arXiv subjects

Xiang Li

Publications and source records attributed to Xiang Li.

At least 163 records · Page 9Linked to original sources

arXiv2026

Exo2EgoPose: Leveraging Exocentric Demonstrations for Vision-Language guided Egocentric 3D Hand Pose Forecasting

Perceiving multimodal cues and forecasting fine-grained actions from an egocentric (Ego) perspective is vital for applications like robot manipulation. However, previous studies either rely mainly on under-informed visual inputs to predict coarse human motions or follow the VRM/VLA paradigm, which suffers from insufficient robot data and the gap between human and robot embodiments. We observe that 3D hand pose naturally serves as a unified representation to bridge human-robot actions. Hence, we investigate an under-explored Vision-Language guided Egocentric 3D Hand Pose Forecasting (VL-EHPF) task, which aims to predict future Ego 3D hand poses from visual observations, a language instruction, and pose states. To overcome the limited field-of-view and highly dynamic motions in the Ego view, we propose a framework dubbed Exo2EgoPose, which innovatively leverages holistic and stable exocentric (Exo) demonstrations as guidance to compensate for partial and dynamic Ego-view cues. Specifically, we introduce a Dual-level Exocentric Reconstruction Module (DERM), which incorporates the paired Exo videos as supervision to reconstruct their video-level and chunked frame-level representations, thereby modeling spatial contexts and temporal dynamics. Then, the Global-to-Local Modulation Module (GLMM) utilizes the reconstructed hierarchical Exo representations for progressive feature refinement via attention mechanisms and adaptive modulation, enabling comprehensive Exo guidance for accurate Ego hand pose forecasting. Extensive experiments on \textit{AssemblyHands}, \textit{Ego-Exo4D}, and our newly constructed \textit{EgoMe-pose} benchmarks show the superiority of our method, which outperforms state-of-the-art methods by a large margin. Moreover, it demonstrates an effective human-to-robot transfer capability and yields improvements on the \textit{CALVIN} dataset.

cs.CV↗

arXiv2026

ASTELD: A Six-Axis Classification Framework for Autonomous AI Agents - Design, Evaluation, and an OpenClaw Case Study

Autonomous AI agent platforms differ substantially in architecture, security, tool integration, execution, autonomy, and deployment, yet the field lacks a common classification scheme for comparing these design choices. We propose ASTELD, an operational six-axis classification framework for autonomous AI agents: Architecture pattern, Security posture, Tool integration model, Execution paradigm, Level of autonomy and human control, and Deployment topology. ASTELD is constructed by synthesizing prior agent taxonomies with observable platform properties and explicit category-assignment rules. We evaluate its discriminative and explanatory utility by mapping eight representative frameworks and by using OpenClaw as an in-depth case study. The resulting profiles separate all eight platforms under their dominant configurations and reveal three cross-platform patterns: a security-accessibility diagonal, strong execution-architecture coupling, and capability convergence with persistent architectural differentiation. We further classify 50+ OpenClaw derivatives and find that innovation concentrates on the Security, Execution, and Deployment axes, indicating that ASTELD can explain where ecosystem fragmentation occurs. The OpenClaw case study also supplies a six-category vulnerability taxonomy, evidence from five institutional assessments, and adoption and governance analyses that connect platform coordinates to observed risks. These results position ASTELD as a reproducible method for comparing agent platforms, identifying unoccupied design regions, guiding framework selection, and organizing future empirical research. The analysis also exposes a consequential empty region: none of the evaluated systems combines local-first deployment with enterprise-grade security.

cs.CR↗

arXiv2026

LongCat Sparse Attention: Taming the Lightning via Streaming-aware Hierarchical Cross-Layer Indexing

