and thus visual tokens). Our goal is to select a Abbreviated paper title5 subset of K frames that is (i) relevant to q and (i) representative in semantic space, so the subset avoids near-duplicates and covers temporally distant events. Concretely, we (1) construct a bounded 1 FPS candidate pool with exact timestamp alignment, (2) embed candidates in two complementary spaces (query relevance vs. semantic coverage), and (3) maximize a relevance+coverage objec- tive via a greedy algorithm with a standard approximation guarantee. 3.2 Candidate construction and timestamp alignment 1 FPS candidate seconds. For a video with average FPS f and T decoded frames, define duration in seconds as ⌊T/f⌋ and construct integer-second candidates: U =0, 1,...,⌊T/f⌋− 1. To bound compute, we cap the candidate count at 1000. If |U| > 1000, we uniformly downsample the second indices to 1000 (e.g., linspace(0,|U|−1, 1000) followed by integer casting). Positions vs. seconds (alignment invariant). Embeddings are stored in the or- der of the selected seconds. Therefore, the selector operates on positions i ∈ 1,...,N (where N ≤ 1000), which index the embedding set. A fixed mapping π(i) 7→ s i converts each position back to its corresponding second index using the same downsampling rule used during embedding extraction. Loading selected frames. Given selected seconds s i , we load frames at indices ⌊s i ·f⌋ (clipped to the valid range), ensuring exact alignment between embedding positions and the decoded frames used by the VLM. 3.3 Two embedding spaces We precompute two complementary embedding spaces over the N candidates. Query relevance space (SigLIP). For each candidate i, we compute a SigLIP visual embedding v i ∈R d s and for question q a SigLIP text embedding t∈R d s . We ℓ 2 -normalize both so v ⊤ i t is cosine similarity. Semantic representativeness space (DINOv2). For each candidate i, we compute a DINOv2 embedding d i ∈R d d and ℓ 2 -normalize it. These embeddings define a semantic similarity space used by a facility-location coverage objective. 3.4 Objective: relevance plus facility-location coverage Let the ground set of candidates be G = 1,...,N. We seek a subset S ⊆ G with |S|≤ K maximizing F(S) = αR(S) + β C(S),(1) with trade-off weights α,β ≥ 0. 6F. Author et al. Relevance term (modular). We define a non-negative relevance score r i ≜ max(v ⊤ i t, 0), (optionally with per-video normalization such as z-score across candidates fol- lowed by ReLU and max-normalization, as long as r i ≥ 0 and preserves within- video ordering). The relevance set function is modular: R(S) = X i∈S r i .(2) Coverage term (facility-location). Define the DINOv2 cosine similarity s j,i ≜ d ⊤ j d i , s j,i ∈ [−1, 1], and the facility-location coverage objective C(S) = X j∈G max b, max i∈S s j,i − b ,(3) where b≜−1 is a baseline chosen to match our implementation (initial coverage is −1) while ensuring C(∅) = 0. 3.5 Greedy maximization with efficient coverage updates We maximize F(S) under |S| ≤ K using the standard greedy algorithm. Main- tain a coverage vector c j (S)≜ max b, max i∈S s j,i , initialized as c j (∅) = b. The marginal coverage gain of adding i /∈ S is ∆ C (i| S) = N X j=1 max c j (S),s j,i − c j (S) , and the total marginal gain is ∆(i| S) = αr i + β ∆ C (i| S). After selecting i ⋆ = arg max i /∈S ∆(i | S), update c j ← max(c j ,s j,i ⋆ ) for all j. We return the selected positions sorted in temporal order (equivalently, by their mapped seconds). Complexity. Precomputing the similarity matrix is O(N 2 ). Each greedy step evaluates all remaining candidates with an O(N) coverage-gain computation, yielding O(KN 2 ) time in the straightforward implementation (highly vectoriz- able on GPU for N ≤ 1000). Abbreviated paper title7 Algorithm 1: Greedy Relevance + Facility-Location Coverage Selec- tion Input: SigLIP frame embeddings v i N i=1 (normalized), SigLIP text embedding t (normalized); DINOv2 embeddings d i N i=1 (normalized); subset size K; weights α,β Output: selected positions S (sorted) Compute relevance r i ← max(v ⊤ i t, 0) for all i; Compute similarity matrix s j,i ← d ⊤ j d i ; Initialize S ←∅, and c j ← b (=−1) for all j; for ℓ = 1 to min(K,N) do foreach candidate i /∈ S do ∆ C (i| S)← P N j=1 [max(c j ,s j,i )− c j ]; ∆(i| S)← αr i + β∆ C (i| S); i ⋆ ← arg max i/∈S ∆(i| S); S ← S∪i ⋆ ; c j ← max(c j ,s j,i ⋆ ) for all j; return sort(S); 3.6 Submodularity and greedy approximation guarantee We now justify the greedy procedure theoretically. Monotonicity. R(S) is monotone since r i ≥ 0. For coverage, for each j, the