PaperlayerView exact revision on arXiv

Source-linked AI summary

Rethinking Long-Video Efficiency: A Joint Allocation Perspective on Frames, Pixels, and Front-End Latency

Sixun Dong, Wei Li, Andong Deng, Qi Qian, Victor Zhu, Zhengping Ji, Chen Chen

arXiv:2610.04318v1cs.CV

TL;DR

Existing efficient long-video methods largely select frames or tokens at native resolution, leaving resolution and front-end latency outside the allocation. LoHi combines dense low-resolution video with sparse high-resolution images; at matched visual-token budgets it improves average accuracy by 10.6% over the native-resolution baseline.

  • Problem

    Existing efficient long-video VLMs largely select frames or visual tokens at fixed native resolution, leaving per-frame resolution and front-end latency outside the budget allocation.

    Source: §1 · p. 1

  • Method

    LoHi is a training-free, single-pass framework that combines dense low-resolution video for temporal coverage with sparse high-resolution images for spatial detail through native VLM pathways.

    Source: §1 · p. 2

  • Results

    10.6% higher average accuracy than the native-resolution baseline at the same visual-token budget, with front-end latency reduced by up to 7× on hour-scale clips.

    Source: §6 · p. 9

  • Takeaways & Limitations

    The decoupled streams separate temporal coverage from spatial detail, while plug-and-play selectors support different cost–accuracy trade-offs.

    Source: §1 · p. 2, §6 · p. 9

    Findings apply most directly to modern and future architectures; older or smaller models may need higher base resolution or larger Hi-I budgets.

    Source: E Limitations and Outlook · p. 25

Abstract

from arXiv · show

Efficient long-video understanding with vision-language models (VLMs) is often framed as selecting informative frames or visual tokens at a fixed native resolution. We show that per-frame resolution can instead be traded for denser temporal coverage, while front-end decoding latency depends on the size of the candidate pool rather than the final token budget. An empirical study across multiple VLMs and long-video benchmarks yields three findings: dense low-resolution sampling outperforms sparse native-resolution sampling at matched token budgets; resolution-sensitive tasks benefit from selected high-resolution frames; and front-end decoding dominates wall time for hour-long videos. Motivated by these findings, we introduce LoHi, a training-free, single-pass framework that combines a dense low-resolution video stream with sparse high-resolution image frames through the VLM's native video and image pathways. LoHi-Anchor selects high-resolution frames using codec I-frame metadata, while LoHi-SemDiv uses query relevance and visual diversity over CLIP features. Across three long-video benchmarks, LoHi improves average accuracy by 10.6 percentage points over the native-resolution baseline at a matched token budget and by 5.2 percentage points over the strongest prior efficiency method. It also reduces front-end decoding latency by up to 7x on hour-long videos. Project page: https://sixundong.com/projects/lohi

1 Introduction

Long-video efficiency requires jointly allocating frame count, per-frame resolution, and front-end latency, rather than selecting tokens at fixed native resolution. The section distills empirical lessons into LoHi, which separates temporal coverage from spatial detail.

1 Introduction

  • The central efficiency question is how to jointly allocate an end-to-end budget across frame count, per-frame resolution, and front-end latency.Earlier approaches largely treated efficiency as selecting frames or visual tokens at fixed native resolution.Source: §1 · p. 1
  • At matched token budgets, dense low-resolution sampling beats sparse native-resolution sampling; aggressive downscaling hurts OCR and attribute perception, while uniform fixed-budget sampling keeps latency constant.Decode-then-select latency scales linearly with video duration and dominates wall time on hour-scale videos.Source: Figure 1 · p. 2, §1 · p. 2
  • LoHi combines dense low-resolution video for timeline coverage with sparse high-resolution images for fine spatial evidence through native VLM pathways, without fine-tuning or architectural changes.The streams are linked by an inline text marker.Source: §1 · p. 2
  • LoHi improves accuracy over the strongest keyframe-selection and token-pruning baselines at matched token budgets, while reducing front-end decoding latency by up to 7× on hour-scale videos.Source: §1 · p. 2

2 Related Work

Prior long-video efficiency methods largely retain native per-frame resolution, while recent resolution-aware approaches address only parts of the frame–resolution allocation problem.