DeepSeek Sparse Attention (DSA) enables efficient long-context modeling through its Lightning Indexer. However, practical deployment remains constrained by the indexer's expensive $O(L^2)$ scoring overhead and the hardware-inefficient, discontinuous memory-access patterns induced by its outputs. To address these system-level bottlenecks, we introduce LongCat Sparse Attention (LSA), a hardware-algorithm co-designed framework comprising three complementary and orthogonal strategies: (1) Streaming-Aware Indexing, which selectively converts scattered KV entries into hardware-aligned contiguous layouts to enable coalesced HBM access; (2) Cross-Layer Indexing, which amortizes indexing overhead by reusing the results produced by a single layer across consecutive layers, supported by cross-layer distillation; and (3) Hierarchical Indexing, which adopts a coarse-to-fine scoring scheme to progressively narrow the candidate set for each query, thereby substantially reducing indexing computation. Extensive scaling experiments, ranging from 69B-A3B to 560B-A27B models, demonstrate that LSA consistently achieves performance on par with full attention across both general-purpose and long-context benchmarks. Moreover, LSA supports native training with context lengths of up to one million tokens and underpins the development of LongCat-2.0 (1.6T-A48B). To facilitate further research, we also introduce and open-source LongCat-Flash-Lite-Sparse (69B-A3B), which integrates LSA into LongCat-Flash-Lite and incorporates an updated long-context training corpus.

cs.AI↗

arXiv2026

Pin Once, Swap Light: Subspace-Aligned Centroid-Residual Training for Efficient Ultra-LoRA Serving

Modern multi-tenant Low-Rank Adapters (LoRAs) serving systems concurrently host tens to hundreds of LoRA adapters. Though powerful, this introduces a critical system dilemma between serving efficiency and task performance: higher-rank adapters generally achieve better downstream task performance, but their GPU VRAM footprint and Host-to-Device PCIe swapping overhead severely constrain scalability. Conversely, ultra-low-rank adapters ($r \le 2$) minimize both VRAM footprint and PCIe transfer overhead, but suffer from downstream task performance degradation. To solve this problem, we propose Subspace-Aligned LoRA Training (SALT), a serving efficiency-aware hierarchical fine-tuning framework. Our solution operates in three phases. First, a provider jointly trains high-capacity domain centroids on public data within the domain using a novel alignment regularizer that coheres in-domain task subspaces into a unified basis. Next, users fine-tune ultra-low-rank task residual adapters on private data atop those frozen centroids. Finally, during inference, the provider pins the centroid in GPU VRAM and dynamically swaps in each user's task residual on demand. Across LLMs of varying scales, SALT recovers high-rank accuracy using $r \le 2$ residuals, achieving up to 18.5% absolute accuracy gains over state-of-the-art compression baselines and reducing per-adapter memory by up to 16x. When integrated into vLLM, SALT improves serving throughput by up to 51% under PCIe bandwidth pressure and 28% under GPU VRAM constraints for Llama-3.2-3B.

cs.LG↗

arXiv2026

Accelerated quantum Monte Carlo simulations of the attractive Hubbard model on the kagome lattice

The recent discovery of several families of kagome materials and experimental realization of optical kagome lattices have stimulated growing numerical studies of interaction-driven correlated states on the kagome lattice. Among the available numerical approaches, determinant quantum Monte Carlo (DQMC) is a powerful method for investigating such strongly correlated states. However, the accessible system sizes of existing DQMC simulations remain limited, preventing reliable finite-size scaling analyses. Here we develop a general acceleration scheme based on fast Fourier transform (FFT) for propagator multiplications on composite lattices and combine it with the delay-update algorithm, enabling simulations on system sizes twice as large as those of previous DQMC studies, allowing reliable finite-size scaling analyses of the attractive kagome-lattice Hubbard model. Our large-scale simulations reveal the interaction-driven zero-temperature superfluid quantum criticality at the Dirac filling and provide reliable estimates of the associated critical exponents. Besides, we find no evidence that the previously proposed triangle-rule charge-density-wave order survives in the thermodynamic limit, suggesting that it is likely a finite-size effect. Moreover, for system sizes accessible in current two-dimensional optical lattice experiments, the combined FFT and delay-update scheme exhibits an effective computational cost scaling as $N^{2.49}$, substantially below the $\mathcal{O}(N^3)$ computational cost of conventional DQMC simulations.

cond-mat.str-el↗

arXiv2026

GradientStabilizer:Fix the Norm, Not the Gradient

Training instability in modern deep learning systems is frequently triggered by rare but extreme gradient-norm spikes, which can induce oversized parameter updates, corrupt optimizer state, and lead to slow recovery or divergence. Widely used safeguards such as gradient clipping mitigate these failures but require threshold tuning and indiscriminately truncate large updates. We propose GradientStabilizer, a lightweight, drop-in gradient transform that preserves the instantaneous gradient direction while replacing the update magnitude with a statistically stabilized estimate derived from running gradient-norm statistics. We prove that the resulting stabilized magnitude is uniformly bounded on spike steps, independent of the spike size, and show how this boundedness controls optimizer state evolution in adaptive methods. Across LLM pre-training (FP16), quantization-aware pre-training (FP4), ImageNet classification, reinforcement learning, and time-series forecasting, GradientStabilizer consistently improves training stability, widens stable learning-rate regions, and reduces divergence relative to clipping-based baselines, even substantially reducing Adam's sensitivity to weight-decay strength. Code will be released soon.

