Paper deep dive
ObjectStream: Latent Objects as Memory Anchors for Streaming Video Understanding
Mingkang Dong, Muxin Pu, Jie Li, Bohan Guo, Songruo Chen, Bin Ren, Xu Zheng, Chen Zhao, Tianwen Qian, Mohamed Elhoseiny, Yuqian Fu
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 8/4/2026, 3:45:11 AM
Summary
ObjectStream is a training-free framework for streaming video understanding that uses latent objects as memory anchors. It organizes visual evidence into persistent object histories, transient changes, and recent context, improving efficiency and accuracy on benchmarks like OVO-Bench and StreamingBench without modifying the underlying Video-LLM.
Entities (8)
Relation Signals (7)
ObjectStream → uses → Latent Object-Anchored Memory
confidence 95% · ObjectStream organizes the incoming stream through a Latent Object-Anchored Memory.
ObjectStream → uses → Object-Conditioned Temporal Residuals
confidence 95% · Object-Conditioned Temporal Residuals preserve short-lived but important evidence around abrupt object-level changes.
ObjectStream → uses → Recent Visual Grounding Window
confidence 95% · the Recent Visual Grounding Window retains the latest observations for timely visual grounding.
ObjectStream → improves → Qwen2.5-VL-7B
confidence 92% · ObjectStream improves Qwen2.5-VL-7B by 10.0 points on OVO-Bench Real-Time Visual Perception
Latent Object-Anchored Memory → discovers → latent objects
confidence 90% · Latent Object-Anchored Memory discovers and tracks latent object anchors and stores their persistent evidence
ObjectStream → evaluatedon → OVO-Bench
confidence 90% · Extensive experiments on online streaming and offline long-video benchmarks demonstrate both effectiveness and efficiency.
ObjectStream → evaluatedon → StreamingBench
confidence 90% · improves ... 2.9 points on ... StreamingBench
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Streaming video understanding requires models to continuously retain useful visual evidence before future questions are known. Existing approaches primarily manage the growing visual context according to token importance, temporal redundancy, or segment-level relevance, but rarely organize evidence around objects that persist and evolve over time. Thus, in this paper, we introduce ObjectStream, a training-free framework that treats latent objects as memory anchors for streaming video understanding. ObjectStream induces spatially coherent latent objects directly from frozen Video-LLM representations, links them across frames into persistent anchors, and maintains their histories under a bounded memory budget, without requiring external object detectors or segmentation models. Built on these anchors, ObjectStream preserves three complementary forms of evidence: persistent object histories, transient object changes, and recent visual context. This design enables existing Video Large Language Models (Video-LLMs) to reason over object identities, interactions, and state changes while leaving the underlying model unchanged. Extensive experiments on online streaming and offline long-video benchmarks demonstrate both effectiveness and efficiency. In online streaming evaluation, ObjectStream improves Qwen2.5-VL-7B by 10.0 points on OVO-Bench Real-Time Visual Perception, while reducing peak GPU mem-ory and TTFT by approximately 50%. On offline long-video benchmarks, it surpasses the full-token baseline while discarding 82.5% of visual tokens. These results highlight latent objects as a practical and effective organizing principle for compact streaming video memory.
Tags
Links
- Source: https://arxiv.org/abs/2607.28312v2
- Canonical: https://arxiv.org/abs/2607.28312v2
Trouble viewing inline? Open PDF directly →
Full Text
49,056 characters extracted from source content.
Expand or collapse full text
ObjectStream: Latent Objects as Memory Anchors for Streaming Video Understanding Mingkang Dong1 , Muxin Pu2 , Jie Li3, Bohan Guo1, Songruo Chen4, Bin Ren5, Xu Zheng6, Chen Zhao7, Tianwen Qian8, Mohamed Elhoseiny7, Yuqian Fu 7 Abstract Streaming video understanding requires models to continuously retain useful visual evidence before future questions are known. Existing approaches primarily manage the growing visual context according to token importance, temporal redundancy, or segment-level relevance, but rarely organize evidence around objects that persist and evolve over time. Thus, in this paper, we introduce ObjectStream, a training-free framework that treats latent objects as memory anchors for streaming video understanding. ObjectStream induces spatially coherent latent objects directly from frozen Video-LLM representations, links them across frames into persistent anchors, and maintains their histories under a bounded memory budget, without requiring external object detectors or segmentation models. Built on these anchors, ObjectStream preserves three complementary forms of evidence: persistent object histories, transient object changes, and recent visual context. This design enables existing Video Large Language Models (Video-LLMs) to reason over object identities, interactions, and state changes while leaving the underlying model unchanged. Extensive experiments on online streaming and offline long-video benchmarks demonstrate both effectiveness and efficiency. In online streaming evaluation, ObjectStream improves Qwen2.5-VL-7B by 10.0 points on OVO-Bench Real-Time Visual Perception, while reducing peak GPU memory and TTFT by approximately 50%. On offline long-video benchmarks, it surpasses the full-token baseline while discarding 82.5% of visual tokens. These results highlight latent objects as a practical and effective organizing principle for compact streaming video memory. Code is available at https://github.com/DMK041218/ObjectStream. Figure 1: Comparison between existing streaming memory and ObjectStream. Existing methods compress streaming observations into coarse latent representations, whereas ObjectStream organizes memory around latent objects to preserve fine-grained object evidence for long-term reasoning. Introduction Streaming video understanding [27, 53, 1, 42] requires models to process continuously arriving visual inputs and respond to questions that may be issued at any time. This capability is essential