2 Related Work

  • Table 1 qualitatively positions LoHi against other long-video efficiency methods by per-frame cost and front-end latency.Source: §2 · p. 2, Table 1 · p. 3
  • Fixed-resolution methods use pre-encoder keyframe selection or post-encoder token pruning, limiting frame count and temporal coverage; keyframe methods also incur duration-dependent front-end decode latency.Source: §2 · p. 3
  • Resolution-aware approaches include VisionThink [39], image-only MMTok [6], inter-frame VideoLLaMA3 [41], and query-adaptive Q-frame [43]; LoHi instead pairs fixed-resolution Lo-V with sparse Hi-I.Source: §2 · p. 3

3 Lessons on Frames, Resolution, and Front-End Latency

The study treats frame count, resolution, and front-end latency as a joint allocation problem under a fixed visual-token budget. It finds that dense low-resolution sampling often helps, but the preferred frame–resolution allocation depends on task sensitivity, and dense candidate-pool decoding can be costly.

3.2 Lesson 2 – Optimal frame-resolution allocation is task-dependent

  • At a matched token budget, (64, 0.5) wins on resolution-sensitive tasks while (256, 0.25) wins on the remaining tasks; no single pair suits both subsets.The aggregate winner therefore reflects benchmark subset proportions, and the low-resolution baseline is suboptimal on the resolution-sensitive subset.Source: Figure 2 · p. 4, §3.2 · p. 4

3.1 Lesson 1 – Dense, low-resolution sampling is the new recipe

  • Under ~5,760 tokens, Qwen3-VL-4B at (256, 0.25) beats the default by +6.66% on VideoMME, +10.47% on MLVU, and +5.10% on LVBench.The comparison uses (16, 1) as the default configuration.Source: §3.1 · p. 3, Table 2 · p. 4

3.3 Lesson 3 – Front-end latency cannot be overlooked in long videos

  • 7.4× is the latency gap reached by dense sampling versus uniform sampling on 60-minute videos, and it remains substantial on VideoMME.Dense sampling at FPS=1 increases GoPs touched and in-GoP depths with duration, whereas uniform sampling at N=256 caps both.Source: Figure 2 · p. 4, §3.3 · p. 5

4 The LoHi Framework

LoHi addresses frame–resolution allocation under token and decode budgets by combining dense low-resolution video with sparse high-resolution images. The framework uses native video and image pathways, with diversity-aware selection and an optional entropy-triggered variant.

4.1 Cross-Resolution Decomposition

  • LoHi splits one configuration into N low-resolution frames at rℓ and K ≪ N high-resolution frames at rh > rℓ, with Hi-I indices drawn from the Lo-V grid so each frame is decoded once.The streams jointly satisfy token and decode budgets.Source: §4.1 · p. 5

4.2 LoHi: Instantiating the Decomposition via Video and Image Pathways

  • Lo-V uses the video pathway’s globally consistent temporal positions, while Hi-I uses independent image encodings, leaving the video timeline undisturbed.LoHi-Uniform selects regular intervals with zero selection overhead; LoHi-Anchor uses precomputed I-frame indices without pixel decoding.Source: §4.2 · p. 6

4.3 LoHi-SemDiv: Diversity-Aware Selection via Quality-Similarity DPP

  • LoHi-SemDiv selects K high-resolution frames to balance query relevance and visual diversity through a DPP quality–similarity objective over frame-level CLIP features.Temporal coverage is already supplied by Lo-V, so Hi-I slots target distinct, query-relevant spatial evidence.Source: §4.3 · p. 6

4.4 Adaptive LoHi via Entropy Triggering

  • Entropy-triggered LoHi runs Lo-V first and invokes Hi-I only when answer uncertainty exceeds θ, nearly matching full LoHi while using Hi-I on one-third of queries.The variant bypasses the high-resolution stream for simpler queries.Source: §4.4 · p. 7

5 Experiments

On three long-video benchmarks, LoHi-SemDiv leads prior methods at matched token cost, while experiments examine its efficiency, task-dependent resolution needs, and component contributions. The ablations also test selector robustness and illustrate selected high-resolution frames.