cs.LG↗

arXiv2026

CoorGrasp: Coordinated Contact Control for Adaptive Dexterous Grasping Under Uncertainty

While recent research has focused heavily on dexterous grasp pose generation, less attention has been devoted to the execution of planned grasps. Under shape and position uncertainty, open-loop execution often yields uncoordinated contacts, causing undesired in-hand object motion and even grasp failures. To address this, this paper proposes a tactile-driven model predictive controller for adaptive and delicate execution of diverse dexterous grasps. Our approach emphasizes multi-contact coordination across both approaching and grasping phases, with three key novelties: (i) coordination-aware phase separation, (ii) arm-hand coordination to compensate for position errors, and (iii) adaptive force coordination to increase contact forces in a balanced manner. An analytical model is employed to relate contact forces to robot joint motions for predictive control. Our formulation imposes no restrictions on grasp types or contact configurations and integrates seamlessly with state-of-the-art grasp pose generation methods. We validate the approach through large-scale simulations involving 15k grasps across 478 objects on three robotic hands, and real-world experiments on 8 objects. Results demonstrate that our method achieves higher grasp success rates and reduced undesired object movements.

cs.RO↗

arXiv2026

Evaluating the Usability of Microgestures for Text Editing Tasks in Virtual Reality

As virtual reality (VR) continues to evolve, traditional input methods such as handheld controllers and gesture systems often face challenges with precision, social accessibility, and user fatigue. These limitations motivate the exploration of microgestures, which promise more subtle, ergonomic, and device-free interactions. We introduce microGEXT, a lightweight microgesture-based system designed for text editing in VR without external sensors, which utilizes small, subtle hand movements to reduce physical strain compared to standard gestures. We evaluated microGEXT in three user studies. In Study 1 ($N=20$), microGEXT reduced overall edit time and fatigue compared to a ray-casting + pinch menu baseline, the default text editing approach in commercial VR systems. Study 2 ($N=20$) found that microGEXT performed well in short text selection tasks but was slower for longer text ranges. In Study 3 ($N=10$), participants found microGEXT intuitive for open-ended information-gathering tasks. Across all studies, microGEXT demonstrated enhanced user experience and reduced physical effort, offering a promising alternative to traditional VR text editing techniques.

cs.HC↗

arXiv2026

CurveShift: Is Agent Progress Scalar? Separating Level from Shape

Progress in large language models is often summarized using a single scalar measure, such as a time horizon, a latent ability estimate, or an aggregate benchmark score. These summaries capture the overall performance, but they do not test whether progress is distributed differently across task difficulty. We find that most of the apparent shift in gains toward harder tasks does not reflect a change in the shape of the difficulty-response curve. On METR time-horizon data, a single Rasch model with rising ability reproduces this pattern, so it is largely explained by ceiling effects rather than a qualitative change in capability. This echoes how the choice of metric can make claimed emergent abilities look like a property of the models themselves. We then identify a smaller hard-task effect that survives this control. Isolating it is difficult on agentic benchmarks, because newer models are usually run with newer agentic harnesses, so a gain on hard tasks cannot be assigned to the model or its scaffold. We break the confound with LiveCodeBench, a public competitive programming benchmark that runs no agentic scaffold while pairing dated models with an exogenous difficulty ordering. After accounting for the rise in overall ability, models released after September 2024 still gain on the hardest problems beyond what their easy and medium performance predicts, by about +0.40 logits under our most conservative assumption, raising the hard-problem solve rate from roughly 18% to 25%. The effect is led by the strongest reasoning models and holds for hard tasks that need only short reasoning, not autonomy over long horizons. We present this as a result specific to competitive programming, since our clean identification rests on a single coding benchmark. We release the LiveCodeBench Difficulty Panel (66 dated models x 1,055 problems) and our analysis code.

