Pith. sign in

REVIEW 4 major objections 5 minor 17 references

A single memory grown from a streaming video, stratified into a perceptual trace, moments, episodes, and recurring patterns, and read through one retrieval skill per scale, can both answer questions about the past and decide when past exper

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-04 08:06 UTC pith:3CLUSVJK

load-bearing objection A genuinely interesting memory architecture whose headline benchmark wins are mostly backbone effect — the fixed-backbone ablations carry the paper. the 4 major comments →

arxiv 2608.02392 v1 pith:3CLUSVJK submitted 2026-08-03 cs.CV cs.AI

GROVE: Growing and Reasoning over Temporally Stratified Memory from Streaming Video Experience

Sitong Gong , Caixin Kang , Tianyu Yan , Guo Chen , Bo Zheng , Kaipeng Zhang , Yunzhi Zhuge , Xiang Ruan
show 2 more authors
Huchuan Lu Yifei Huang
This is my paper
classification cs.CV cs.AI
keywords streaming video memorytemporal stratificationproactive assistanceegocentric videoagentic retrievallifelong video understandingmemory consolidationquestion answering
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

GROVE claims that a wearable assistant's streaming video history can be organized into one memory stratified by temporal scale—a perceptual trace, time-stamped moments, coherent episodes, and recurring cross-day patterns—each paired with a dedicated retrieval skill. This one substrate serves both reactive question answering and proactive assistance, where retrieval is kicked off by a user query or by the current situation. The paper shows on five benchmarks, including long-horizon QA and proactive service tests, that this design beats flat or single-scale memory baselines, and its ablations show each stratum and each skill contributes, with patterns mattering most when evidence spans days. The deeper claim is that temporal scale is the right interface between stored experience and agentic retrieval, making separate memory and control mechanisms for proactive helpers unnecessary.

Core claim

On its own terms, GROVE establishes that a causally grown, temporally stratified textual memory—built online from sampled frames through a frozen vision-language model plus a lightweight consolidation model—matches or beats flat caption indexes and prior memory agents on long-horizon video question answering and proactive service benchmarks. The hierarchy is the mechanism: moments carry second-level detail, episodes provide activity-boundary spans, patterns encode recurrences absent from any single episode, and each stratum has a scale-native retrieval entry point. Ablations show that removing a stratum while replacing it with equal caption volume hurts accuracy, and restoring the full struc

What carries the argument

The load-bearing object is the stratified memory M = {R, F, E, P} — a perceptual trace of captions and structured entity/action/text records, timestamped moments, coherent episodes, and pattern groups over non-contiguous episodes — together with a library of four scale-native retrieval skills: Perception Lookup, Moment Recall, Episode Replay, and Pattern Traversal. The identity that carries the argument is the one-memory-two-policies design: a reactive user query or an unasked situational trigger both initiate retrieval through the same skills, so the memory does not care whether the demand is question-driven or situation-driven.

Load-bearing premise

The entire memory is built only from what a fixed captions model sees at fixed sampling intervals (one frame every 2–5 seconds), so any event, object, or text the model misses at ingestion is permanently absent and no later consolidation can recover it.