for applications such as live monitoring [9], autonomous systems [16, 2, 54], and egocentric vision and wearable assistant [12, 31, 49, 55, 45, 48, 17, 13, 30], and embodied agents and robotic systems [41, 56, 8, 38, 37, 22, 23, 40, 21]. Unlike offline video understanding, where the complete video can be revisited after a question is given, streaming models must decide what information to retain before future queries are known, while operating under strict memory and latency constraints. Recent studies [43, 19, 51, 15] have extended Video Large Language Models (Video-LLMs) to streaming scenarios from various perspectives, including online adaptation [33, 26], streaming-oriented inference [50, 46], and memory-based context management [44, 19, 51, 14, 43]. Among them, training-free approaches [44, 4, 19] are particularly attractive because they can adapt existing offline Video-LLMs without additional training or modification to the underlying model. These methods typically control the growing visual context through frame selection [34], visual-token pruning [44], KV-cache compression [7, 5], and temporal or segment-level memory organization [19, 51]. By reducing redundant historical information and prioritizing potentially relevant evidence over long video streams, they improve memory efficiency and response latency, facilitating the transition of Video-LLMs from offline video understanding to practical online streaming. However, deciding which visual units or temporal portions to retain addresses only part of the streaming memory problem; how the retained evidence is organized is equally important for subsequent reasoning. While existing designs are effective at preserving coarse-grained contextual and temporal semantics, finer-grained object-level information is often less explicitly organized across time. This limitation is particularly problematic for questions that require tracing an object’s identity, interactions, or state changes. These observations motivate an object-centric organization of streaming visual memory. Based on this insight, we introduce ObjectStream, a training-free and plug-and-play visual memory framework that treats latent objects as memory anchors for streaming Video-LLMs. Rather than managing historical visual tokens or KV states in isolation, ObjectStream organizes the incoming stream through a Latent Object-Anchored Memory. Specifically, it induces spatially coherent object candidates directly from frozen Video-LLM representations, associates them across frames, and maintains their evolving histories under a bounded memory budget. Since these object-like units emerge entirely from the model’s latent feature space, without requiring external object detectors or segmentation models, we refer to them as latent objects. The resulting cross-frame object states provide a compact semantic index over the stream, enabling the model to preserve object-related evidence beyond locally salient visual content. To complement persistent object histories, ObjectStream introduces two additional components. Object-Conditioned Temporal Residuals preserve short-lived but important evidence around abrupt object-level changes, while the Recent Visual Grounding Window retains the latest observations for timely visual grounding. Together, these components capture three complementary forms of memory: persistent object histories, transient object changes, and recent visual context, while leaving the underlying Video-LLM unchanged. We validate ObjectStream on both online streaming and offline long-video understanding benchmarks. It improves Qwen2.5-VL-7B by 10.0, 4.9, and 2.9 points on OVO-Bench Real-Time Visual Perception, OVO-Bench Backward Tracing, and StreamingBench, respectively. It also surpasses the full-token baseline on offline long-video benchmarks while discarding 82.5% of visual tokens. These results well demonstrate the effectiveness and token efficiency of ObjectStream. Our main contributions are summarized as follows: • We propose ObjectStream, a training-free and plug-and-play visual memory framework for efficient streaming video understanding, without architectural modifications. • We introduce Latent Object-Anchored Memory, using latent objects as persistent anchors to organize long-term visual evidence without external detectors or segmenters. • We complement persistent object memory with Object-Conditioned Temporal Residuals and a Recent Visual Grounding Window, achieving strong performance and token efficiency across online streaming and offline video understanding benchmarks. Related Work Video Large Language Models. Many Video Large Language Models (Video-LLMs) have been proposed to address video-language alignment, temporal modeling, and long-context reasoning. Representative proprietary models include Gemini 1.5 Pro [35] and GPT-4o [29], while widely used open-source models include LLaVA-Video [53], Qwen2-VL [39], InternVL2 [6], and LongVU [33]. Despite their strong performance, these models primarily follow an offline inference paradigm, assuming that the complete video is available before question answering. This assumption limits their direct applicability to the more realistic and challenging streaming setting, where visual inputs arrive continuously and future observations are unavailable. Streaming Video-LLMs and Visual Memory. Existing streaming video understanding models explore either training-based adaptation or training-free memory management. Training-based methods include VideoLLM-Online [3], which introduces online video instruction tuning for continuous video perception and interaction; Dispider [32], which enables active real-time interaction by disentangling perception, decision, and reaction; and ThinkStream [25], which adopts a Watch–Think–Speak paradigm and learns compressed semantic memory for long-horizon streaming. ViSpeak [11] and StreamForest [47] further improve online video understanding through streaming-oriented datasets, training objectives, and model adaptations. These methods generally require additional optimization or model-specific modifications. Training-free approaches instead preserve the underlying Video-LLM and manage historical visual information directly at inference time. StreamChat [24] maintains