cs.CL↗

arXiv2026

Strategy, Not Payoffs: A Behavioural Embedding of Normal-Form Games

Learning a strategic task changes more than what is directly taught: fine-tuning on one game can either enhance or degrade an agent's ability to reason in another. Understanding and predicting this transfer of strategic capabilities, however, remains a key challenge for large language models (LLMs). Normal-form games provide an ideal testbed for analyzing this phenomenon, as they feature explicitly defined payoffs and well-characterized equilibrium behaviours. In this work, we investigate whether game embeddings can explain and predict changes in LLM strategic capabilities following fine-tuning across different games. We propose a lightweight two-feature embedding that captures fundamental behavioural demands: the entropy of the Nash equilibrium and the sensitivity of optimal responses to an opponent's action. We show that while existing published structural embeddings primarily memorize game identities and fail to generalize, our behavioural embedding reliably predicts performance changes on held-out games. These results demonstrate that the transfer of strategic capabilities in LLMs is not dictated by the payoff geometry of a game, but by the underlying structure of the decision-making behaviour it requires.

cs.GT↗

arXiv2026

TEA-AgriVLN: Traversability Estimation Alarm for Agricultural Vision-and-Language Navigation

Vision-and-Language Navigation in Continuous Environments (VLN-CE) requires an agent to follow a natural language instruction, predicting a sequence of low-level actions to navigate a robot from a starting point to a target location. The A2A benchmark and the AgriVLN method pioneeringly extended VLN-CE from indoor scenes to agricultural scenes, while we observed a challenging distinction: In indoor scenes, whether a zone is traversable tends to be clear to classify, such as wood floors are traversable but concrete walls are not. In agricultural scenes, however, this issue tends to be ambiguous, such as an unripe cornfield might be traversable for a robotic dog but might be non-traversable for a human. To address this issue, we propose the TEA module, which estimates the traversability of the camera image, then alarm the decision-maker for rethinking when the predicted action does not align with the traversability map. We integrate it into the AgriVLN backbone to build our TEA-AgriVLN method. When evaluated on A2A, it improves Success Rate (SR) from 0.47 to 0.54 and Navigation Error (NE) from 2.91 m to 2.70 m, showing the state-of-the-art performance in the agricultural VLN-CE domain. We further implement the ablation studies and the case study, discussing the effectiveness and limitations of TEA on different ground categories and scene classes. Code: https://github.com/AlexTraveling/TEA-AgriVLN.

cs.RO↗

arXiv2026

DIRECTOR: Dynamic Index-based Recommendation with Transport-Optimized Retrieval

Reranking is a combinatorial decision problem that aims to select and order a high-utility slate from a request-specific candidate set. A major line of generative rerankers adopts autoregressive (AR) models, which construct the slate one position at a time to capture inter-position dependencies. However, under practical greedy or bounded-width decoding, prefix-based search may prematurely prune globally promising permutations and incurs inherently sequential latency, restricting the effective search space under a fixed serving budget. Non-autoregressive (NAR) alternatives alleviate this efficiency bottleneck through position-parallel prediction, but naive position-wise factorization treats different positions too independently, leading to insufficient cross-position coordination and potentially duplicate or conflicting item selections. To retain parallel efficiency while introducing global structural coordination, we propose Dynamic Index-based RECommendation with Transport-Optimized Retrieval (DIRECTOR), a transport-guided parallel reranking framework. DIRECTOR maps candidate items into a continuous latent space and generates request-conditioned dynamic retrieval indices for all target positions in parallel. During training, it uses entropy-regularized OT to provide conflict-aware supervision; at inference, it directly performs global hard matching on similarity matrix, producing duplicate-free slates without iterative transport. To further align the generator with an opaque list-wise evaluator that returns only a scalar utility, we introduce a prefix-anchored credit assignment mechanism that converts the global reward into position-specific training signals. Extensive offline and online experiments demonstrate that DIRECTOR consistently outperforms strong reranking baselines, achieving significant improvement in large-scale industrial recommendation scenarios.