What would settle it

Take a video where a critical object or action appears only in the gap between two sampled frames, build GROVE's memory from that stream, and ask a question about that missing evidence. If the stratified memory cannot recover it and a flat caption index performs no worse, the claim that stratification preserves faithful long-horizon memory fails at the perception ceiling. Alternatively, if a flat caption index with the same evidence volume and an unrestricted retrieval budget matches GROVE's accuracy and latency, then the stratification itself is not load-bearing.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • One shared memory substrate suffices for both user-initiated recall and situation-initiated assistance; separate control modules are not needed.
  • Memory construction is causal and can keep pace with a live stream, reaching build times faster than real time at sparse sampling rates.
  • Each stratum contributes different evidence: patterns give the largest gains at month horizons, moments matter most for second-level proactive services, and removing the episode layer hurts day-scale recall most.
  • Stratification improves retrieval efficiency—fewer reasoning rounds and lower latency than a flat caption index—because the agent can enter memory at the right scale directly.
  • If the central claim is right, improving ingestion (denser sampling or a stronger captioner) should lift all four strata nearly proportionally, since the perception ceiling bounds everything above it.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The scale-native retrieval interface could transfer to other serialized experience streams, such as audio logs or sensor time series, where moments, episodes, and patterns have natural analogues; this is an extension, not a claim in the paper.
  • A testable prediction follows from the month-split ablations: on routine-poor data where activities rarely recur, the pattern layer should add little, locating GROVE's benefit specifically in regularity-rich egocentric life.
  • The results imply that agentic retrieval systems should be designed scale-aware from the start; a router trained to pick temporal scale first should beat flat tool lists on long-horizon tasks, which the paper leaves implicit.
  • Because the paper identifies perception as the ceiling, a direct stress test is to feed a stream with events that fall between sampled frames and measure whether any amount of stratification recovers them—if not, the memory's fidelity is bounded by the frozen captioner.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper introduces GROVE, a training-free framework that grows a temporally stratified memory from a streaming video and uses it for both reactive question answering and proactive assistance. The memory consists of a perceptual trace, time-stamped moments, episodes, and cross-episode patterns, each exposed through a scale-native retrieval skill; a multi-round agent composes these skills, and a separate schema-based policy handles proactive triggering. The system is evaluated on five benchmarks (MM-Lifelong, OVO-Bench, StreamingBench, ESTP-Bench, EgoServe), with controlled ablations on memory structure, retrieval skills, retrieval details, and efficiency. The central claim is that temporal scale is the right interface between stored experience and agentic retrieval, and that a single memory substrate suffices for both query-driven and situation-driven assistance. The paper also includes detailed appendices with prompts, hyperparameters, per-service ablations, and memory-scale statistics, and it explicitly acknowledges several limitations, including the frozen-perception ceiling and the use of different reasoning backbones across main-table comparisons.

Significance. If the central claim holds, the idea of making temporal scale the organizing interface for video memory is a useful and reasonably novel contribution, and the unified treatment of reactive and proactive access is an appealing design. The paper's strengths include a fully causal online construction pipeline, controlled ablations that hold information volume fixed when removing a stratum (Tables 6 and 9), a detailed retrieval-skill analysis (Tables 7, 8, 12, 19), and unusual transparency about benchmark-specific protocols and hyperparameters. The efficiency analysis (Figure 3, Table 9) is also a valuable practical datapoint. However, the published absolute results are weakened by the backbone confound: main-table comparisons mix GPT-5.2 and GPT-5-mini across benchmarks while baselines use their own backbones, and the paper's own Limitations section concedes that component-level conclusions rely on the fixed-backbone ablations. The ablations themselves are the main support for the architectural claim, and they are mostly well designed, but they lack repeated-run variance. The proactive evaluation also relies on benchmark protocols that expose future queries or task priors during const

