arxiv: 2605.09874 · v1 · submitted 2026-05-11 · 💻 cs.CV · cs.AI· cs.CL

Recognition: no theorem link

EgoMemReason: A Memory-Driven Reasoning Benchmark for Long-Horizon Egocentric Video Understanding

Ziyang Wang , Yue Zhang , Shoubin Yu , Ce Zhang , Zengqi Zhao , Jaehong Yoon , Hyunji Lee , Gedas Bertasius

show 1 more author

Mohit Bansal

Authors on Pith no claims yet

Pith reviewed 2026-05-12 04:32 UTC · model grok-4.3

classification 💻 cs.CV cs.AIcs.CL

keywords egocentric videolong-horizon memorymemory-driven reasoningvideo benchmarkmultimodal modelstemporal reasoningquestion answering

0 comments

The pith

Current AI models reach only 39.6 percent accuracy on memory-driven reasoning over week-long egocentric videos.

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

The paper presents a benchmark designed to measure how well AI systems integrate and reason over sparse visual information spread across days of continuous egocentric footage. It separates the problem into three memory demands: tracking how objects change state, recalling and sequencing distant events, and extracting repeated behavioral patterns from infrequent observations. Testing seventeen existing methods shows the strongest result at 39.6 percent overall accuracy, with clear drops as the time gap between relevant clips grows. This gap matters because future visual assistants and life-logging devices will need exactly this kind of long-range recall to be useful in everyday settings.

Core claim

EgoMemReason supplies 500 questions across entity memory, event memory, and behavior memory. Each question draws on an average of 5.1 video segments and requires backtracking through as much as 25.9 hours of footage. Evaluation of seventeen multimodal models and agent frameworks establishes that no current approach exceeds 39.6 percent accuracy. The three memory categories exhibit different failure modes, and accuracy declines steadily with longer temporal distances between evidence clips.

What carries the argument

EgoMemReason benchmark, organized around three memory types that each require accumulation, recall, and abstraction of evidence scattered across multiple days of egocentric video.

If this is right

Models will require new architectures that explicitly accumulate and index information over multi-day spans.
Performance falls as the interval between relevant evidence increases, so temporal modeling must scale beyond current context windows.
Entity, event, and behavior memory each break for different reasons, implying that progress needs targeted components rather than uniform scaling.
Practical always-on visual assistants cannot be deployed until accuracy on these tasks rises substantially.

Where Pith is reading between the lines

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

Success on this benchmark could directly improve reliability of personal AI systems that maintain ongoing context from a user's daily visual stream.
The same question format could be adapted to test memory in other continuous data streams such as audio or sensor logs.
Better long-horizon memory mechanisms would also raise new questions about selective forgetting and privacy management in always-recording devices.

Load-bearing premise

The benchmark questions force genuine cross-day memory use rather than being answerable from short clips or surface patterns alone.

What would settle it

A model that scores above 70 percent on the full set of questions while processing only short local segments without any explicit long-term storage would falsify the claim that long-horizon memory remains unsolved.

Figures

Figures reproduced from arXiv: 2605.09874 by Ce Zhang, Gedas Bertasius, Hyunji Lee, Jaehong Yoon, Mohit Bansal, Shoubin Yu, Yue Zhang, Zengqi Zhao, Ziyang Wang.

**Figure 1.** Figure 1: Illustration of our EGOMEMREASON for week-long egocentric video memory. Given a query at a specific time, answering requires retrieving and aggregating evidence from multiple temporally distant observations across days. We categorize memory into three types: entity memory (tracking persistent objects and states, tracking the same object for a long temo), event memory (ordering and linking events, e.g. link… view at source ↗

**Figure 2.** Figure 2: Comparison with existing weeklong video benchmarks. The x-axis shows the average number of distinct video segments needed to answer a question (i.e., evidence), and the y-axis shows temporal certification in hours (i.e., the total video duration one must search to locate all ground-truth evidence. Bubble size is proportional to the number of questions. As shown in [PITH_FULL_IMAGE:figures/full_fig_p002… view at source ↗