cs.IR↗

arXiv2026

PSG: Pair-Space Generation for Efficient Generative Reranking

Modern recommender systems adopt Generator-Evaluator (G-E) for list-wise reranking: a generator produces sequences from candidates and an evaluator scores them at sequence-level to filter out the optimal one for exposure. Auto-Regressive(AR), working as the backbone for generative recommendation, suffers two limitations. First, its complexity grows linearly with list length, forcing the system to generate fewer lists under rigorous latency constraints and thus limiting exploration. Second, teacher-forcing creates a train-test mismatch; cumulative errors worsen with length and degrade quality. To address these problems, we propose Pair-Space Generation (PSG), a reformulation that elevates the generation atom from individual items to ordered item pairs. Given $n$ candidate items, PSG operates over pair vocabulary of size $n(n-1)$ per request, generates only $L/2$ tokens. Pair token representations are produced on-the-fly by a pretrained pair-token representation module optimized over large scale exposure logs, eliminating the data sparsity that would otherwise plague a quadratic sized vocabulary. We establish three theoretical guarantees: (i) PSG is bijective with item-space generation and induces an equivalent family of sequence distributions, thus incurring no loss of expressiveness; (ii) generation in pair-token space achieves approximately a $2\times$ to $4\times$ speedup theoretically under moderate settings and $1.83\times$ in the real industrial environmental settings; and (iii) under outcome-only rewards, the worst-case suboptimality of PSG is bounded by $O((L/2)^2 \barε)$, representing a nearly $4\times$ improvement over item-space generation. Beyond benchmark-based validation, PSG has also been deployed on Kuaishou, delivering a 0.178\% lift in per-user stay time on the platform, which serves over 400 million daily active users.

cs.IR↗

arXiv2026

FlashEvaluator: Expanding Search Space with Parallel Sequence-Level Evaluation

The Generator-Evaluator (G-E) framework generates K candidate sequences and uses an evaluator to select the highest-scoring one, which is widely used in recommender systems (RecSys) and natural language processing (NLP). Existing evaluators commonly score candidates independently. Although such evaluations can be batched, independent scoring neither models interactions among candidates nor eliminates repeated computation of request-level context and recurring candidate elements, causing the total evaluation work to grow approximately linearly with K. To handle with, we propose FlashEvaluator, a joint evaluator that scores all candidate sequences in a single forward pass. FlashEvaluator factorizes evaluation into shared request-level encoding, reusable candidate-side computation, sequence assembly by indexing, and cross-sequence interaction for setwise comparison. We call this request-local reuse scheme QKV-Cache: inspired by autoregressive KV caching, it reuses context-side key/value representations across candidate sequences and, when candidate elements recur, reuses their request-conditioned representations on the query side. In repeated-item settings, the dominant item-encoding cost therefore depends on the number of distinct items rather than their total occurrences across sequences, reducing the marginal cost of evaluating additional candidates. We provide a computational analysis and evaluate FlashEvaluator on recommendation and text summarization. The results show lower latency and higher throughput with competitive recommendation and summarization quality. In an online deployment at Kuaishou with K=50, FlashEvaluator reduces inference latency by 44% and increases QPS by 114% relative to the production baseline, while yielding statistically significant gains in retention, engagement, and ecosystem metrics.

cs.IR↗

arXiv2026

LazyMem: Retrieve Broadly, Construct Selectively for Efficient Long-Term Agent Memory

Long-term memory enables LLM agents to leverage past interactions, but dialogue histories quickly exceed the context window, forcing agents to retrieve relevant subsets at query time. Because useful evidence is sparse and scattered across verbose conversations, retrieval faces a fundamental tension: broadening recall improves coverage but floods downstream reasoning with noise, while compressing memories at write time eases retrieval but irreversibly discards details that future queries may need. We introduce LazyMem, which resolves this tension by deferring all memory construction to query time. Given a retrieved candidate pool, a lightweight model processes it in overlapping parallel windows, selectively retaining and compressing only query-relevant content. The model is trained with supervised fine-tuning followed by reinforcement learning, using a reward that jointly encourages the identification of relevant messages and the generation of compressions that are faithful to the source and useful for answering the query. On LongMemEval, LazyMem-4B achieves an LLM-judge accuracy of 0.85, outperforming the strongest non-oracle baseline while using only 213 answer-context memory tokens, 21.0 times fewer than the baseline. It further generalizes to LoCoMo without target-domain training and reduces mean latency relative to the prior query-time baseline. Code is available at https://github.com/allacnobug/LazyMem.

cs.AI↗

arXiv2026

HAIGEN: Towards Human-AI Collaboration for Facilitating Creativity and Style Generation in Fashion Design