5 Experiments

  • The evaluation uses VideoMME without subtitles, MLVU, and LVBench, primarily with Qwen3-VL-4B, at a shared token budget of approximately 5,760.LoHi uses 128 frames at 0.25 resolution plus 8 high-resolution images; baselines decode 256 frames.Source: §5 · p. 7

5.1 Performance Comparison across different benchmarks

  • LoHi-SemDiv reaches 66.81 / 74.38 / 45.90% on VideoMME / MLVU / LVBench, gaining +10.60% over (16F,1.0) and +5.23% over BOLT[17] at identical token cost.It achieves the best performance on all three benchmarks.Source: §5.1 · p. 7, Table 3 · p. 8
  • LoHi reduces front-end latency from 5.4 s to 3.3 s and ViT latency from 2,486 to 195 ms, approximately 12×, versus native-resolution token-pruning pipelines.LoHi decodes 128 frames rather than the 256 decoded by prior baselines; its measured GPU memory footprint is 9.6 GB.Source: Table 5 · p. 8, §5.1 · p. 8
  • All three LoHi variants outperform single-configuration baselines on both resolution-sensitive and resolution-insensitive task subsets.The passage attributes the gains to complementary video and image pathways, not a specific Hi-I selector.Source: §5.1 · p. 8, Table 4 · p. 8

5.2 Rethinking Token Pruning for Video-VLMs

  • Under fixed token budgets, simple resizing matches or outperforms dedicated token-pruning methods, which process native-resolution frames before discarding tokens.The paper also argues that fragmented patch pruning harms spatial continuity and temporal consistency in video.Source: §5.2 · pp. 8–9, Table 6 · p. 8

5.3 Ablations

  • Across 12 Hi-I count and selector configurations, VideoMME accuracy is 65.22–66.81%, above the 63.00% Lo-V-only baseline.SemDiv improves over Uniform at every tested Hi-I budget; on EgoLongQA it matches a higher-resolution single-stream baseline with 29% fewer visual tokens.Source: Table 7 · p. 9, §5.3 · p. 9
  • At a controlled 5,760-token budget, combining 128 Lo-V frames and 8 Hi-I images reaches 66.81%, versus 64.44% for Lo-V alone and 63.74% for Hi-I alone.The reported results characterize dense temporal context and spatial clarity as complementary, non-replaceable benefits.Source: Table 8 · p. 9, §5.3 · p. 9
  • Figure 4 shows Anchor recovering a scoreboard after replacing a black frame, while SemDiv selects query-relevant frames before and after a museum bombing.These are qualitative examples of the selected high-resolution images.Source: §5.3 · p. 9, Figure 4 · p. 9

6 Conclusion

The conclusion frames long-video VLM efficiency as a joint allocation problem over frame count, resolution, and front-end latency, and presents LoHi as a training-free combination of dense low-resolution video and sparse high-resolution images. At matched visual-token budget, it reports higher average accuracy than the native-resolution baseline and lower front-end latency on hour-scale clips.

6 Conclusion

  • LoHi addresses joint allocation across frame count, per-frame resolution, and front-end latency with a training-free, single-pass combination of dense low-resolution video and sparse high-resolution images.Source: §6 · p. 9
  • 10.6% higher average accuracy than the native-resolution baseline at the same visual-token budget, with front-end latency reduced by up to 7× on hour-scale clips.Source: §6 · p. 9
  • The plug-and-play selectors support different cost–accuracy trade-offs; future work can explore richer selectors and adaptive allocation.Source: §6 · p. 9

A Extra Method and Experiment Details

The appendix details codec-guided and semantic selectors for LoHi, plus exploratory variants that combine codec metadata or condition high-resolution selection on the low-resolution stream. The combined selectors show small gains, while plain SemDiv remains the default.

A Extra Method and Experiment Details

  • The appendix also specifies two exploratory extensions: combining codec and semantic signals, and conditioning Hi-I selection on Lo-V content.The authors present these as variants extending the SemDiv selector.Source: A Extra Method and Experiment Details · p. 14

A.1 LoHi-Anchor

  • LoHi-Anchor starts from uniformly spaced positions, checks nearby I-frames to avoid duplicate selections, then maps each selected I-frame to the nearest Lo-V grid index.It uses codec keyframe metadata, does not use the question or run a CLIP pass, and reuses the decoded Lo-V candidate pool.Source: §A.1 · p. 14

A.2 LoHi-SemDiv

  • LoHi-SemDiv greedily selects frames that maximize the regularized log-determinant objective, with a (1−1/e) approximation guarantee.With rank-one Cholesky updates it runs in O(NK^2), adds one CLIP pass over the N low-resolution frames, and its costs are described as negligible compared with VLM prefill.Source: §A.2 · p. 14