external visual memory to incorporate previous observations, while ReKV and StreamKV [7, 5] reduce memory overhead through KV-cache retention and compression. More recent methods explore structured memory management: FluxMem [44] introduces hierarchical visual memory to balance long-term history and recent observations; QueryStream [52] performs query-aware token pruning to retain request-relevant evidence; and OASIS [19] organizes streaming history into event-level structures for on-demand retrieval. However, these methods do not explicitly maintain persistent latent object representations across frames, limiting their ability to preserve object identities and state evolution over long video streams. Method Figure 2: Overview of ObjectStream. It comprises three memory modules: Latent Object-Anchored Memory discovers and tracks latent object anchors and stores their persistent evidence; Object-Conditioned Temporal Residuals preserve transient changes; and the Recent Visual Grounding Window retains recent observations for immediate grounding. Overview. Our design follows three core objectives: maintaining persistent object histories, preserving transient object-level changes, and retaining recent visual context for real-time grounding. The overall framework of our proposed ObjectStream is illustrated in Fig. 2 and consists of three components. First, Latent Object-Anchored Memory discovers spatially coherent latent object candidates from frozen Video-LLM representations, associates them across frames, and maintains their evolving histories under a bounded memory budget. Second, Object-Conditioned Temporal Residuals preserve short-lived evidence around abrupt object-level changes. Third, the Recent Visual Grounding Window retains the latest observations for timely question answering. Formally, let τ=Itt=1τV_τ=\I_t\_t=1^τ denote the video stream observed up to the current timestamp τ. ObjectStream processes each frame incrementally upon arrival. At timestamp t≤τt≤τ, the newly observed frame ItI_t is encoded into a set of patch-level visual tokens: t=xt,jj=1NX_t=\x_t,j\_j=1^N, where N is the number of visual tokens extracted from frame ItI_t, and each token xt,j∈ℝdx_t,j ^d is a d-dimensional visual representation. After processing all observations up to timestamp τ, ObjectStream maintains the retained visual memory as: ℛτ=ℛobjectτ∪ℛtemporalτ∪ℛrvgwτ,R^τ=R^τ_object ^τ_temporal ^τ_rvgw, (1) where ℛobjectτR^τ_object, ℛtemporalτR^τ_temporal, and ℛrvgwτR^τ_rvgw are produced by the aforementioned modules, retaining features associated with persistent object histories, transient object-level changes, and recent visual context, respectively. Latent Object-Anchored Memory To use latent objects as memory anchors for object-centric evidence retention, this module comprises four stages, denoted as S1–S4: S1: Query-Agnostic Token Saliency Estimation, S2: Spatial Clustering for Latent Object Anchor Discovery, S3: Cross-Frame Association and Bounded Object-Set Maintenance and S4: Object-Centric Evidence Retention. We describe each stage below. Figure 3: Illustration of latent object discovery. ObjectStream discovers latent object candidates directly from frozen visual tokens by estimating query-agnostic saliency and spatially clustering salient regions. The resulting latent objects provide the semantic object anchors for memory construction. S1: Query-Agnostic Token Saliency Estimation. For each frame ItI_t, we first summarize its overall visual content using a frame-level context vector ctc_t. Specifically, we ℓ2 _2-normalize each visual token and aggregate the normalized tokens into a global context representation : x^t,j x_t,j =xt,j‖xt,j‖2, = x_t,j\|x_t,j\|_2, (2) ct c_t =∑j=1Nx^t,j‖∑j=1Nx^t,j‖2. = _j=1^N x_t,j \| _j=1^N x_t,j \|_2. The context vector ctc_t captures the dominant visual semantics of frame ItI_t. We then estimate the saliency of each token by its cosine dissimilarity from this frame-level context: st,j=1−cos(x^t,j,ct)=1−x^t,j⊤ct.s_t,j=1-cos( x_t,j,c_t)=1- x_t,j c_t. (3) The score st,js_t,j measures how visually distinctive token xt,jx_t,j is within the current frame. Tokens that deviate more from the global frame representation are more likely to correspond to informative objects, persons, or regions. We then select active token candidates using a per-frame quantile threshold: t=j∣st,j≥Quantileq(t),A_t= \j s_t,j _q (S_t ) \, (4) where t=st,jj=1NS_t=\s_t,_j\_j=1^N is the set of saliency scores for all visual tokens in frame t, and Quantileq(t)Quantile_q(S_t) denotes the q-th quantile of saliency scores within frame t; thus larger q selects fewer, more salient tokens as active object candidates. S2: Spatial Clustering for Latent Object Anchor Discovery. As shown in Fig. 3, the active tokens indexed by tA_t are still patch-level candidates and may be spatially fragmented. To form coherent object-level regions, ObjectStream exploits the 2D patch-grid structure of visual tokens. Let pt,jp_t,j denote the spatial position of token xt,jx_t,j. We build an undirected graph tG_t over active tokens, where edges connect spatially adjacent tokens: t _t =(t,ℰt), =(A_t,E_t), (5) (j,k)∈ℰt (j,k) _t ⟺j,k∈t∧pt,k∈(pt,j), j,k _t p_t,k (p_t,j), t _t =C(t). =C(G_t). Here, ℰtE_t is the edge set, 4(pt,j)N_4(p_t,j) denotes the 4-connected spatial neighborhood of patch pt,jp_t,j on the patch grid, including its upper, lower, left, and right neighbors. C(⋅)C(·) denotes connected components, and t=Ct,mm=1MtC_t=\C_t,m\_m=1^M_t is the set of frame-level object candidates. μ=f(Ct,m)=∑j∈Ct,mst,jxt,j‖∑j∈Ct,mst,jxt,j‖2.