Beyond the Timeline: Augmenting Long-Video Memory with Grounded Entity Biographies

ID: 61105b8a
原文链接: https://arxiv.org/abs/2609.38155
PDF: https://arxiv.org/pdf/2609.38155
作者: H, u, i, , R, e, n, ,, , L, e, i, , F, a, n, ,, , H, e, n, r, y, , P, a, o, ,, , H, a, n, , G, u, o, ,, , Z, e, e, s, h, a, n, , Z, i, a, ,, , Y, i, n, g, , C, h, e, n, ,, , A, l, e, x, a, n, d, e, r, , S, c, h, w, i, n, g, ,, , G, a, n, g, , H, u, a
日期: 2026-09-29
更新: 2026-09-29
分类: learning
来源类型: paper
标签: arxiv, video-memory, entity-grounding, long-video-qa, memory-consolidation
质量评分: 4/5
抓取时间: 2026-10-01T04:22:00+00:00

中文导读

针对长视频问答中同一物体跨小时/天的事件串联难题，提出 Grounded Entity Biographies（GEB）：在记忆构建阶段把跨片段的同一物理实例的视觉接地观察归组成可检索的实体生平，同时保留每个时刻的上下文问答时将生平与情节性证据一并检索，让模型能沿着记忆巩固时建立的身份链追踪实体解决了时序描述与文本实体无法解决物理身份（不同物体共享描述同一物体观察断开）的检索失效问题

为什么值得关注

长视频记忆新形态：按物理实例把跨片段观察归组成可检索实体生平，修补身份断链的检索失效

Grounded in the abstract: chronological descriptions and text-derived entities leave physical identity unresolved - different objects share descriptions while one object's observations stay disconnected - so retrieval does not recover the right events; GEB fixes this by grouping grounded observations per physical instance into retrievable biographies.

关键信息

论文标题：Beyond the Timeline: Augmenting Long-Video Memory with Grounded Entity Biographies
作者：H, u, i, , R, e, n, ,, , L, e, i, , F, a, n, ,, , H, e, n, r, y, , P, a, o, ,, , H, a, n, , G, u, o, ,, , Z, e, e, s, h, a, n, , Z, i, a, ,, , Y, i, n, g, , C, h, e, n, ,, , A, l, e, x, a, n, d, e, r, , S, c, h, w, i, n, g, ,, , G, a, n, g, , H, u, a
arXiv: https://arxiv.org/abs/2609.38155
发布时间：2026-09-29
arXiv 分类：c, s, ., C, V, ,, , c, s, ., A, I, ,, , c, s, ., C, L, ,, , c, s, ., I, R, ,, , c, s, ., L, G
关联标签：arxiv, video-memory, entity-grounding, long-video-qa, memory-consolidation

English Abstract

Answering questions about long videos often requires connecting events involving the same objects across hours or days. Chronological descriptions and text-derived entities can leave physical identity unresolved: different objects may share a description, while observations of the same object remain disconnected across events. Retrieving relevant events therefore does not necessarily recover the "biography" of the particular entity a question concerns. To address this, we introduce Grounded Entity Biographies (GEB), a long-video memory framework that groups visually grounded observations of the same physical instance across clips into retrievable biographies while preserving the context of each moment. During question answering, the biography is retrieved alongside episodic evidence, allowing the model to follow an entity through events using identity links established during memory construction. Evaluations across four benchmarks, including day-long and week-long recordings, demonstrate improvements over prior memory frameworks in both multiple-choice and open-ended question answering. On EgoLifeQA, GEB achieves 72.0% accuracy, 4.4 percentage points above the best published result. Ablations show that grounded identity association and biography reading both contribute to the gains, which additional descriptions alone do not fully recover.

English Summary

Targets long-video QA that requires connecting events involving the same physical object across hours or days. Grounded Entity Biographies (GEB) groups visually grounded observations of the same physical instance across clips into retrievable biographies while preserving per-moment context; at QA time the biography is retrieved alongside episodic evidence so the model can follow an entity via identity links established during memory consolidation, fixing retrieval failures where different objects share a description and observations of one object stay disconnected.

Obsidian Notes

内容由 opencli arxiv paper 拉取 arXiv 元数据与摘要生成。
中文导读与价值判断均锚定在条目已有摘要、论文摘要、作者、日期与分类信息上；未补充论文摘要之外的实验细节。