max- imum max(b, max i∈S s j,i ) cannot decrease as S grows, hence C(S) is monotone. Therefore, for α,β ≥ 0, F(S) is monotone. Submodularity (diminishing returns). R(S) is modular and thus submodular. It suffices to show C(S) is submodular. For a fixed j, define g j (S)≜ max b, max i∈S s j,i − b, so C(S) = X j∈G g j (S). Take any A ⊆ B ⊆ G and e /∈ B. Let m A = max(b, max i∈A s j,i ) and m B = max(b, max i∈B s j,i ), so m A ≤ m B . Then the marginal gains satisfy g j (A∪e)−g j (A) = max(0,s j,e −m A ) ≥ max(0,s j,e −m B ) = g j (B∪e)−g j (B), which proves each g j is submodular, hence C(S) is submodular. A nonnegative linear combination preserves submodularity, so F(S) is monotone submodular. Greedy guarantee. Maximizing a normalized monotone submodular F under |S|≤ K with greedy yields the classic bound F(S greedy ) ≥ 1− 1 e F(S ⋆ ), (4) where S ⋆ is the optimal subset of size at most K. 8F. Author et al. Match to implementation. Our maintained coverage c j equals c j (S) = max b, max i∈S s j,i , and the per-step score ∆(e| S) = αr e + β X j∈G max c j (S),s j,e − c j (S) is exactly the marginal gain F(S∪e)−F(S), so Alg. 1 is the greedy algorithm for Eq. (1). 3.7 Strategy presets (four modes) We instantiate four strategies by choosing (α,β) while keeping the frame budget fixed (e.g., K=32 by default): – Relevance-only: (α=1,β=0). – Coverage-only: (α=0,β=1). – Relevance-oriented: (α=1,β=λ) with λ∈ (0, 1). – Coverage-oriented: (α=λ,β=1) with λ∈ (0, 1). These presets form a simple family that interpolates between query-focused se- lection and representativeness-focused selection. 3.8 Question-type adaptive selection Different question types benefit from different trade-offs: topic/summary ques- tions typically require broader representativeness, while needle-in-a-haystack questions benefit from higher relevance. Question-type classifier. We train a lightweight question-type classifier using only the question text, supervised by the benchmark’s official taxonomy. At inference, it predicts a type ˆy for each question. Type-to-strategy mapping. On a validation split, we evaluate the four presets per question type and select the best preset for each type: s ⋆ (y) = arg max s∈S Acc val (s| y), where S is the set of four presets. During testing, given ˆy we apply greedy selection with preset s ⋆ (ˆy), yielding an interpretable question-type → strategy policy. 4 Experiments 4.1 Question-Type Classifier for Adaptive Routing Our adaptive selector routes each question to a fixed relevance–coverage trade- off preset (Sec. 3). This requires a reliable question-type predictor. We therefore train a lightweight text-only classifier on the MLVU question text to predict the question taxonomy used in our evaluation split (7 types: plotQA, needle, ego, count, order, anomaly_reco, topic_reasoning). We optimize standard cross-entropy and train for 10 epochs. Abbreviated paper title9 (a) Training loss and test accuracy over 10 epochs.(b) Confusion matrix on the test set. Fig. 1: Question-type classifier used for adaptive strategy routing. The clas- sifier achieves 97.7% test accuracy (425/435), with most errors concentrated between topic_reasoning and plotQA. Learning dynamics. Fig. 1 (left) shows that the classifier converges quickly: the average training loss decreases from ∼1.39 at epoch 1 to ∼0.04 by epoch 10, while test accuracy jumps from ∼75% to ∼95% after one additional epoch and stabilizes around∼97–98% thereafter. This indicates that the routing signal can be learned from question text alone with minimal training. Confusions and per-type behavior. Fig. 1 (right) reports the confusion matrix on the held-out test set. The overall accuracy is 97.7% (425/435). Most types are nearly perfectly separated (e.g., count, order, and anomaly_reco have no observed confusions in this split). The dominant errors are concentrated between topic_reasoning and plotQA (4 instances of topic_reasoning predicted as plotQA), with a small number of plotQA questions misclassified as needle or ego. Importantly, these confusions occur among semantically adjacent types, suggesting that routing noise is limited and unlikely to dominate the overall VLM performance. Usage in subsequent experiments. Unless stated otherwise, we use the predicted type to select the routing preset. To isolate the effect of routing errors, we addi- tionally report an oracle routing variant (ground-truth type → preset) in abla- tions. 