μ=f(C_t,m)= _j∈ C_t,ms_t,jx_t,j \| _j∈ C_t,ms_t,jx_t,j \|_2. (6) This transforms scattered salient patch tokens into coarse object-level units without requiring external detectors or segmentation annotations. S3: Cross-Frame Association and Bounded Object-Set Maintenance. After obtaining frame-level object candidates, ObjectStream associates them across frames to build persistent object tracks. We maintain a dynamic object set =OiO=\O_i\, where each object track OiO_i stores an object feature μi _i, the last observed timestamp τi _i, a temporal consistency score ρi _i, an object saliency score aia_i, and a historical token set iT_i. The temporal consistency score ρi _i is computed as the cosine similarity between the track feature and the candidate feature at its most recent match, indicating the confidence of cross-frame association. Upon each successful match, the temporal consistency score ρi _i is updated as the cosine similarity. The object saliency score aia_i is inherited from the token-level saliency introduced in S2 by running averaging the saliency values of the visual tokens belonging to the corresponding object candidate. For each frame-level candidate Ct,mC_t,m, ObjectStream first finds the most similar existing object track: i⋆=argmaxicos(f(Ct,m),μi).i = _i (f(C_t,m), _i ). (7) If cos(f(Ct,m),μi⋆t−1)≥θ (f(C_t,m),μ^t-1_i )≥θ, the candidate is assigned to Oi⋆O_i and the object feature is updated by Exponential Moving Average (EMA): μi∗t=βμi∗t−1+(1−β)f(Ct,m)‖βμi∗t−1+(1−β)f(Ct,m)‖2. _i^*^t= β _i^*^t-1+(1-β)f(C_t,m) \|β _i^*^t-1+(1-β)f(C_t,m) \|_2. (8) If no existing object satisfies the matching threshold, ObjectStream initializes a new object track. For a matched object, we set τi∗ _i^* to the current timestamp, update its temporal consistency ρi∗ _i^* and average saliency ai∗a_i^*, and insert all tokens in the matched component into i∗T_i^*. Since streaming videos continuously introduce new objects, the object set must remain bounded. When the number of maintained tracks exceeds the budget M, ObjectStream evicts low-priority objects according to recency, saliency, and temporal consistency. rank(Oi) (O_i) =(τi,ai,ρi), =( _i,\ a_i, _i), (9) =Top-MOi∈(rank(Oi)). =Top -M_O_i (rank(O_i) ). The ranking is applied in lexicographic order, prioritizing object anchors that are recently observed, visually salient, and temporally consistent. S4: Object-Centric Evidence Retention. Once persistent object tracks in the set O are maintained under a fixed budget, ObjectStream constructs the long-term object-centric evidence ℛobjectτR^τ_object by selecting representative tokens from each retained object anchor. Object eviction and token retention operate at different granularities: the bounded object set O determines which object tracks are maintained up to timestamp τ, while token retention is performed over the historical token pool iτT_i^τ of each retained track Oi∈O_i . For each retained object OiO_i, its historical token pool contains token indices (t,j)(t,j) assigned to this object up to timestamp τ. Instead of globally selecting top-scoring tokens from the whole video, which may be dominated by a few salient regions or frames, ObjectStream allocates token retention capacity at the object level: ℐiτ ^τ_i =Top-K(t,j)∈iτ,t≤τ(st,j,ki), =Top -K_(t,j) ^τ_i,\;t≤τ (s_t,j,k_i ), (10) ℛobjectτ ^τ_object =⋃Oi∈xt,j∣(t,j)∈ℐiτ. = _O_i \x_t,j (t,j) ^τ_i \. The operator Top-KTop -K ranks tokens within the object-specific pool iτT_i^τ by their saliency scores st,js_t,j and returns the top-kik_i token indices. The retained object memory ℛobjectτR^τ_object then collects the corresponding visual tokens from all maintained object tracks, forming a compact yet comprehensive representation of persistent visual histories. Object-Conditioned Temporal Residuals Latent Object-Anchored memory preserves persistent visual semantics by smoothing object representations over time. However, this smoothing may suppress short-lived but important changes, such as motion, action transitions, object appearance or disappearance, state changes, and interaction changes. Thus, this module is further proposed to detect abrupt object-level feature changes and preserve additional raw tokens around the corresponding timestamps. For each object track OiO_i, we denote its matched sequence of historical observations as zi,ll=1Li\z_i,l\_l=1^L_i. Specifically, if the frame-level object candidate f(Ct,m)f(C_t,m) is successfully assigned to track OiO_i at a timestamp t, this observation is mapped as zi,l=f(Ct,m)z_i,l=f(C_t,m), where l indexes the chronologically ordered observation sequence up to length LiL_i, where 2≤l≤Li2≤ l≤ L_i, and ti,l=t_i,l=t preserves its corresponding timestamp. We then compute the feature change between neighboring observations and use an object-specific adaptive threshold to detect temporal events: Δi,l _i,l =1−cos(zi,l,zi,l−1), =1- (z_i,l,z_i,l-1), (11) γi,l _i,l =mean(Δi,kk=2l)+std(Δi,kk=2l), =mean (\ _i,k\_k=2^l )+std (\ _i,k\_k=2^l ), ℬτ ^τ =(i,ti,l)∣Δi,l>γi,ti,l≤τ. = \(i,t_i,l) _i,l> _i,\;t_i,l≤τ \. Here, Δi,l _i,l measures the temporal change of object OiO_i between two consecutive matched observations, γi,l _i,l is its adaptive temporal threshold, and ℬτB^τ denotes the temporal events detected up to timestamp τ. For each temporal event (i,t)∈ℬτ(i,t) ^τ, ObjectStream retrieves the original visual tokens assigned to the same object within a local temporal window (t,r)=u∣t−r≤u≤t+r,u≤τW(t,r)=\u t-r≤ u≤ t+r,\;u≤τ\, where r denotes the temporal window frame offset that defines the local neighborhood around event timestamp t. Within this object-specific temporal neighborhood, we select token indices according to their saliency scores: ℐi,tτ ^τ_i,t =TopK(u,j)∈iτ,u∈(t,r)(su,j,kb), =TopK_(u,j) ^τ_i,\;u (t,r) (s_u,j,k_b ), (12) ℛtemporalτ ^τ_temporal =⋃(i,t)∈ℬτxu,j∣(u,j)∈ℐi,tτ. = _(i,t) ^τ \x_u,j (u,j) ^τ_i,t \. In Eq.12 Top-KTop -K returns the indices of the top-kbk_b tokens ranked by saliency within the local temporal neighborhood, and ℛtemporalτR^τ_temporal stores the corresponding raw visual tokens. Unlike object-level representative tokens, temporal residual tokens preserve raw visual evidence around rapid object-level changes. Therefore, they complement the smooth long-term object memory and improve the ability to reason about short-term temporal events. Method Frames OVO-Bench Real-Time BT StreamingBench OCR ACR ATR STU FPD OJR Avg. Avg. OP CR CS ATP EU TR PR SU ACP CT Avg. Proprietary Models Gemini 1.5 Pro 1 fps 85.9 67.0 79.3 58.4 63.4 62.0 69.3 62.5 79.0 80.5 83.5 79.7 80.0 84.7 77.8 64.2 72.0 48.7 75.7 GPT-4o 64 69.8 64.2 71.6 51.1 70.3 59.8 64.5 60.8 77.1 80.5 83.9 76.5 70.2 83.8 66.7 62.2 69.1 49.2 73.3 Open-source Offline MLLMs LLaVA-Video 64 69.8 59.6 66.4 50.6 72.3 61.4 63.3 41.7 – – – – – – – – – – – Qwen2-VL 64 69.1 53.2 63.8 50.6 66.3 60.9 60.7 48.6 – – – – – – – – – – – InternVL2 64 68.5 58.7 69.0 44.9 67.3 56.0 60.7 44.0 68.1 60.9 69.4 