major comments (4)
  1. [Experiments: Main Results (Tables 2–5) and Table 12] The headline 'best results among the compared methods' is not a clean test of the memory architecture. GROVE uses GPT-5.2 on MM-Lifelong and GPT-5-mini elsewhere, while baselines use their own backbones. Table 12 shows that holding memory, prompts, and tools fixed, replacing gpt-5-mini with gpt-5.2 raises MM-Lifelong day accuracy from 18.75 to 23.50 and week from 19.75 to 22.75. The 4.75-point day gain is comparable to the full-vs-flat ablation gap in Table 9 (18.75 vs 12.75–15.50) and to the reported margin over ReMA (6.75). The paper's own Limitations acknowledge that component-level conclusions rely on fixed-backbone ablations. Please either provide same-backbone comparisons for the strongest baselines (at least ReMA and one flat-memory agentic baseline) or restrict the superiority claims to the fixed-backbone ablations and rephrase the abstract and introduction accordingly.
  2. [Ablations (Tables 6–8) and Inference-Time Analysis (Fig. 3, Table 9)] All numerical claims rest on single runs with greedy decoding; no error bars, confidence intervals, or repeated-run variance are reported. Some ablation deltas are small (e.g., EgoServe 12.62 vs 11.77 in Table 6; MM-Lifelong week 19.75 vs 20.50 for w/o Perceptual Trace), so the complementarity claim could change under noise. Please report variance over repeated runs or at least a statistical significance test for the key ablations, and state clearly how many questions underlie the 20-question efficiency figures in Table 9 and Figure 3.
  3. [Appendix A.1 'Dataset-specific Instantiations' and proactive/forward tracks] For OVO-Bench forward, ESTP-Bench, and the instructional EgoServe subsets, the construction stage is given the candidate user questions, task type, or canonical step sequence; on HoloAssist and CaptainCook4D it receives the correct step sequence. This is disclosed as following the benchmark protocols, but it means these results do not test query-free discovery from a raw stream, which the abstract's 'recognize when that history is useful' claim implies. Please either provide a fully query-free evaluation or clearly separate 'protocol-assisted proactive' results in the abstract and claims. The fact that EgoServe and the EgoMemo baseline are from the same lab (Gong et al. 2026) also argues for an independent baseline or an explicit statement of this relationship.
  4. [Limitations (perception ceiling)] As the paper states, GROVE inherits errors from its frozen perception model and evidence missed during ingestion cannot be recovered through later consolidation. This is a legitimate scope restriction, but because all higher strata are built from VLM captions at 2s/5s sampling and structured traces, the reported gains are bounded by caption quality. A sensitivity analysis (e.g., varying the sampling rate or using a weaker/stronger captioner on a subset of each benchmark) would clarify how much of the benefit comes from stratification rather than from the specific perception model chosen. At minimum, the main text should state this ceiling more prominently than a single Limitations paragraph.
minor comments (5)
  1. [Abstract and Introduction] The abstract says 'best results among the compared methods' without the caveat about differing reasoning backbones; the Introduction also says 'all three benchmarks' after listing five benchmarks. Please correct the inconsistency and qualify the claim.
  2. [Table 4] Some baseline entries are '—'; please add a footnote explaining whether these are unavailable or not applicable. Also clarify how 'StreamingBench Avg' is computed (arithmetic mean of real-time and contextual?).
  3. [Table 5] EyeWO* is shown in gray and excluded from the bold-faced 'best non-gray' comparison. The caption should explain why it is treated as a benchmark-specific model and whether the comparison protocol differs.
  4. [Table 6 vs Table 9] The flat baseline in Table 6 ('Flat (w/o Hier.)' = 14.75/15.50) and the flat-caption baseline in Table 9 (12.75) differ. Please clarify the exact configuration difference (e.g., equal-volume raw captions vs. flat caption index) so readers can interpret both tables correctly.
  5. [General editorial] Typo: 'evlauate' in Section 4. Also, 'MM-lifelong' in the abstract is inconsistent with 'MM-Lifelong' elsewhere. The paper would benefit from a careful pass for such inconsistencies.

Circularity Check

2 steps flagged

Fixed-backbone ablations are self-contained, but headline cross-benchmark claims carry disclosed confounds: forward-track test questions/options are visible to memory construction, EgoServe/EgoMemo are same-lab, and two EgoServe ablation rows coincide by definition.

specific steps
  1. other [Appendix A.1 'Dataset-specific Instantiations'; Appendix D.2 FRAME_CAPTION_WINDOW_FORWARD_EGO_PROMPT and FRAME_CAPTION_WINDOW_ESTP_QAWARE_V2_PROMPT]
    "On the forward-looking tracks of OVO-Bench and ESTP-Bench, ... the candidate user questions of a video are visible to the construction stage; this follows the protocol of these tracks, in which the forward query is known in advance and only its firing time is evaluated. ... This is OFFLINE evaluation: the question and ALL OPTIONS are already shown to you. Your captions will later be used to pick the correct option, so you MUST help disambiguate between the options."