4.2 Main Results: Accuracy vs. Frame Budget We report the core performance comparison under a controlled frame bud- get, where each method selects K frames from the same 1 FPS candidate pool (capped at 1000). Fig. 2 plots average accuracy as a function of selected frame count. 10F. Author et al. Fig. 2: Main result: average accuracy vs. selected frame count. Our greedy relevance+coverage selection consistently improves over uniform sampling and AKS (CVPR 2025), with the largest gains under tight budgets. The Optimized (task-wise best) curve shows the additional benefit from question-type adaptive strategy selection. Consistent gains over uniform sampling. Across the entire budget range, our greedy selector substantially outperforms uniform sampling. The improvement is especially pronounced at small budgets (e.g., 5–20 frames), demonstrating that simply adding more frames is not sufficient—the choice of which frames to keep is critical for long-video VLM inference. While uniform sampling improves steadily with larger K, it remains consistently below our selected subsets. Outperforming AKS. We further compare against AKS (CVPR 2025), a re- cent strong baseline for frame selection. Our method achieves higher accuracy across nearly all budgets, with the largest margin again appearing in the low-to- mid budget regime. This indicates that jointly optimizing query relevance and semantic-space representativeness yields a stronger subset than AKS under the same frame constraints. Benefit of question-type adaptivity. The Optimized (task-wise best) variant— which selects the best relevance–coverage preset per question type—provides the highest performance curve, offering an upper bound on the gains achievable via strategy routing. Importantly, this curve is consistently above any single fixed preset, confirming that no single trade-off is universally optimal and motivating our question-type adaptive design. Takeaway. Overall, the results show that our greedy relevance+coverage selec- tion is a practical and effective token-reduction mechanism: it delivers higher long-video question-answering accuracy than both uniform sampling and the recent AKS baseline under identical frame budgets. Abbreviated paper title11 5 Conclusion We presented a query-adaptive greedy frame selection method for long-video VLM inference that explicitly balances query relevance with semantic-space cov- erage to avoid both redundancy collapse and coverage collapse. Our selector optimizes a simple parameterized objective that combines a modular relevance term with a facility-location coverage term, enabling efficient greedy maximiza- tion with a standard (1 − 1/e) approximation guarantee under a cardinality (frame-budget) constraint. To handle the fact that the optimal relevance–coverage trade-off is question dependent, we introduced an interpretable routing mechanism that maps pre- dicted question types to one of four preset trade-offs; a lightweight text-only classifier achieves 97.7% test accuracy, making this adaptive routing practical. Experiments on MLVU show that our greedy relevance+coverage selection consistently improves long-video QA accuracy over uniform sampling and the recent AKS baseline, with the largest gains under tight frame budgets; moreover, the task-wise optimized (type-adaptive) variant forms the strongest curve and confirms that no single fixed trade-off is universally optimal. Future work includes extending evaluation beyond MLVU to additional long- video benchmarks (e.g., Video-MME, LongVideoBench, EgoSchema), and ex- ploring richer routing policies that go beyond coarse question types to incorpo- rate uncertainty, video length, and iterative evidence acquisition. References 1. Buch, S., Nagrani, A., Arnab, A., Schmid, C.: Flexible frame selection for efficient video reasoning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). p. 29071–29082 (June 2025) 2. Carbonell, J., Goldstein, J.: The use of MMR, diversity-based reranking for re- ordering documents and producing summaries. In: Proceedings of the 21st Annual International ACM SIGIR Conference on Research and Development in Informa- tion Retrieval (SIGIR). p. 335–336. ACM (1998). https://doi.org/10.1145/ 290941.291025, https://doi.org/10.1145/290941.291025 3. Carbonell, J., Goldstein, J.: The use of mmr, diversity-based reranking for reorder- ing documents and producing summaries. In: Proceedings of the 21st Annual In- ternational ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR). p. 335–336. ACM (1998) 4. Fu, C., Dai, Y., Luo, Y., Li, L., Ren, S., Zhang, R., Wang, Z., Zhou, C., Shen, Y., Zhang, M., Chen, P., Li, Y., Lin, S., Zhao, S., Li, K., Xu, T., Zheng, X., Chen, E., Shan, C., He, R., Sun, X.: Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). p. 