A.3 LoHi-SemDiv-Anchor

  • LoHi-SemDiv-Anchor snaps greedy candidates to nearby I-frame-associated Lo-V grid positions within a duration- and budget-dependent radius.On short clips snapping is disabled, reducing the variant to plain SemDiv; the procedure checks nearby I-frames to avoid duplicate grid selections.Source: §A.3 · pp. 14–15
  • LoHi-SemDiv-Anchor improves over plain SemDiv by +0.19% at K=4 and +0.08% at K=8 on VideoMME, with the two selectors on par at K=2.The authors describe the variant as exploratory because the gains are small and its constants were set by rough reasoning rather than tuning.Source: §A.3 · p. 15, Table 9 · p. 15

A.4 LoHi-SemDiv++: Lo-V-Conditioned Hi-I Selection

  • LoHi-SemDiv++ projects Hi-I candidate embeddings away from the principal Lo-V subspace before computing the SemDiv kernel.It improves slightly over plain SemDiv, but the authors retain plain SemDiv as the default and leave deeper study of Lo-V conditioning to future work.Source: §A.4 · p. 15, Table 10 · p. 16

B.2 Front-end decoding latency profiling

The appendix profiles front-end decoding and TTFT components, then tests whether LoHi’s allocation strategy transfers across backbones and frame counts. The reported results show transfer to additional models and continued gains from combining denser low-resolution coverage with high-resolution frames.

B.2 Front-end decoding latency profiling

  • The decoding comparison uses VideoMME clips with a median duration of about 8.1 minutes and averages clips near 10-, 30-, and 60-minute reference durations.CPU decord was faster than NVDEC in the reported comparisons, and the authors report latency using CPU decord; they state that front-end decoding dominates VLM compute on long clips for both backends.Source: §B.2 · p. 16, Table 11 · p. 16
  • LoHi’s shared decode pass produces 128 Lo-V frames and K Hi-I frames; its decode workload therefore uses fewer sampled frames than the 256-frame comparison.The appendix notes that decoding remains at native resolution before resizing, and resizing accounts for roughly 5% of decode wall time.Source: §B.2 · pp. 16–17, §B.4 · p. 17

B.3 TTFT pipeline profile

  • The TTFT profile separates decode, selector, ViT, and LLM prefilling, measuring model components on a dedicated H100 node.The profile reuses a fixed decoded tensor for component comparisons and uses separately measured CPU decord timings for decode.Source: §B.3 · p. 17

C.1 LoHi on Qwen3-VL-8B

  • All three LoHi variants outperform prior efficiency baselines on average while using approximately 25% fewer tokens; LoHi-Anchor scores 67.70% on VideoMME, and LoHi-SemDiv scores 75.34% on MLVU and 46.80% on LVBench.The Table 12 comparison uses 128F Lo-V at r_l=0.25 paired with K=4 Hi-I frames at r_h=1.0.Source: §C.1 · p. 17, Table 12 · p. 18

C.2 Generalization to Qwen3.5

  • At the same token budget, 256F@0.25 beats 16F@1.0 on all three Qwen3.5-4B benchmarks; LoHi-SemDiv also beats the dense-base setting on every benchmark.The largest reported LoHi-SemDiv gain over the 256F@0.25 Low-Res-Base is +5.23% on LVBench.Source: §C.2 · p. 18, Table 13 · p. 18, Table 14 · p. 18

C.3 Generalization to Additional VLM Backbones

  • Across Qwen2.5-VL-7B, VideoChat3-4B, and VideoLLaMA3-7B, the dense low-resolution base beats the 16-frame vanilla setting at matched token budgets.LoHi-SemDiv improves over the dense base in seven of nine model–benchmark combinations, with decreases of −0.26% and −0.13% in the other two.Source: §C.3 · p. 18, Table 15 · p. 19

C.4 Pushing the Lo-V to N=256

  • With 256 Lo-V frames and four Hi-I frames, denser Lo-V coverage improves VideoMME accuracy, and adding Hi-I frames still brings a clear gain.The authors describe the two streams as complementary as the Lo-V grows denser.Source: §C.4 · p. 19, Table 16 · p. 19

C.5 Evaluation under Larger Token Budgets

The appendix evaluates larger token budgets, adaptive two-pass inference, model-specific evaluation details, prompt placement, and SemDiv’s sharpening coefficient. It reports gains from adding Hi-I frames and finds that interleaving the prompt cue modestly benefits some selectors.