77.1 67.7 62.9 59.3 53.3 55.0 56.5 63.7 LongVU 1 fps 55.7 49.5 59.5 48.3 68.3 63.0 57.4 39.5 – – – – – – – – – – – Open-source Online MLLMs (Training-Based) VideoLLM-Online-8B 2 fps 8.1 23.9 12.1 14.0 45.5 21.2 20.8 17.7 39.1 40.1 34.5 31.1 46.0 32.4 31.5 34.2 42.5 27.9 36.0 Dispider-7B 1 fps 57.7 49.5 62.1 44.9 61.4 51.6 54.6 36.1 74.9 75.5 74.1 73.1 74.4 59.9 76.1 62.9 62.2 45.8 67.6 Flash-VStream-7B 1 fps 25.5 32.1 29.3 33.7 29.7 28.8 29.9 25.4 25.9 43.6 24.9 23.9 27.3 13.1 18.5 25.2 23.9 48.7 23.2 ViSpeak 1 fps 75.2 58.7 71.6 51.1 74.3 66.9 66.3 57.5 79.8 88.3 83.3 81.1 76.4 75.1 70.4 65.9 77.3 34.2 74.4 ThinkStream 1 fps 85.2 64.2 69.8 49.4 69.3 64.1 67.0 52.3 – – – – – – – – – – – TimeChat-Online-7B 1 fps 75.2 46.8 70.7 47.8 69.3 61.4 61.9 41.7 80.8 79.7 80.8 83.3 74.8 78.8 78.7 64.2 68.8 58.0 75.3 StreamForest-7B 1 fps 68.5 53.2 71.6 47.8 65.4 60.9 61.2 52.0 83.1 82.8 82.7 84.3 77.5 78.2 76.9 69.1 75.6 54.4 77.3 Open-source Online MLLMs (Training-Free) Qwen2.5-VL-3B† 1 fps 76.5 44.0 67.2 42.1 68.3 62.0 60.0 42.0 76.2 68.8 75.4 79.2 73.0 72.3 71.3 61.4 71.6 26.1 68.0 + FluxMem 1 fps 83.2 56.9 67.2 47.8 68.3 63.6 64.5 42.1 72.5 73.8 73.8 79.4 72.6 79.7 76.9 64.2 65.1 42.6 70.9 + ObjectStream (Ours) 1 fps 87.3 68.8 73.3 54.5 66.3 65.2 68.6 42.5 72.7 69.8 79.8 80.4 73.9 82.8 65.7 71.1 72.0 34.6 72.4 Qwen2.5-VL-7B† 1 fps 79.2 53.2 67.2 51.7 71.3 57.1 63.3 44.6 78.3 80.5 79.8 82.4 75.5 80.4 74.1 62.6 67.6 51.1 73.9 + FluxMem 1 fps 81.2 59.6 70.7 53.4 75.2 63.0 67.2 46.8 80.2 81.1 81.4 85.3 78.0 83.8 80.6 65.9 69.6 52.1 76.4 + QueryStream 1 fps 75.2 49.5 69.8 50.0 71.3 62.5 63.1 44.9 82.4 84.4 79.2 82.4 78.0 81.3 78.7 65.0 69.3 47.3 75.3 + OASIS – 85.2 72.5 66.4 52.3 67.3 64.7 64.7 52.6 – – – – – – – – – – 70.6 + ObjectStream (Ours) 1 fps 89.9 70.6 74.1 59.0 75.3 70.7 73.3 49.5 74.1 80.2 88.6 82.6 77.1 85.0 82.4 73.6 73.2 44.2 76.8 Table 1: Performance comparison on OVO-Bench and StreamingBench. For OVO-Bench, we report real-time visual perception scores and the average backward tracing score. For StreamingBench, we report real-time category scores and the average score. Bold indicates the best result, and underlining indicates the second-best result. † indicates the backbone model. Component VME Ego OVO Streaming RVGW LOAM OCTR Long Schema Bench Bench ✓ 52.3 58.6 62.3 73.9 ✓ 51.1 58.3 62.0 74.8 ✓ ✓ 53.4 59.4 60.6 76.5 ✓ ✓ 53.8 60.4 62.1 74.9 ✓ ✓ ✓ 54.0 60.8 65.4 76.8 Table 2: Component ablation of ObjectStream. Model #Frames VideoMME-L EgoSchema Proprietary MLLMs Gemini 1.5 Pro 1 fps 62.5 69.3 GPT-4o 64 60.7 64.5 Open-source Offline MLLMs LLaVA-Video-7B 32 – 57.3 Qwen2-VL-7B 64 – 66.7 LongVU-7B 1 fps 60.6 58.2 Open-source Online MLLMs TimeChat-Online-7B 1 fps 48.4 61.9 Dispider-7B 1 fps 49.7 55.6 Vista 1 fps – 58.7 Training-free Offline-to-Online Methods Qwen2.5-VL-3B† 1 fps 49.3 54.9 + FluxMem 1 fps 51.2 57.8 + ObjectStream (Ours) 1 fps 52.9 57.7 Qwen2.5-VL-7B† 1 fps 51.3 58.5 + FluxMem 1 fps 53.9 60.1 + QueryStream 1 fps 52.9 – + ObjectStream (Ours) 1 fps 54.0 60.8 Table 3: Performance comparison (%) on offline benchmarks. Recent Visual Grounding Window Although object memory and temporal residuals preserve long-term semantics and short-term dynamics, streaming questions may also require the latest visual state. For example, questions about what is currently happening or which object is now visible depend strongly on the most recent frames. To support such real-time questions, ObjectStream maintains a lightweight Recent Visual Grounding Window (rvgw), which directly preserves raw visual tokens from the latest L frames: ℛrvgwτ=⋃t=T−L+1Tt.R_rvgw^τ= _t=T-L+1^TX_t. (13) The rvgw complements the compressed object memory by providing high-fidelity local evidence for the current visual state, while the object and temporal residual memories preserve compact historical context. Inference. Given a textual question Q, the language model generates the answer by conditioning on the retained visual memory: Y^=LLM(Q,ℛτ). Y=LLM(Q,R^τ). (14) By integrating persistent object histories, transient object-level changes, and recent visual context, ObjectStream provides the underlying Video-LLM with a compact yet expressive representation of the observed stream. This unified memory supports both historical reasoning and grounding in the current visual state under a bounded visual-token budget. Experiments Implementation Details. For all experiments, we build our method on Qwen2.5-VL-3B and 7B [1]. For online benchmarks, we sample videos at 1 fps and set the maximum video length to 256 frames. We use 8 frames as the Recent Visual Grounding Window (rvgw) for real-time question answering and compress older frames into object-centric memory. We keep up to 64 objects with 8 tokens per object, and further enable temporal residuals with at most 32 events, radius 1, and extra 2 tokens per selected frame. For offline benchmarks, we limit the maximum sequence length to 1024 frames. We use a larger object memory budget to preserve longer-range temporal context, keeping up to 512 objects with 80 tokens per object. We preserve the most recent 80 frames as the rvgw, while compressing older visual content into object-centric long-term memory. Benchmarks. We evaluate our method on both online and offline benchmarks. For online evaluation, we use StreamingBench [20] and OVO-Bench [18]. StreamingBench evaluates streaming video question answering under causal access, requiring the model to answer questions incrementally as video frames arrive over time. OVO-Bench focuses on online video understanding with an emphasis on real-time reasoning, covering abilities such as object perception, temporal reasoning, causal understanding, and event-based decision making under constrained memory. For offline evaluation, we use EgoSchema [28] and VideoMME-Long [10]. These benchmarks evaluate holistic video understanding when the entire video is available, with a particular focus on long-range temporal reasoning, egocentric understanding, and comprehensive multi-step video QA. Results on Online Benchmark. We compare ObjectStream with strong proprietary models, open-source offline Video-LLMs, training-based online Video-LLMs, and training-free online methods, including Gemini 1.5 Pro [35], GPT-4o [29], LLaVA-Video [53], Qwen2-VL [39], InternVL2 [6], LongVU [33], VideoLLM-Online [3], Dispider [32], Flash-VStream [50], ViSpeak [11], ThinkStream [25], TimeChat-Online [36], StreamForest [47], QueryStream [52], OASIS [19] and FluxMem [44]. Main Results on Streaming Benchmarks. As shown in Tab. 1, ObjectStream consistently improves streaming video understanding across online benchmarks without fine-tuning or architectural modification, demonstrating its effectiveness as a lightweight and plug-and-play visual memory mechanism for Video-LLMs. On OVO-Bench, ObjectStream achieves substantial improvements over Qwen2.5-VL baselines. With the 7B backbone, it improves the real-time visual perception score from 63.3 to 73.3 (+10.0) and the Backward Tracing score from 44.6 to 49.5, indicating better preservation of historical visual cues. Compared with FluxMem, ObjectStream further improves the overall score from 60.4 to 65.4 on the 7B backbone and from 54.0 to 60.2 on the 3B backbone; On StreamingBench, ObjectStream improves the average score from 73.9 to 76.8 with the 7B backbone and from 68.0 to 72.4 with the 3B backbone. These results demonstrate that latent object memory construction provides a general and effective solution for online perception and temporal reasoning. Effect of Memory Components. To evaluate the contributions of Recent Visual Grounding Window (RVGW), Latent Object-Anchored Memory (LOAM), and Object-Conditioned Temporal Residuals (OCTR), we progressively enable each module. As shown in Tab. 2, RVGW mainly improves StreamingBench by preserving recent visual observations, while LOAM enhances OVO-Bench through structured object memory. Combining the two consistently improves performance across both streaming and offline settings. Adding OCTR further preserves abrupt object changes, resulting in the best overall performance. Extension to Offline Long-video Benchmarks. We also evaluate ObjectStream on offline long-video benchmarks to verify whether the proposed memory design remains effective in the offline scenario. As shown in Tab. 3, ObjectStream consistently improves the original Qwen2.5-VL baselines on both VideoMME-Long and EgoSchema. With Qwen2.5-VL-7B, our method improves VideoMME-Long from 51.3 to 54.0 and EgoSchema from 58.5 to 60.8, achieving gains of 2.7 and 2.3 points, respectively. With Qwen2.5-VL-3B, ObjectStream also improves VideoMME-Long from 49.3 to 52.9 and EgoSchema from 54.9 to 57.7, showing that the benefit is consistent across model scales. These results demonstrate that ObjectStream is not limited to online streaming inference. Figure 4: Qualitative analysis of latent object discovery. ObjectStream maintains task-relevant object states across consecutive observations, while SAM+DINOv2 produces finer masks but less consistent temporal associations. Object Memory Construction and Efficiency Analysis. We further analyze different object memory construction strategies under the same Qwen2.5-VL-7B backbone on OVO-Bench. As shown in Tab. 4, Patch Top-K directly selects salient patch tokens without explicit object modeling. Although it keeps memory usage low, it yields lower performance and much higher TTFT than our method, showing that unstructured token selection is insufficient for streaming visual memory. DINOv2 Clustering introduces external patch features for object memory construction but brings additional memory and latency overhead while still underperforming our method. SAM+DINOv2 achieves the highest accuracy by using segmentation-based regions, yet it requires expensive segmentation and feature extraction, resulting in substantially higher latency and memory cost. In contrast, ObjectStream discovers objects directly in the backbone visual token space and combines them with temporal retention. It achieves 73.3 performance with only 19.89 GB peak memory and 6.697s TTFT, offering a better balance between accuracy and efficiency. Fig. 4 also illustrates this difference qualitatively. Although SAM+DINOv2 produces finer object masks, its outputs are optimized for pixel-level segmentation and do not directly provide a temporally reusable memory state for VLM reasoning. ObjectStream produces coarser but sufficiently localized object regions and maintains task-relevant objects across consecutive observations despite changes in scale, pose, and camera viewpoint. These results suggest that ObjectStream does not aim to replace segmentation models; instead, it provides a lightweight and temporally grounded object construction mechanism that better satisfies the low-latency and memory-bounded requirements of streaming video understanding. Variant Peak Mem↓ TTFT↓ TPOT↓ Perf↑ Reference Baselines Qwen2.5-VL-7B 40.50GB 13.431s 0.336s 64.2 + FluxMem 25.78GB 6.641s 0.252s 67.7 object Construction Variants Patch Top-K 19.89GB 16.223s 0.320s 72.8 DINOv2 Cluster 23.09GB 18.120s 1.230s 72.3 SAM+DINOv2 26.78GB 19.060s 0.785s 73.8 Latent Objects (Ours) 19.89GB 6.697s 0.240s 73.3 Table 4: Analysis on different object memory construction strategies with Qwen2.5-VL-7B as backbone. Conclusion To conclude, in this paper, we propose ObjectStream, a training-free object-centric memory mechanism for efficient streaming video understanding. By discovering object-latent anchors and linking them into persistent tracks, ObjectStream builds a compact semantic object related memory over incoming video streams. This design preserves persistent visual evidence while remaining sensitive to dynamic changes and the latest visual state, enabling effective reasoning under memory and latency constraints. Experiments on both streaming and offline long-video benchmarks show that ObjectStream substantially improves online and offline video understanding while maintaining strong robustness under limited token budgets. References [1] S. Bai, K. Chen, X. Liu, J. Wang, W. Ge, S. Song, K. Dang, P. Wang, S. Wang, J. Tang, H. Zhong, Y. Zhu, M. Yang, Z. Li, J. Wan, P. Wang, W. Ding, Z. Fu, Y. Xu, J. Ye, X. Zhang, T. Xie, Z. Cheng, H. Zhang, Z. Yang, H. Xu, and J. Lin (2025) Qwen2.5-vl technical report. External Links: 2502.13923, Link Cited by: Introduction, Experiments. [2] T. Brödermann, C. Sakaridis, Y. Fu, and L. Van Gool (2024) Cafuser: condition-aware multimodal fusion for robust