    The forward/contextual predictions are produced by a memory whose perception stage was handed the evaluated question and its full option set and instructed that captions 'will later be used to pick the correct option.' The claimed gains on these tracks (OVO forward 56.5; ESTP CQ 41.0 vs 26.6) are therefore not a clean test of query-free or even unassisted memory: evidence is collected under instructions to disambiguate the exact answer set, so part of the 'retrieval' result is fed in at ingestion. Baselines that see only raw video receive no such steering, so the cross-system comparison is confounded by construction. This is a disclosed, protocol-permitted leak rather than a full identity, since the video content must still be observed, which is why it is not treated as a fitted-parameter

  2. self definitional [Ablations, Table 7 note; Appendix B.1 'Retrieval Skills across Service Types' and Table 17]
    "On EgoServe, the schema injects evidence blocks instead of exposing tools, so removing a skill amounts to removing its stratum for perceptual trace and pattern layer. ... without Pattern Traversal memory recall (2.7) and routine optimization (4.2) fall to exactly the level seen when the PATTERN stratum itself is deleted, since on this benchmark the two interventions coincide."

    On EgoServe the 'skill' ablation is defined as the stratum deletion, so Table 17's w/o Perception Lookup and w/o Pattern Traversal rows are identical by construction to Table 16's w/o Perceptual Trace and w/o Pattern rows (12.27 and 11.49 in both tables). The paper's stated inference that this 'separates content value from entry-point value' is therefore not testable on EgoServe for these two layers; only the MM-Lifelong ablations, where content remains reachable when a skill is deleted, genuinely separate the two interventions. The authors disclose the coincidence, so this is a definitional overlap rather than hidden double-counting, but the corresponding claim is proportionately weaker.

full rationale

GROVE's core derivation is empirical and training-free: no parameter is fitted and then re-predicted, and the central architectural claim (temporally stratified memory plus scale-native retrieval beats a flat caption index at equal volume and budget) is supported by controlled ablations that are fixed-backbone and question-free. Tables 6, 7, and 9 all use GPT-5-mini on MM-Lifelong day/week (month in the appendix), and the paper states 'No question is visible during construction on MM-Lifelong, StreamingBench, and backward tracks of OVO-Bench.' Table 9's monotone accuracy restoration (12.75 to 18.75 with fewer rounds, 6.25 to 4.95, and lower latency) is a self-contained comparison: the flat baseline receives the same caption volume, so the result does not reduce to its inputs. The main circularity burden is therefore evaluative rather than derivational. First, the main-table 'best results' are not clean tests of the memory: MM-Lifelong is run with GPT-5.2 while the ablations use GPT-5-mini, and Table 12 shows that swapping only the backbone moves day accuracy 18.75 to 23.50 and week 19.75 to 22.75, a day gain of 4.75 points comparable to the full-vs-flat gap and to the 6.75-point margin over ReMA. The paper itself concedes this in Limitations: 'because the main benchmarks use different reasoning backbones, component-level conclusions rely on the fixed-backbone ablations rather than absolute cross-benchmark results.' Second, on the forward-looking tracks of OVO-Bench and ESTP-Bench the evaluated questions and options are visible to the construction stage, and the caption prompt instructs the captioner to help disambiguate the options, which steers evidence collection toward the evaluated answers and partly manufactures the reported contextual/forward gains relative to baselines that lack this steering. Third, EgoServe, its service taxonomy, and the EgoMemo baseline come from the same lab (Gong et al. 2026), so the query-free proactive claim rests substantially on a same-lab benchmark, and the EgoServe skill-vs-content ablations coincide by construction for the perceptual-trace and pattern layers. None of these are hidden, and none make the fixed-backbone ablations circular; they bound the strength of the cross-benchmark superiority claims. No uniqueness theorem, ansatz-by-citation, or renamed-known-result pattern was found. On the strength of the clean fixed-backbone ablations and the transparency of the disclosed confounds, the paper is not circular in its

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

GROVE introduces no new physical or mathematical entities; its strata are software structures. The main ledger entries are hand-set hyperparameters and domain assumptions about LLM perception/consolidation reliability and benchmark validity.