C.5 Evaluation under Larger Token Budgets

  • Accuracy rises monotonically from 82.14% to 87.57% as Hi-I budget K increases from 0 to 128 on EgoLongQA.At matched accuracy of 85.86%, 512F@0.25 with K=64 uses 29% fewer visual tokens (71K vs. 100K) and 29% fewer ViT patches (285K vs. 401K) than 512F@0.5.Source: §C.5 · pp. 19–20, Table 17 · p. 20

C.6 Adaptive cost control via multi-turn triggering

  • On VideoMME, only about a third of questions enter the Hi-I second pass, while accuracy nearly matches always running Hi-I.The entropy threshold is tuned on a held-out 10% split, and the authors call the result exploratory because threshold selection without benchmark-specific tuning remains future work.Source: §C.6 · p. 20, Table 18 · p. 20

C.7 Answer-Format Normalization for VideoLLaMA3 on MLVU

  • VideoLLaMA3’s MLVU scores are increased by 16–17% after normalizing predictions that end in a period instead of the expected closing parenthesis.The passage attributes the change to answer extraction formatting, not accuracy, and says the frame-count trend is unchanged.Source: §C.7 · p. 20, Table 19 · p. 20

C.8 The inter-frame token compressor in VideoLLaMA3 removes almost no tokens

  • VideoLLaMA3-7B’s DiffFP keeps a mean token ratio above 0.959 across the tested configurations, removing at most about 4% of tokens on average.The authors attribute this saturation to reduced-resolution merged patches covering larger image regions, where motion tends to push pixel differences above the token-drop threshold.Source: §C.8 · p. 21, Table 20 · p. 21

C.9.1 Prompt template: appended cue vs. interleaved cue

  • Across selectors and Hi-I budgets, the cross-resolution decomposition helps under either prompt cue, and the gap between prompt variants never exceeds 0.8%.The interleaved cue adds +0.44 to +0.74% for Uniform and SemDiv, but Anchor does not show the same benefit.Source: §C.9.1 · p. 21, Table 21 · p. 21

C.9.2 SemDiv sharpening coefficient α

  • With α set to 1, 5, and 10, VideoMME accuracy is 66.67%, 66.81%, and 66.48%, respectively, indicating limited sensitivity over the tested range.The authors use α=5 for all LoHi-SemDiv runs without per-task tuning.Source: §C.9.2 · p. 22

C.10 Further Analysis of Hi-I Selection

Further analysis finds an overall advantage for SemDiv over Uniform, with gains that vary by benchmark and query type. The appendix also reports low selection overhead and keeps Uniform as an option when semantic-selection overhead is undesirable.

Statistical significance

  • On MLVU, SemDiv significantly improves over Uniform by +6.76% (p < 10^-16), while differences on VideoMME and LVBench are not statistically significant.The paired comparison uses video-level bootstrap confidence intervals.Source: §C.10 · p. 22, Table 22 · p. 22

Benchmark dependence

  • Query-aware selection gains over uniform selection are largest on MLVU and small or negative on VideoMME for SemDiv, CLIP-TopK, and AKS.The passage cautions that gain magnitudes are not directly comparable between keyframe methods and SemDiv because their selected inputs differ.Source: §C.10 · p. 22, Table 23 · p. 22

Task-level benefits

  • On MLVU action counting, SemDiv reaches 53.40% versus Uniform’s 29.13%, a +24.27% improvement (95% CI: [+16.5, +32.0]).Uniform is below the Lo-V-only baseline of 41.75%; leave-one-category-out differences remain +4.93 to +8.06%, all with p < 10^-9.Source: §C.10 · p. 22, Table 24 · p. 22

Variation across query types

  • Within MLVU’s ego category, SemDiv reaches 69.03% against Uniform’s 59.66%, while static entity counting improves from 18.52% to 37.04% with either selector.Transient-action accuracy rises from 66.67% with Uniform to 83.33% with SemDiv; the authors leave query-aware adaptive frame–resolution allocation for future work.Source: §C.10 · p. 23, Table 25 · p. 23

Selection overhead

  • SemDiv adds approximately 70 ms over Uniform, or 1.8% of TTFT, and its 120 ms selector-stage latency is below the keyframe baselines’ 130 ms.The passage describes SemDiv as scoring only the 128 Lo-V candidates.Source: §C.10 · p. 23

Performance across benchmarks and task categories

  • SemDiv is never significantly worse than Uniform on evaluated benchmarks or task categories, and a stratified exact test finds an overall advantage (p = 5 × 10^-13).The authors choose SemDiv as the default while retaining Uniform as an option without semantic-selection overhead.Source: §C.10 · p. 23

C.11 Evaluation with Subtitles

Subtitle-free evaluation isolates visual allocation because transcripts alone provide substantial information and consume context; the discussion positions LoHi as complementary to codec-aware systems and explains its separation of temporal coverage from spatial detail.