The process of fashion design usually involves sketching, refining, and coloring, with designers drawing inspiration from various images to fuel their creative endeavors. However, conventional image search methods often yield irrelevant results, impeding the design process. Moreover, creating and coloring sketches can be time-consuming and demanding, acting as a bottleneck in the design workflow. In this work, we introduce HAIGEN (Human-AI Collaboration for GENeration), an efficient fashion design system for Human-AI collaboration developed to aid designers. Specifically, HAIGEN consists of four modules. T2IM, located in the cloud, generates reference inspiration images directly from text prompts. With three other modules situated locally, the I2SM batch generates the image material library into a certain designer-style sketch material library. The SRM recommends similar sketches in the generated library to designers for further refinement, and the STM colors the refined sketch according to the styles of inspiration images. Through our system, any designer can perform local personalized fine-tuning and leverage the powerful generation capabilities of large models in the cloud, streamlining the entire design development process. Given that our approach integrates both cloud and local model deployment schemes, it effectively safeguards design privacy by avoiding the need to upload personalized data from local designers. We validated the effectiveness of each module through extensive qualitative and quantitative experiments. User surveys also confirmed that HAIGEN offers significant advantages in design efficiency, positioning it as a new generation of aid-tool for designers.

cs.HC↗

arXiv2026

High-Order Pole-Skipping in Near-Extremal Holography

We develop a systematic analytic method for studying high-order pole-skipping in near-extremal holographic black holes. In the near-extremal regime, approaching the limit $T\to0$, the near-horizon geometry develops an approximately $\mathrm{AdS}_2 \times \mathbb{R}^{d-1}$ structure; we show that the mode index $q$ labeling pole-skipping points is identified with the IR conformal dimension $Δ_{\mathrm{IR}} = q$ in the emergent $\mathrm{AdS}_2/\mathrm{CFT}_1$ correspondence, providing a concrete physical interpretation of the subleading pole-skipping tower. The method reorganizes the near-horizon Frobenius expansion according to powers of temperature. This reveals a temperature-graded hierarchical structure that reduces the $n$-th-order pole-skipping condition to a factorized algebraic equation:each pole-skipping momentum depends only on the mode index $q$, not on the order $n$. This $n$-independence produces a high degeneracy as $T\to 0$, where pole-skipping momenta at all orders collapse onto a discrete set of values determined by near-horizon geometry and the scalar field mass; these values can be expressed in terms of thermodynamic quantities such as entropy density and specific heat. In the limit $n \gg 1$ (with $nT$ remaining small), the leading pole-skipping momenta grow asymptotically as $k_{n,n} \propto n$. We compute leading temperature corrections and verify our predictions through numerical analysis of the Dyonic Gubser--Rocha model. The results confirm that high-order pole-skipping at low temperature is governed by near-horizon physics. This provides analytic access to pole-skipping points well beyond those accessible by standard determinant methods and clarifies the structure of holographic Green's functions in the low-temperature regime.

hep-th↗

arXiv2026

Memory-Augmented Multimodal Large Language Models for Small Object Understanding in Streaming Aerial Videos

Language-guided aerial perception aims to understand user-specified tiny targets in complex unmanned aerial vehicle (UAV) scenes. In real UAV deployment, the UAV must respond while it flies, so such perception runs in an online streaming manner, where frames arrive sequentially and the model responds to each one without access to future frames. However, applying current Multimodal Large Language Models (MLLMs) to this setting raises two challenges. First, targets viewed from the air are often tiny, yet the visual compression in existing MLLMs treats all regions equally and discards their fine-grained details. Second, understanding a continuous stream requires past-frame context, yet retaining the entire history is infeasible on resource-constrained onboard hardware, whereas discarding it causes the target to drift or disappear. We address the tiny object and streaming challenges from both data and method perspectives. From the data perspective, we present \textbf{DroneEyes}, the \textbf{first} pixel-level and open-vocabulary referring-segmentation dataset for tiny aerial targets, comprising $2,140$ high-definition videos and $176,623$ pairs across Object Description and Referring Expression tasks, with dense per-frame masks. From the method perspective, we propose \textbf{SkyAnchor}, an MLLM with two designs to the above challenges: a Semantics-Aware Token Router that preserves small-target under a reduced visual-token budget, and a Hierarchical Memory Bank that keeps the target consistently understood on streams.

cs.CV↗