**Figure 3.** Figure 3: Overview of the six core challenges across three memory types in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

**Figure 4.** Figure 4: Overview of the 4-stage benchmark construction pipeline. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗

**Figure 5.** Figure 5: Dataset composition by memory type. EGOMEMREASON is formulated as a multiple-choice question-answering benchmark, with each question paired with one correct answer and several semantically competitive distractors. As shown in [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

**Figure 6.** Figure 6: Effect of input frame count on accuracy across memory types. Direct QA ICL CoT 30 Accuracy (%) 20 Entity Event Behavioral Overall 40 10 50 [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗

**Figure 8.** Figure 8: Human verification interface. Left: the annotator reviews the query-time context video and expandable evidence clips alongside the question and option set. Right: multidimensional quality assessment panel with structured ratings for query quality, correctchoice quality, option quality, and an overall accept/revise/reject decision. evidence beyond the pre-selected set. Annotators follow five guidelines: w… view at source ↗

**Figure 9.** Figure 9: Qualitative error analysis across three memory types. [PITH_FULL_IMAGE:figures/full_fig_p022_9.png] view at source ↗

read the original abstract

Next-generation visual assistants, such as smart glasses, embodied agents, and always-on life-logging systems, must reason over an entire day or more of continuous visual experience. In ultra-long video settings, relevant information is sparsely distributed across hours or days, making memory a fundamental challenge: models must accumulate information over time, recall prior states, track temporal order, and abstract recurring patterns. However, existing week-long video benchmarks are primarily designed for perception and recognition, such as moment localization or global summarization, rather than reasoning that requires integrating evidence across multiple days. To address this gap, we introduce EgoMemReason, a comprehensive benchmark that systematically evaluates week-long egocentric video understanding through memory-driven reasoning. EgoMemReason evaluates three complementary memory types: entity memory, tracking how object states evolve and change across days; event memory, recalling and ordering activities separated by hours or days; and behavior memory, abstracting recurring patterns from sparse, repeated observations over the whole week period. EgoMemReason comprises 500 questions across three memory types and six core challenges, with an average of 5.1 video segments of evidence per question and 25.9 hours of memory backtracking. We evaluate EgoMemReason on 17 methods across MLLMs and agentic frameworks, revealing that even the best model achieves only 39.6% overall accuracy. Further analysis shows that the three memory types fail for distinct reasons and that performance degrades as evidence spans longer temporal horizons, revealing that long-horizon memory remains far from solved. We believe EgoMemReason establishes a strong foundation for evaluating and advancing long-context, memory-aware multimodal systems.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

EgoMemReason is a new benchmark that shows current models top out at 39.6% on week-long egocentric memory reasoning, but its strength rests on unshown details about question design.

read the letter

The main takeaway is that this paper presents EgoMemReason, a benchmark for long-horizon memory reasoning in egocentric videos, and demonstrates that current models are still weak at it, with the best one at 39.6% overall accuracy. The new part is structuring the evaluation around three memory types—entity, event, and behavior—over week-long data. Existing benchmarks handle perception or summarization but not this kind of cross-day reasoning. They collected 500 questions with an average of 5.1 evidence segments and 25.9 hours of backtracking, then ran 17 methods on them. That setup lets them show how performance drops with longer time spans and that each memory type fails differently. The paper does well in identifying the gap and providing a concrete testbed. The numbers give a clear signal that more work is needed on memory accumulation and recall. Where it could be stronger is in the details of benchmark creation. The abstract mentions the design but skips how questions were generated, whether they were validated by multiple people, or if there are controls to prevent models from guessing via biases. If those are missing or weak in the full paper, the results might overstate the memory challenge. This work is for the video understanding and multimodal reasoning community. Researchers testing new long-context models or agents would get value from using or extending this benchmark. It is worth a serious referee because it tackles an important practical problem with a fresh evaluation approach. I would send it to peer review, asking the authors to expand on the question construction and any safeguards against shortcuts.