free parameters (6)
  • top-k retrieval width = 20
    Selected from ablation on MM-Lifelong day/week/month (Table 18); the paper uses top-k=20 as the default.
  • max retrieval rounds T_max = 8/6/12 for day/week/month; 3/2 for OVO/StreamingBench
    Chosen by horizon and benchmark; inference-time analysis on the day split motivated the eight-round default but the appendix uses per-split values.
  • BM25 parameters = k1=1.5, b=0.75
    Standard retrieval hyperparameters listed in Table 10 without justification or sensitivity analysis.
  • Traversal fan-out = 3/8/15 for Pattern/Episode/Moment
    Hand-set values in Table 10; no ablation reported for these widths.
  • Moment target density = 5s/10s/20s depending on episode length
    Hand-set adaptive density thresholds in Table 10; no sensitivity analysis.
  • Episode cap and window length = 10min/30s on MM-Lifelong and EgoServe; 2min/10s elsewhere
    Chosen to balance construction cost and temporal resolution; no ablation on these values.
axioms (5)
  • domain assumption Frozen VLM perception faithfully captures all evidence needed for later QA and proactive services.
    Used throughout dual perception; the Limitations explicitly state that evidence missed during ingestion cannot be recovered.
  • domain assumption LLM-based episode segmentation, moment extraction, and pattern matching produce semantically correct structures.
    Equations 3-6 rely on LLM judgments; no independent verification of segment or pattern quality is provided.
  • domain assumption Benchmark LLM judges and task protocols are valid proxies for real assistant usefulness.
    MM-Lifelong uses a GPT-5 judge, EgoServe uses Macro-F1 plus an LLM quality score; no human correlation study is presented.
  • domain assumption Using future/task information where the benchmark protocol supplies it does not invalidate the comparison.
    Appendix A.1 states forward-looking tracks see candidate questions during construction, and EgoServe instructional subsets receive official task graphs; the authors argue this follows benchmark protocol.
  • domain assumption Temporal scale is the correct organizing interface for video memory.
    This is the paper's core design hypothesis; it is supported by ablations but not derived from prior theory.

pith-pipeline@v1.3.0-daily-deepseek · 47904 in / 9999 out tokens · 110813 ms · 2026-08-04T08:06:22.169223+00:00 · methodology

0 comments
read the original abstract

A wearable assistant should both answer questions about its visual history and recognize when that history is useful to the present situation. Existing video-memory systems primarily support question-conditioned recall, whereas proactive assistants typically use separate memory and control mechanisms. We introduce GROVE, a training-free framework that supports both behaviors with one memory grown causally from a continuous video stream. GROVE retains fine-grained perceptual evidence and incrementally consolidates it into time-stamped moments, coherent episodes, and recurring cross-day patterns. Each stratum is paired with a scale-native retrieval skill for locating an observation, replaying an activity, or traversing long-range regularities. Reactive QA and proactive assistance share this memory and access interface, differing in whether retrieval is initiated by a user query or the current situation. Across multiple benchmarks including the challenging MM-lifelong and EgoServe, GROVE achieves the best results among the compared methods. Controlled ablations show that the temporal strata and their access skills are complementary, with patterns providing the largest benefit when evidence spans multiple days. Code will be available at https://github.com/SitongGong/GROVE.

Figures

Figures reproduced from arXiv: 2608.02392 by Bo Zheng, Caixin Kang, Guo Chen, Huchuan Lu, Kaipeng Zhang, Sitong Gong, Tianyu Yan, Xiang Ruan, Yifei Huang, Yunzhi Zhuge.

Figure 1
Figure 1. Figure 1: Performance of GROVE across five video under [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of GROVE. (a) Memory construction proceeds in four incremental steps: dual perception, adaptive segmen [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of GROVE’s predictions. The first [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 3
Figure 3. Figure 3: Inference and construction cost. (a) Accuracy, total [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Temporally Stratified Memory Visualization of EgoServe Benchmark [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Temporally Stratified Memory Visualization of MMLifelong-Day [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Temporally Stratified Memory Visualization of MMLifelong-Month [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Reasoning and Skill Scale Calling Example on the MMLifelong (day). [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

17 extracted references · 12 linked inside Pith

  1. [1]

    Claude3.5Sonnet

    Anthropic.2024. Claude3.5Sonnet. https://www.anthropic. com/news/claude-3-5-sonnet. Accessed: 2026-07-29. Azad, S.; Vineet, V.; and Rawat, Y. S

    2024

  2. [4]