semantic perception of driving scenes. arXiv preprint arXiv:2410.10791. Cited by: Introduction. [3] J. Chen, Z. Lv, S. Wu, K. Q. Lin, C. Song, D. Gao, J. Liu, Z. Gao, D. Mao, and M. Z. Shou (2024) VideoLLM-online: online video large language model for streaming video. External Links: 2406.11816, Link Cited by: Related Work, Experiments. [4] X. Chen, K. Tao, K. Shao, and H. Wang (2026) StreamingTOM: streaming token compression for efficient video understanding. External Links: 2510.18269, Link Cited by: Introduction. [5] Y. Chen, X. Bai, Z. Wang, C. Bai, Y. Dai, M. Lu, and S. Zhang (2025) StreamKV: streaming video question-answering with segment-based kv cache retrieval and compression. External Links: 2511.07278, Link Cited by: Introduction, Related Work. [6] Z. Chen, W. Wang, Y. Cao, Y. Liu, Z. Gao, E. Cui, J. Zhu, S. Ye, H. Tian, Z. Liu, L. Gu, X. Wang, Q. Li, Y. Ren, Z. Chen, J. Luo, J. Wang, T. Jiang, B. Wang, C. He, B. Shi, X. Zhang, H. Lv, Y. Wang, W. Shao, P. Chu, Z. Tu, T. He, Z. Wu, H. Deng, J. Ge, K. Chen, K. Zhang, L. Wang, M. Dou, L. Lu, X. Zhu, T. Lu, D. Lin, Y. Qiao, J. Dai, and W. Wang (2025) Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. External Links: 2412.05271, Link Cited by: Related Work, Experiments. [7] S. Di, Z. Yu, G. Zhang, H. Li, T. Zhong, H. Cheng, B. Li, W. He, F. Shu, and H. Jiang (2025) Streaming video question-answering with in-context video kv-cache retrieval. External Links: 2503.00540, Link Cited by: Introduction, Related Work. [8] Y. Du, T. Lin, Z. Zhong, R. Li, X. Chen, J. Liu, C. Liu, Y. Chen, Y. Fu, and B. Zhao (2026) Focusable monocular depth estimation. External Links: 2605.11756, Link Cited by: Introduction. [9] E. Dumitru and S. Spînu (2026) A multi-task deep learning framework for real-time intelligent video surveillance with temporal event validation. External Links: 2607.03131, Link Cited by: Introduction. [10] C. Fu, Y. Dai, Y. Luo, L. Li, S. Ren, R. Zhang, Z. Wang, C. Zhou, Y. Shen, M. Zhang, P. Chen, Y. Li, S. Lin, S. Zhao, K. Li, T. Xu, X. Zheng, E. Chen, C. Shan, R. He, and X. Sun (2025) Video-mme: the first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. External Links: 2405.21075, Link Cited by: Experiments. [11] S. Fu, Q. Yang, Y. Li, Y. Peng, K. Lin, X. Wei, J. Hu, X. Xie, and W. Zheng (2025) ViSpeak: visual instruction feedback in streaming videos. External Links: 2503.12769, Link Cited by: Related Work, Experiments. [12] Y. Fu, C. Wang, Y. Fu, Y. Wang, C. Bai, X. Xue, and Y. Jiang (2019) Embodied one-shot video recognition: learning from actions of a virtual embodied agent. In Proceedings of the 27th ACM international conference on multimedia, Cited by: Introduction. [13] Y. Fu, R. Wang, B. Ren, G. Sun, B. Gong, Y. Fu, D. P. Paudel, X. Huang, and L. Van Gool (2025) Objectrelator: enabling cross-view object relation understanding across ego-centric and exo-centric perspectives. In Proceedings of the IEEE/CVF International Conference on Computer Vision, Cited by: Introduction. [14] H. Ge, Y. Wang, H. Wu, and Y. Cai (2026) What should a streaming video model remember?. External Links: 2606.16353, Link Cited by: Introduction. [15] Y. Guan, L. Yin, D. Liang, J. Ju, Z. Luo, J. Luan, Y. Liu, and X. Bai (2026) Video streaming thinking: videollms can watch and think simultaneously. External Links: 2603.12262, Link Cited by: Introduction. [16] F. Li, C. Zhang, and G. Chen (2026) Sparse-aware vector quantization for bandwidth-efficient collaborative 3d semantic occupancy prediction. External Links: 2607.01928, Link Cited by: Introduction. [17] Y. Li, Y. Fu, T. Qian, Q. Xu, S. Dai, D. P. Paudel, L. Van Gool, and X. Wang (2026) Egocross: benchmarking multimodal large language models for cross-domain egocentric video question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, Cited by: Introduction. [18] Y. Li, J. Niu, Z. Miao, C. Ge, Y. Zhou, Q. He, X. Dong, H. Duan, S. Ding, R. Qian, P. Zhang, Y. Zang, Y. Cao, C. He, and J. Wang (2025) OVO-bench: how far is your video-llms from real-world online video understanding?. External Links: 2501.05510, Link Cited by: Experiments. [19] Z. Liang, J. Li, W. Chen, Y. Zhang, H. Lu, and G. Li (2026) OASIS: on-demand hierarchical event memory for streaming video reasoning. External Links: 2604.17052, Link Cited by: Introduction, Related Work, Experiments. [20] J. Lin, Z. Fang, C. Chen, Z. Wan, F. Luo, P. Li, Y. Liu, and M. Sun (2024) StreamingBench: assessing the gap for mllms to achieve streaming video understanding. External Links: 2411.03628, Link Cited by: Experiments. [21] T. Lin, Y. Du, J. Liu, N. Zhu, Y. Li, Y. Fu, Y. Chen, H. Cai, Z. Ye, B. Cheng, et al. (2026) Evo-depth: a lightweight depth-enhanced vision-language-action model. arXiv preprint arXiv:2605.14950. Cited by: Introduction. [22] T. Lin, Y. Du, Y. Mao, Z. Ye, Y. Zhong, B. Cheng, Y. Wang, J. Liu, Y. Tian, J. Yan, et al. (2026) LA4VLA: learning to act without seeing via language-action pretraining. arXiv preprint arXiv:2606.27295. Cited by: Introduction. [23] T. Lin, Y. Zhong, Y. Du, J. Zhang, J. Liu, Y. Chen, E. Gu, Z. Liu, H. Cai, Y. Zou, et al. (2025) Evo-1: lightweight vision-language-action model with preserved semantic alignment. arXiv preprint arXiv:2511.04555. Cited by: Introduction. [24] J. Liu, Z. Yu, S. Lan, S. Wang, R. Fang, J. Kautz, H. Li, and J. M. Alvare (2025) StreamChat: chatting with streaming video. External Links: 2412.08646, Link Cited by: Related Work. [25] Z. Liu, L. Guo, H. Li, R. Zhen, X. He, R. Ji, X. Ren, Y. Zhang, H. Lu, and J. Liu (2026) Thinking in streaming video. External Links: 2603.12938, Link Cited by: Related Work, Experiments. [26] X. Lu, Y. Bo, J. Chen, S. Li, X. Guo, H. Guan, F. Liu, D. Xu, P. Sun, H. Sun, R. Liu, and H. Li (2026) AURA: always-on understanding and real-time assistance via video streams. External Links: 2604.04184, Link Cited by: Introduction. [27] M. Maaz, H. Rasheed, S. Khan, and F. S. Khan (2024) Video-chatgpt: towards detailed video understanding via large vision and language models. External Links: 2306.05424, Link Cited by: Introduction. [28] K. Mangalam, R. Akshulakov, and J. Malik (2023) EgoSchema: a diagnostic benchmark for very long-form video language understanding. External Links: 2308.09126, Link Cited by: Experiments. [29] OpenAI:, Y. Dai, and