Referee Report

1 major / 2 minor

Summary. The paper introduces EgoMemReason, a benchmark for memory-driven reasoning over week-long egocentric videos. It defines three memory types (entity memory for object state changes, event memory for ordering activities across hours/days, and behavior memory for abstracting recurring patterns) across six core challenges, comprising 500 questions with an average of 5.1 evidence segments and 25.9 hours of backtracking per question. Evaluation of 17 MLLM and agentic methods shows a best-case accuracy of 39.6%, with further analysis indicating distinct failure modes per memory type and degradation over longer temporal spans, leading to the conclusion that long-horizon memory remains far from solved.

Significance. If the benchmark questions are confirmed to require genuine cross-day evidence integration, this work would provide a valuable new resource for evaluating memory capabilities in multimodal systems, addressing a gap left by existing week-long video benchmarks focused on perception or summarization. The categorization into complementary memory types, the scale of evidence backtracking, and the consistent low performance across diverse methods offer concrete directions for future model development in applications such as smart glasses and life-logging agents. The explicit reporting of average evidence segments and temporal horizons is a positive aspect of the design.

major comments (1)

[§3] §3 (Benchmark Construction): The manuscript provides no details on question generation process, validation that questions necessitate multi-day memory integration (vs. single-segment or superficial cues), inter-annotator agreement, or explicit controls for dataset biases and shortcuts. This is load-bearing for the central claim, as the reported 39.6% ceiling accuracy and the assertion that 'long-horizon memory remains far from solved' (abstract) rest on the 500 questions accurately measuring the three intended memory types without artifacts.

minor comments (2)

[Abstract] Abstract: Specify the identity of the single best model achieving 39.6% overall accuracy and include a brief per-memory-type breakdown to make the headline result more informative.
[§5] §5 (Experiments): Ensure the six core challenges are explicitly enumerated and linked to the three memory types with example questions for reader clarity.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their constructive feedback, which helps strengthen the manuscript. We address the major comment on benchmark construction below and will revise accordingly.

read point-by-point responses

Referee: [§3] §3 (Benchmark Construction): The manuscript provides no details on question generation process, validation that questions necessitate multi-day memory integration (vs. single-segment or superficial cues), inter-annotator agreement, or explicit controls for dataset biases and shortcuts. This is load-bearing for the central claim, as the reported 39.6% ceiling accuracy and the assertion that 'long-horizon memory remains far from solved' (abstract) rest on the 500 questions accurately measuring the three intended memory types without artifacts.

Authors: We agree that the original §3 provided insufficient detail on the construction pipeline, which is critical for validating the benchmark's claims. In the revised manuscript we will expand §3 with a dedicated subsection on question generation: questions were created by human annotators who first reviewed complete week-long timelines to identify sparse evidence spanning multiple days, then formulated queries requiring integration across those segments. We will add quantitative validation showing that >90% of questions cannot be solved from any single segment (via per-question evidence ablation) and that superficial cues were filtered through a multi-stage review. Inter-annotator agreement (Cohen's κ) for memory-type labeling and evidence-segment selection will be reported. Finally, we will include explicit bias controls, such as content diversity checks across subjects and environments plus shortcut analysis (e.g., lexical overlap or temporal heuristics). These additions directly substantiate that the 39.6% ceiling reflects genuine long-horizon memory demands rather than artifacts. revision: yes

Circularity Check

0 steps flagged

No significant circularity: independent benchmark with empirical evaluation only

full rationale

The paper introduces EgoMemReason as a new benchmark with 500 questions across three memory types, constructed from week-long egocentric videos with specified evidence requirements (average 5.1 segments, 25.9h backtracking). It then reports direct accuracy results from evaluating 17 existing MLLM and agentic methods, with the 39.6% top score presented as an empirical observation rather than a derived quantity. No equations, fitted parameters, self-definitional claims, or load-bearing self-citations appear in the abstract or described construction. The evaluation chain is self-contained: question design and model testing are independent of any prior author derivations, making the low-performance conclusion falsifiable on the benchmark itself without reduction to inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that the defined memory categories and question set faithfully represent real-world long-horizon reasoning needs; no free parameters are fitted and no new physical entities are postulated.