    Fu, C.; Lin, H.; Wang, X.; Shen, Y.; Liu, X.; Cao, H.; Long, Z.; Gao, H.; Li, K.; MA, L.; et al

    Nvidia nemotron nano v2 vl.arXiv preprint arXiv:2511.03929. Fu, C.; Lin, H.; Wang, X.; Shen, Y.; Liu, X.; Cao, H.; Long, Z.; Gao, H.; Li, K.; MA, L.; et al

    arXiv

  3. [5]

    arXiv preprint arXiv:2607.11523

    Vinci2: Provid- ing Proactive Assistance in Continuous Egocentric Videos. arXiv preprint arXiv:2607.11523. Hurst, A.; Lerer, A.; Goucher, A. P.; Perelman, A.; Ramesh, A.; Clark, A.; Ostrow, A.; Welihinda, A.; Hayes, A.; Rad- ford, A.; et al

    Pith/arXiv arXiv

  4. [7]

    Li, B.; Zhang, Y.; Guo, D.; Zhang, R.; Li, F.; Zhang, H.; Zhang, K.; Zhang, P.; Li, Y.; Liu, Z.; et al

    Visual Agen- tic Memory: Enabling Online Long Video Understanding via Online Indexing, Hierarchical Memory, and AgenticRe- trieval.arXiv preprint arXiv:2605.16481. Li, B.; Zhang, Y.; Guo, D.; Zhang, R.; Li, F.; Zhang, H.; Zhang, K.; Zhang, P.; Li, Y.; Liu, Z.; et al

    Pith/arXiv arXiv

  5. [8]

    Li,J.;Wu,C.-H.;Liu,Y.;Ding,K.;Li,J.;andZhang,C.2026

    Llava-onevision: Easy visual task transfer.arXiv preprint arXiv:2408.03326. Li,J.;Wu,C.-H.;Liu,Y.;Ding,K.;Li,J.;andZhang,C.2026. Bridging Modalities, Spanning Time: Structured Memory for Ultra-Long Agentic Video Reasoning.arXiv preprint arXiv:2605.08271. Liang,Z.;Li,J.;Chen,W.;Zhang,Y.;Lu,H.;andLi,G.2026. OASIS:On-DemandHierarchicalEventMemoryforStream- i...

    Pith/arXiv arXiv 2026

  6. [9]

    InICASSP 2026-2026 IEEE Interna- tional Conference on Acoustics, Speech and Signal Process- ing (ICASSP), 12147–12151

    Streaming- bench: Assessing the gap for mllms to achieve streaming video understanding. InICASSP 2026-2026 IEEE Interna- tional Conference on Acoustics, Speech and Signal Process- ing (ICASSP), 12147–12151. IEEE. Liu, X. B.; Fang, S.; Shi, W.; Wu, C.-S.; Igarashi, T.; and Chen, X. A. 2025a. Proactive conversational agents with inner thoughts. InProceeding...

    arXiv 2026

  7. [10]

    Ren, X.; Xu, L.; Xia, L.; Wang, S.; Yin, D.; and Huang, C

    Video-xl-2: Towards very long-video understanding through task-aware kv spar- sification.arXiv preprint arXiv:2506.19225. Ren, X.; Xu, L.; Xia, L.; Wang, S.; Yin, D.; and Huang, C

    Pith/arXiv arXiv

  8. [11]

    Song, E.; Chai, W.; Wang, G.; Zhang, Y.; Zhou, H.; Wu, F.; Chi, H.; Guo, X.; Ye, T.; Zhang, Y.; et al

    Openai gpt-5 system card.arXiv preprint arXiv:2601.03267. Song, E.; Chai, W.; Wang, G.; Zhang, Y.; Zhou, H.; Wu, F.; Chi, H.; Guo, X.; Ye, T.; Zhang, Y.; et al

    Pith/arXiv arXiv

  9. [12]

    5-omni technical report.arXiv preprint arXiv:2604.15804

    Qwen3. 5-omni technical report.arXiv preprint arXiv:2604.15804. Tulving,E.1972.Episodicandsemanticmemory.InTulving, E.; and Donaldson, W., eds.,Organization of Memory, 381–

    Pith/arXiv arXiv 1972

  10. [14]