C.11 Evaluation with Subtitles

  • At the 90th percentile, transcripts exceed the 5,760-token visual budget, while transcript-only input reaches 61.56% on subtitled questions versus 63.13% for the 256-frame visual pipeline.The main evaluation excludes subtitles to isolate visual allocation; the comparison uses transcript and question without visual input.Source: §C.11 · pp. 23–24, Table 26 · p. 23, Table 27 · p. 24
  • +2.15% accuracy (p = 0.011) for LoHi-SemDiv over Low-Res-Base without subtitles; with subtitles, their difference is not significant (−0.99%, p = 0.117).On the same subtitle-available subset, adding subtitles improves both methods, with a larger gain for Low-Res-Base (+9.99 vs. +6.85%); LoHi uses 128 rather than 256 sampled frames.Source: §C.11 · p. 24, Table 28 · p. 24

D.1 Codec-Aware and System-Level Video Processing

  • Codec-aware methods reduce per-frame decoding cost but require substantial training or engineering; LoHi instead uniformly samples a fixed frame budget, keeping front-end cost independent of video duration.Uniform timestamps also support multimodal RoPE-based temporal modeling; LoHi is presented as complementary to codec-aware tokenizers and runtime systems.Source: §D.1 · p. 24

D.2 Anyres and Slow–Fast: Two Single-Axis Decompositions

  • Anyres patching and token merging make low-resolution video frames token-efficient, while LoHi combines video and image modalities to balance temporal density and spatial detail.The cited discussion reports about 20–30 merged tokens per Qwen3-VL frame at r=0.25 on a 720p source and says moderate downscaling preserves most accuracy, with losses concentrated in resolution-sensitive tasks.Source: §D.2 · p. 25
  • Slow–Fast-style heterogeneous token counts disrupt unified-stream VLMs’ multimodal RoPE temporal positions and typically require dedicated training, limiting plug-and-play use.LoHi keeps uniform per-frame allocation in the video stream and routes extra spatial detail through the image pathway, aligning with the pathways’ respective training distributions.Source: §D.2 · p. 25

E Limitations and Outlook

LoHi’s demonstrated scope is modern VLM architectures with dynamic resolution, and its integration has prompt-structure sensitivity; the outlook proposes extensions to selection, adaptive computation, and frame reuse.

E Limitations

  • LoHi is validated on frontier VLMs with dynamic-resolution architectures; older or smaller models may degrade more sharply at very low resolutions and need higher base resolution or a larger Hi-I budget.The authors state findings apply most directly to modern and future architectures where low-resolution inputs are highly token-efficient.Source: E Limitations and Outlook · p. 25
  • The training-free integration relies on a textual cue between streams, creating mild sensitivity to prompt structure; the study focuses on front-end and prefill latency, not VLM decoding latency.At matched visual-token budgets, LoHi reduces front-end cost without increasing visual context length during generation.Source: E Limitations and Outlook · p. 25

E Outlook

  • The authors propose combining LoHi with keyframe selection or token pruning, extending Lo-V-first inspection to open-ended tasks, and reusing decoder-processed frames as Hi-I candidates or motion priors.The proposed frame reuse draws on preceding I- and P-frames that the decoder processes when decoding a target P-frame, without increasing decode budget Ndec.Source: E Limitations and Outlook · p. 26

F Qualitative Results

The examples show that high-resolution frames can recover fine spatial evidence, while informed selection can replace uninformative frames or avoid redundant selections to surface task-relevant evidence.

F Qualitative Results

  • Hi-I frames restore the blurred chin detail needed to answer an interview question when the Lo-V-only prediction is wrong.The example illustrates that the image pathway is useful when a question requires fine spatial evidence.Source: F Qualitative Results · p. 26, Figure 5 · p. 26
  • LoHi-Anchor replaces a weakly informative frame with the question-answering scoreboard, while LoHi-SemDiv selects complementary action-centric frames instead of redundant ones.The examples contrast codec I-frame-guided selection with semantic-diversity selection.Source: F Qualitative Results · p. 26, Figure 6 · p. 27, Figure 7 · p. 27
  • Uniform Hi-I selection can miss evidence by choosing an uninformative frame or missing a brief event; Lo-V preserves low-resolution coverage, but fine-detail recovery still depends on informative Hi-I selection.The two pathways therefore address different evidence needs: timeline coverage and selected fine spatial detail.Source: F Qualitative Results · p. 26
Loading 2610.04318v1…