axioms (1)

domain assumption Entity memory, event memory, and behavior memory constitute the primary complementary challenges for long-horizon egocentric video reasoning.
The benchmark is explicitly organized around these three types as the core evaluation axes.

pith-pipeline@v0.9.0 · 5629 in / 1449 out tokens · 56884 ms · 2026-05-12T04:32:43.908341+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

120 extracted references · 120 canonical work pages · 6 internal anchors

[1]

Attention is All you Need , url =

Vaswani, Ashish and Shazeer, Noam and Parmar, Niki and Uszkoreit, Jakob and Jones, Llion and Gomez, Aidan N and Kaiser, ukasz and Polosukhin, Illia , booktitle =. Attention is All you Need , url =

work page
[2]

Gemini 3.1 Pro Model Card , year =

work page
[3]

2025 , eprint=

StreamingVLM: Real-Time Understanding for Infinite Video Streams , author=. 2025 , eprint=

work page 2025
[4]

Evaluating memory in llm agents via incremental multi-turn interactions.arXiv preprint arXiv:2507.05257, 2025

Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions , author=. arXiv preprint arXiv:2507.05257 , year=

work page arXiv
[5]

Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

Islam, Md Mohaiminul and Nagarajan, Tushar and Wang, Huiyu and Bertasius, Gedas and Torresani, Lorenzo , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

work page
[6]

Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

Islam, Md Mohaiminul and Ho, Ngan and Yang, Xitong and Nagarajan, Tushar and Torresani, Lorenzo and Bertasius, Gedas , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

work page
[7]

Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing,

A Simple LLM Framework for Long-Range Video Question-Answering , author=. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing,

work page 2024
[8]

arXiv preprint arXiv:2509.25911 , year=

Mem- : Learning Memory Construction via Reinforcement Learning , author=. arXiv preprint arXiv:2509.25911 , year=

work page arXiv
[9]

arXiv preprint arXiv:2506.13356 , year=

Storybench: A dynamic benchmark for evaluating long-term memory with multi turns , author=. arXiv preprint arXiv:2506.13356 , year=

work page arXiv
[10]

arXiv preprint arXiv:2603.07392 , year=

Can Large Language Models Keep Up? Benchmarking Online Adaptation to Continual Knowledge Streams , author=. arXiv preprint arXiv:2603.07392 , year=

work page arXiv
[11]

2024 , journal=

RULER: What's the Real Context Size of Your Long-Context Language Models? , author=. 2024 , journal=

work page 2024
[12]

Advances in Neural Information Processing Systems , volume=

Babilong: Testing the limits of llms with long context reasoning-in-a-haystack , author=. Advances in Neural Information Processing Systems , volume=

work page
[13]

arxiv , year =

Maharana, Adyasha and Lee, Dong-Ho and Tulyakov, Sergey and Bansal, Mohit and Barbieri, Francesco and Fang, Yuwei , title =. arxiv , year =

work page
[14]

2024 , eprint=

LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory , author=. 2024 , eprint=

work page 2024
[15]

2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

Video-MME: The First-Ever Comprehensive Evaluation Benchmark of Multi-modal LLMs in Video Analysis , author=. 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

work page 2025
[16]

ArXiv , year=

LVBench: An Extreme Long Video Understanding Benchmark , author=. ArXiv , year=

work page
[17]

2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

MLVU: Benchmarking Multi-task Long Video Understanding , author=. 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

work page 2025
[18]