    Yang,Z.;Wang,S.;Zhang,K.;Wu,K.;Leng,S.;Zhang,Y.; Li, B.; Qin, C.; Lu, S.; Li, X.; et al

    Streamagent: Towardsanticipatoryagentsforstreamingvideounderstand- ing.arXiv preprint arXiv:2508.01875. Yang,Z.;Wang,S.;Zhang,K.;Wu,K.;Leng,S.;Zhang,Y.; Li, B.; Qin, C.; Lu, S.; Li, X.; et al

    Pith/arXiv arXiv

  11. [15]

    In Proceedings of the 33rd ACM International Conference on Multimedia, 10807–10816

    Timechat-online: 80% visualtokensarenaturallyredundantinstreamingvideos. In Proceedings of the 33rd ACM International Conference on Multimedia, 10807–10816. Yao, Y.; Yu, T.; Zhang, A.; Wang, C.; Cui, J.; Zhu, H.; Cai, T.;Li,H.;Zhao,W.;He,Z.;etal.2024. Minicpm-v:Agpt-4v levelmllmonyourphone.arXivpreprintarXiv:2408.01800. Zeng, X.; Qiu, K.; Zhang, Q.; Li, X...

    Pith/arXiv arXiv 2024

  12. [16]

    Zhang, X.; Jia, Z.; Guo, Z.; Li, J.; Li, B.; Li, H.; and Lu, Y

    Flash-vstream: Memory-based real- time understanding for long video streams.arXiv preprint arXiv:2406.08085. Zhang, X.; Jia, Z.; Guo, Z.; Li, J.; Li, B.; Li, H.; and Lu, Y

    Pith/arXiv arXiv

  13. [17]

    Zhang, Y.; Dong, X

    Deep video discovery: Agentic search with tool use for long-form video understanding.Advances in Neural Information Processing Systems, 38: 89863–89895. Zhang, Y.; Dong, X. L.; Lin, Z.; Madotto, A.; Kumar, A.; Damavandi, B.; Chai, J.; and Moon, S. 2025a. Proac- tiveassistantdialoguegenerationfromstreamingegocentric videos. InProceedingsofthe2025Conference...

    arXiv

  14. [403]

    Academic Press. Wang,H.;Feng,B.;Lai,Z.;Xu,M.;Li,S.;Ge,W.;Dehghan, A.;Cao,M.;andHuang,P.2026.Streambridge:Turningyour offline video large language model into a proactive stream- ing assistant.Advances in Neural Information Processing Systems, 38: 132332–132359. Wang, P.; Bai, S.; Tan, S.; Wang, S.; Fan, Z.; Bai, J.; Chen, K.; Liu, X.; Wang, J.; Ge, W.; et ...

    Pith/arXiv arXiv 2026

  15. [2024]

    Gpt-4o system card.arXiv preprint arXiv:2410.21276. Li, A. Y.; Numan, N.; and Steed, A

    Pith/arXiv arXiv

  16. [2025]

    Chen,C.;Gan,G.;Ji,K.;Zhang,C.;Yang,Z.;Yao,G.;Chen, H.; Chen, J.; Yuan, Y.; and Shen, C

    Qwen3-vl technical report.arXiv preprint arXiv:2511.21631. Chen,C.;Gan,G.;Ji,K.;Zhang,C.;Yang,Z.;Yao,G.;Chen, H.; Chen, J.; Yuan, Y.; and Shen, C. 2026a. Memdreamer: Decoupling perception and reasoning for long video under- standingviahierarchicalgraphmemoryandagenticretrieval mechanism.arXiv preprint arXiv:2606.07512. Chen, G.; Li, Z.; Wang, S.; Jiang, J...

    Pith/arXiv arXiv

  17. [2026]

    arXiv preprint arXiv:2603.08620

    Streamready: Learningwhattoanswerandwheninlongstreamingvideos. arXiv preprint arXiv:2603.08620. Bai, S.; Cai, Y.; Chen, R.; Chen, K.; Chen, X.; Cheng, Z.; Deng, L.; Ding, W.; Gao, C.; Ge, C.; et al

    arXiv