24108–24118 (June 2025) 5. Krause, A., Golovin, D.: Submodular function maximization. In: Tractability: Prac- tical Approaches to Hard Problems. Cambridge University Press (Feb 2014) 6. Krause, A., Golovin, D.: Submodular function maximization. In: Bordeaux, L., Hamadi, Y., Kohli, P. (eds.) Tractability: Practical Approaches to Hard Problems, p. 71–104. Cambridge University Press (2014) 12F. Author et al. 7. Kulesza, A., Taskar, B.: Determinantal point processes for machine learning. Foundations and Trends in Machine Learning 5(2-3), 123–286 (2012). https: //doi.org/10.1561/2200000044, https://doi.org/10.1561/2200000044 8. Lin, H., Bilmes, J.: A class of submodular functions for document summariza- tion. In: Lin, D., Matsumoto, Y., Mihalcea, R. (eds.) Proceedings of the 49th An- nual Meeting of the Association for Computational Linguistics: Human Language Technologies. p. 510–520. Association for Computational Linguistics, Portland, Oregon, USA (Jun 2011), https://aclanthology.org/P11-1052/ 9. Liu, S., Zhao, C., Xu, T., Ghanem, B.: Bolt: Boost large vision-language model without training for long-form video understanding. In: Proceedings of the Com- puter Vision and Pattern Recognition Conference (CVPR). p. 3318–3327 (June 2025) 10. Ma, Z., Gou, C., Shi, H., Sun, B., Li, S., Rezatofighi, H., Cai, J.: Drvideo: Document retrieval based long video understanding. In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR). p. 18936–18946 (June 2025) 11. Mangalam, K., Akshulakov, R., Malik, J.: Egoschema: A diagnostic benchmark for very long-form video language understanding (2023), https://arxiv.org/abs/ 2308.09126 12. Nemhauser, G.L., Wolsey, L.A., Fisher, M.L.: An analysis of approximations for maximizing submodular set functions - I. Mathematical Programming 14(1), 265– 294 (1978). https://doi.org/10.1007/BF01588971, https://doi.org/10.1007/ BF01588971 13. Song, E., Chai, W., Wang, G., Zhang, Y., Zhou, H., Wu, F., Chi, H., Guo, X., Ye, T., Zhang, Y., Lu, Y., Hwang, J.N., Wang, G.: Moviechat: From dense token to sparse memory for long video understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). p. 18221– 18232 (June 2024) 14. Tang, X., Qiu, J., Xie, L., Tian, Y., Jiao, J., Ye, Q.: Adaptive keyframe sampling for long video understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). p. 29118–29128 (June 2025) 15. Wang, Z., Yu, S., Stengel-Eskin, E., Yoon, J., Cheng, F., Bertasius, G., Bansal, M.: Videotree: Adaptive tree-based video representation for llm reasoning on long videos. In: Proceedings of the Computer Vision and Pattern Recognition Confer- ence (CVPR). p. 3272–3283 (June 2025) 16. Wu, H., Li, D., Chen, B., Li, J.: Longvideobench: A benchmark for long-context interleaved video-language understanding (2024), https://arxiv.org/abs/2407. 15754 17. Yao, L., Wu, H., Ouyang, K., Zhang, Y., Xiong, C., Chen, B., Sun, X., Li, J.: Generative frame sampler for long video understanding. In: Che, W., Nabende, J., Shutova, E., Pilehvar, M.T. (eds.) Findings of the Association for Computational Linguistics: ACL 2025. p. 17900–17917. Association for Computational Linguis- tics, Vienna, Austria (Jul 2025). https://doi.org/10.18653/v1/2025.findings- acl.921, https://aclanthology.org/2025.findings-acl.921/ 18. Zhang, X., Wu, Z., Li, Z., Xu, H., Gong, L., Boussaid, F., Werghi, N., Bennamoun, M.: Adard-key: Adaptive relevance-diversity keyframe sampling for long-form video understanding (Oct 2025). https://doi.org/10.48550/arXiv.2510.02778, https://arxiv.org/abs/2510.02778 19. Zhang, X., Wu, Z., Li, Z., Xu, H., Gong, L., Boussaid, F., Werghi, N., Bennamoun, M.: Adard-key: Adaptive relevance-diversity keyframe sampling for long-form video understanding (2025), https://arxiv.org/abs/2510.02778 Abbreviated paper title13 20. Zhou, J., Shu, Y., Zhao, B., Wu, B., Liang, Z., Xiao, S., Qin, M., Yang, X., Xiong, Y., Zhang, B., Huang, T., Liu, Z.: Mlvu: Benchmarking multi-task long video understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). p. 13691–13701 (June 2025) 21. Zou, J., Huang, Z., Zhang, S., Zhang, L., Shen, W.: Videobrain: Learning adaptive frame sampling for long video understanding (Feb 2026). https://doi.org/10. 48550/arXiv.2602.04094, https://arxiv.org/abs/2602.04094 22. Zou, J., Huang, Z., Zhang, S., Zhang, L., Shen, W.: Videobrain: Learning adaptive frame sampling for long video understanding (2026). https://doi.org/10.48550/ arXiv.2602.04094, https://arxiv.org/abs/2602.04094

Instant research discovery