ArXiv , year=

LongVideoBench: A Benchmark for Long-context Interleaved Video-Language Understanding , author=. ArXiv , year=

work page
[19]

arXiv preprint arXiv:2510.23981 , year=

TeleEgo: Benchmarking Egocentric AI Assistants in the Wild , author=. arXiv preprint arXiv:2510.23981 , year=

work page arXiv
[20]

2026 , eprint=

Towards Multimodal Lifelong Understanding: A Dataset and Agentic Baseline , author=. 2026 , eprint=

work page 2026
[21]

Organization of Memory , editor =

Tulving, Endel , title =. Organization of Memory , editor =. 1972 , pages =

work page 1972
[22]

, author=

The reviewing of object files: object-specific integration of information. , author=. Cognitive psychology , year=

work page
[23]

ArXiv , year=

MemGPT: Towards LLMs as Operating Systems , author=. ArXiv , year=

work page
[24]

Yuxia Wang, Minghan Wang, Muhammad Arslan Man- zoor, Fei Liu, Georgi Nenkov Georgiev, Rocktim Jy- oti Das, and Preslav Nakov

Retrieval-augmented generation with conflicting evidence , author=. arXiv preprint arXiv:2504.13079 , year=

work page arXiv
[25]

CORG: Generating answers from complex, interrelated contexts , author=. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , pages=

work page 2025
[26]

Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: Long papers) , pages=

When not to trust language models: Investigating effectiveness of parametric and non-parametric memories , author=. Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: Long papers) , pages=

work page
[27]

Advances in Neural Information Processing Systems , year=

HourVideo: 1-Hour Video-Language Understanding , author=. Advances in Neural Information Processing Systems , year=

work page
[28]

Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =

Nagrani, Arsha and Menon, Sachit and Iscen, Ahmet and Buch, Shyamal and Mehran, Ramin and Jha, Nilpa and Hauth, Anja and Zhu, Yukun and Vondrick, Carl and Sirotenko, Mikhail and Schmid, Cordelia and Weyand, Tobias , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =. 2025 , pages =

work page 2025
[29]

Deep video dis- covery: Agentic search with tool use for long-form video understanding.arXiv preprint arXiv:2505.18079, 2025

Deep Video Discovery: Agentic Search with Tool Use for Long-form Video Understanding , author=. arXiv preprint arXiv:2505.18079 , year=

work page arXiv
[30]

Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =

Wang, Weihan and He, Zehai and Hong, Wenyi and Cheng, Yean and Zhang, Xiaohan and Qi, Ji and Ding, Ming and Gu, Xiaotao and Huang, Shiyu and Xu, Bin and Dong, Yuxiao and Tang, Jie , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =. 2025 , pages =

work page 2025
[31]

Cambrian-S: Towards Spatial Supersensing in Video.arXiv preprint arXiv:2511.04670, 2025

Cambrian-S: Towards Spatial Supersensing in Video , author=. arXiv preprint arXiv:2511.04670 , year=

work page arXiv
[32]

Adversarial NLI : A New Benchmark for Natural Language Understanding

Nie, Yixin and Williams, Adina and Dinan, Emily and Bansal, Mohit and Weston, Jason and Kiela, Douwe. Adversarial NLI : A New Benchmark for Natural Language Understanding. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. 2020. doi:10.18653/v1/2020.acl-main.441

work page doi:10.18653/v1/2020.acl-main.441 2020
[33]

Movie Facts and Fibs (MF \^

Zaranis, Emmanouil and Farinhas, Ant. Movie Facts and Fibs (MF \^. arXiv preprint arXiv:2506.06275 , year=

work page arXiv
[34]

arXiv preprint arXiv:2505.22657 , year=

3dllm-mem: Long-term spatial-temporal memory for embodied 3d large language model , author=. arXiv preprint arXiv:2505.22657 , year=

work page arXiv
[35]

Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

3D-mem: 3D scene memory for embodied exploration and reasoning , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

work page
[36]

Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology (UIST) , year =

Generative Agents: Interactive Simulacra of Human Behavior , author =. Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology (UIST) , year =

work page
[37]