Y. Malkov (2024) GPT-4o system card. External Links: 2410.21276, Link Cited by: Related Work, Experiments. [30] J. Pan, R. Wang, T. Qian, M. Mahdi, Y. Fu, X. Xue, X. Huang, L. Van Gool, D. P. Paudel, and Y. Fu (2025) V2-sam: marrying sam2 with multi-prompt experts for cross-view object correspondence. arXiv preprint arXiv:2511.20886. Cited by: Introduction. [31] C. Plizzari, G. Goletto, A. Furnari, S. Bansal, F. Ragusa, G. M. Farinella, D. Damen, and T. Tommasi (2024) An outlook into the future of egocentric vision. External Links: 2308.07123, Link Cited by: Introduction. [32] R. Qian, S. Ding, X. Dong, P. Zhang, Y. Zang, Y. Cao, D. Lin, and J. Wang (2025) Dispider: enabling video llms with active real-time interaction via disentangled perception, decision, and reaction. External Links: 2501.03218, Link Cited by: Related Work, Experiments. [33] X. Shen, Y. Xiong, C. Zhao, L. Wu, J. Chen, C. Zhu, Z. Liu, F. Xiao, B. Varadarajan, F. Bordes, Z. Liu, H. Xu, H. J. Kim, B. Soran, R. Krishnamoorthi, M. Elhoseiny, and V. Chandra (2024) LongVU: spatiotemporal adaptive compression for long video-language understanding. External Links: 2410.17434, Link Cited by: Introduction, Related Work, Experiments. [34] Y. Shen, S. Tian, J. Yang, and Z. Liu (2026) A simple baseline for streaming video understanding. External Links: 2604.02317, Link Cited by: Introduction. [35] G. Team and P. G. O. Vinyals (2024) Gemini 1.5: unlocking multimodal understanding across millions of tokens of context. External Links: 2403.05530, Link Cited by: Related Work, Experiments. [36] TimeChat-online: 80. Cited by: Experiments. [37] K. Wang, K. Fan, Y. Fu, S. Lin, H. Luo, D. Seita, Y. Fu, Y. Jiang, and X. Xue (2026) Ocra: object-centric learning with 3d and tactile priors for human-to-robot action transfer. arXiv preprint arXiv:2603.14401. Cited by: Introduction. [38] K. Wang, K. Fan, C. Qiu, Z. Shangguan, Y. Fu, Y. Fu, D. Seita, and X. Xue (2026) Oflow: injecting object-aware temporal flow matching for robust robotic manipulation. arXiv preprint arXiv:2604.17876. Cited by: Introduction. [39] P. Wang, S. Bai, S. Tan, S. Wang, Z. Fan, J. Bai, K. Chen, X. Liu, J. Wang, W. Ge, Y. Fan, K. Dang, M. Du, X. Ren, R. Men, D. Liu, C. Zhou, J. Zhou, and J. Lin (2024) Qwen2-vl: enhancing vision-language model’s perception of the world at any resolution. External Links: 2409.12191, Link Cited by: Related Work, Experiments. [40] R. Wang, Y. Fu, Y. Li, T. Lin, T. Qian, M. Elhoseiny, B. Zhao, Y. Fu, Y. Jiang, and X. Xue (2026) Afford-vla: action-aligned visual planning via internalized affordance. arXiv preprint arXiv:2605.24203. Cited by: Introduction. [41] X. Wang, T. Kwon, M. Rad, B. Pan, I. Chakraborty, S. Andrist, D. Bohus, A. Feniello, B. Tekin, F. V. Frujeri, N. Joshi, and M. Pollefeys (2023) HoloAssist: an egocentric human interaction dataset for interactive ai assistants in the real world. External Links: 2309.17024, Link Cited by: Introduction. [42] Y. Wang, Z. Li, T. Qian, H. Zheng, Z. Wang, Y. Fu, and X. Wang (2025) StreamEQA: towards streaming video understanding for embodied scenarios. arXiv preprint arXiv:2512.04451. Cited by: Introduction. [43] H. Wu, S. M. Mathews, Y. Cai, M. Yang, and Y. Wang (2026) Semantic-aware adaptive visual memory for streaming video understanding. External Links: 2605.07897, Link Cited by: Introduction. [44] Y. Xie, B. He, J. Wang, X. Zheng, Z. Ye, and Z. Wu (2026) FluxMem: adaptive hierarchical memory for streaming video understanding. External Links: 2603.02096, Link Cited by: Introduction, Related Work, Experiments. [45] Q. Xu, T. Qian, Y. Fu, K. Li, Y. Jiao, J. Zhang, X. Wang, and L. He (2025) ToG-bench: task-oriented spatio-temporal grounding in egocentric videos. arXiv preprint arXiv:2512.03666. Cited by: Introduction. [46] R. Xu, G. Xiao, Y. Chen, L. He, Y. Lu, and S. Han (2026) StreamingVLM: real-time understanding for infinite video streams. External Links: 2510.09608, Link Cited by: Introduction. [47] X. Zeng, K. Qiu, Q. Zhang, X. Li, J. Wang, J. Li, Z. Yan, K. Tian, M. Tian, X. Zhao, Y. Wang, and L. Wang (2025) StreamForest: efficient online video understanding with persistent event memory. External Links: 2509.24871, Link Cited by: Related Work, Experiments. [48] D. Zhang, Y. Fu, R. Yang, Y. Miao, T. Qian, X. Zheng, G. Sun, A. Chhatkuli, X. Huang, Y. Jiang, et al. (2025) Egonight: towards egocentric vision understanding at night with a challenging benchmark. arXiv preprint arXiv:2510.06218. Cited by: Introduction. [49] G. Zhang, D. Zhang, L. Duan, G. Han, Y. Fu, D. P. Paudel, L. V. Gool, and E. Vollenweider (2026) EgoSpot:egocentric multimodal control for hands-free mobile manipulation. External Links: 2306.02393, Link Cited by: Introduction. [50] H. Zhang, Y. Wang, Y. Tang, Y. Liu, J. Feng, J. Dai, and X. Jin (2024) Flash-vstream: memory-based real-time understanding for long video streams. External Links: 2406.08085, Link Cited by: Introduction, Experiments. [51] H. Zhang, S. Yang, J. Fu, S. Ng, and X. Qiu (2026) HERMES: kv cache as hierarchical memory for efficient streaming video understanding. External Links: 2601.14724, Link Cited by: Introduction. [52] K. Zhang, Z. Yang, B. Wang, S. Qian, and C. Xu (2026) QueryStream: advancing streaming video understanding with query-aware pruning and proactive response. In The Fourteenth International Conference on Learning Representations, External Links: Link Cited by: Related Work, Experiments. [53] Y. Zhang, J. Wu, W. Li, B. Li, Z. Ma, Z. Liu, and C. Li (2025) LLaVA-video: video instruction tuning with synthetic data. External Links: 2410.02713, Link Cited by: Introduction, Related Work, Experiments. [54] J. Zhao, Y. Wu, R. Deng, S. Xu, J. Gao, and A. Burke (2025) A survey of autonomous driving from a deep learning perspective. ACM Computing Surveys. Cited by: Introduction. [55] B. Zhu, Y. Fu, Q. Dong, G. Sun, T. Qian, Y. Wu, D. P. Paudel, X. Xue, and Y. Fu (2026) Egosound: benchmarking sound understanding in egocentric videos. arXiv preprint arXiv:2602.14122. Cited by: Introduction. [56] Z. Zou, G. Sun, Z. Wei, J. Pan, Y. Li, M. Peng, and W. Xu (2026) Self in space: benchmarking self-awareness and spatial cognition in uav embodied intelligence. External Links: 2607.12477, Link Cited by: Introduction.