ACM Transactions on Information Systems , year =

A Survey on the Memory Mechanism of Large Language Model based Agents , author =. ACM Transactions on Information Systems , year =

work page
[38]

Tan, Haoran and Zhang, Zeyu and Ma, Chen and Chen, Xu and Dai, Quanyu and Dong, Zhenhua , booktitle =

work page
[39]

Psychological Review , volume =

Why there are complementary learning systems in the hippocampus and neocortex: Insights from the successes and failures of connectionist models of learning and memory , author =. Psychological Review , volume =. 1995 , doi =

work page 1995
[40]

, journal =

Kumaran, Dharshan and Hassabis, Demis and McClelland, James L. , journal =. What learning systems do intelligent agents need?. 2016 , doi =

work page 2016
[41]

Philosophical Transactions of the Royal Society B , volume =

Distinct mechanisms and functions of episodic memory , author =. Philosophical Transactions of the Royal Society B , volume =. 2024 , doi =

work page 2024
[42]

2025 , issue_date =

A Survey on the Memory Mechanism of Large Language Model based Agents , author =. ACM Transactions on Information Systems , year =. doi:10.1145/3748302 , note =

work page doi:10.1145/3748302
[43]

From Human Memory to

Liu, Shichun and others , journal =. From Human Memory to. 2025 , url =

work page 2025
[44]

Memory in large language models: Mechanisms, eval- uation and evolution.arXiv preprint arXiv:2509.18868, 2025a

Memory in Large Language Models: Mechanisms, Evaluation and Evolution , author =. arXiv preprint arXiv:2509.18868 , year =

work page arXiv
[45]

2025 , url =

Liang, Jiafeng and Li, Hao and Li, Chang and Zhou, Jiaqi and Jiang, Shixin and Wang, Zekun and Ji, Changkai and Zhu, Zhihao and Liu, Runxuan and Ren, Tao and Fu, Jinlan and Ng, See-Kiong and Liang, Xia and Liu, Ming and Qin, Bing , journal =. 2025 , url =

work page 2025
[46]

Advances in Neural Information Processing Systems (NeurIPS) , volume =

Guti. Advances in Neural Information Processing Systems (NeurIPS) , volume =

work page
[47]

Proceedings of the AAAI Conference on Artificial Intelligence , volume =

A Machine with Short-Term, Episodic, and Semantic Memory Systems , author =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. 2023 , doi =

work page 2023
[48]

Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

Movieqa: Understanding stories in movies through question-answering , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

work page
[49]

Proceedings of the 2018 conference on empirical methods in natural language processing , pages=

Tvqa: Localized, compositional video question answering , author=. Proceedings of the 2018 conference on empirical methods in natural language processing , pages=

work page 2018
[50]

Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

Tgif-qa: Toward spatio-temporal reasoning in visual question answering , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

work page
[51]

ICCV , year=

TALL: Temporal Activity Localization via Language Query , author=. ICCV , year=

work page
[52]

ICCV , year=

Dense-Captioning Events in Videos , author=. ICCV , year=

work page
[53]

Proceedings of the 58th annual meeting of the association for computational linguistics , pages=

Tvqa+: Spatio-temporal grounding for video question answering , author=. Proceedings of the 58th annual meeting of the association for computational linguistics , pages=

work page
[54]

Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

Next-qa: Next phase of question-answering to explaining temporal actions , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

work page
[55]

STAR: A Benchmark for Situ- ated Reasoning in Real-World Videos.arXiv e-prints, art

Star: A benchmark for situated reasoning in real-world videos , author=. arXiv preprint arXiv:2405.09711 , year=

work page arXiv
[56]

Icml , volume=

Is space-time attention all you need for video understanding? , author=. Icml , volume=

work page
[57]

Proceedings of the IEEE/CVF international conference on computer vision , pages=

Vivit: A video vision transformer , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

work page
[58]

Proceedings of the IEEE/CVF international conference on computer vision , pages=

Multiscale vision transformers , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

work page
[59]

Proceedings of the IEEE/CVF international conference on computer vision , pages=

Slowfast networks for video recognition , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

work page
[60]

Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

Long-term feature banks for detailed video understanding , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

work page
[61]

World Models

World models , author=. arXiv preprint arXiv:1803.10122 , volume=

work page internal anchor Pith review Pith/arXiv arXiv
[62]

MA-EgoQA: Question answer- ing over egocentric videos from multiple embodied agents.arXiv preprint arXiv:2603.09827, 2026

MA-EgoQA: Question Answering over Egocentric Videos from Multiple Embodied Agents , author=. arXiv preprint arXiv:2603.09827 , year=

work page arXiv
[63]

The eleventh international conference on learning representations , year=

React: Synergizing reasoning and acting in language models , author=. The eleventh international conference on learning representations , year=

work page
[64]

Advances in neural information processing systems , volume=

Reflexion: Language agents with verbal reinforcement learning , author=. Advances in neural information processing systems , volume=

work page
[65]

Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

Episodic memory question answering , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

work page
[66]

arXiv preprint arXiv:2510.12422 , year=

VideoLucy: Deep Memory Backtracking for Long Video Understanding , author=. arXiv preprint arXiv:2510.12422 , year=

work page arXiv
[67]

Transactions of the association for computational linguistics , volume=

Lost in the middle: How language models use long contexts , author=. Transactions of the association for computational linguistics , volume=

work page
[68]

Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

Ma-lmm: Memory-augmented large multimodal model for long-term video understanding , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

work page
[69]

Advances in Neural Information Processing Systems , volume=

Streaming long video understanding with large language models , author=. Advances in Neural Information Processing Systems , volume=

work page
[70]

Seeing, listening, remem- bering, and reasoning: A multimodal agent with long-term memory.arXiv:2508.09736, 2025

Seeing, listening, remembering, and reasoning: A multimodal agent with long-term memory , author=. arXiv preprint arXiv:2508.09736 , year=

work page arXiv
[71]

arXiv preprint arXiv:2601.03515 , year=

Mem-gallery: Benchmarking multimodal long-term conversational memory for mllm agents , author=. arXiv preprint arXiv:2601.03515 , year=

work page arXiv
[72]

Ryoo, and Juan Carlos Niebles

Active Video Perception: Iterative Evidence Seeking for Agentic Long Video Understanding , author=. arXiv preprint arXiv:2512.05774 , year=

work page arXiv
[73]

arXiv preprint arXiv:2510.09608 , year=

Streamingvlm: Real-time understanding for infinite video streams , author=. arXiv preprint arXiv:2510.09608 , year=

work page arXiv
[74]

arXiv preprint arXiv:2512.04540 , year=

VideoMem: Enhancing Ultra-Long Video Understanding via Adaptive Memory Management , author=. arXiv preprint arXiv:2512.04540 , year=

work page arXiv
[75]

arXiv preprint arXiv:2510.04428 , year=

AIR: Enabling Adaptive, Iterative, and Reasoning-based Frame Selection For Video Question Answering , author=. arXiv preprint arXiv:2510.04428 , year=

work page arXiv
[76]

arXiv preprint arXiv:2603.05484 , year=

Towards Multimodal Lifelong Understanding: A Dataset and Agentic Baseline , author=. arXiv preprint arXiv:2603.05484 , year=

work page arXiv
[77]

Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

Adaptive keyframe sampling for long video understanding , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

work page
[78]

European Conference on Computer Vision , pages=

Videoagent: Long-form video understanding with large language model as agent , author=. European Conference on Computer Vision , pages=. 2024 , organization=

work page 2024
[79]

European Conference on Computer Vision , pages=

Videoagent: A memory-augmented multimodal agent for video understanding , author=. European Conference on Computer Vision , pages=. 2024 , organization=

work page 2024
[80]

Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

Moviechat: From dense token to sparse memory for long video understanding , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

work page

Showing first 80 references.