Searcharxiv⌕ Search

arXiv subjects

Wenbo Zhu

Publications and source records attributed to Wenbo Zhu.

At least 19 recordsLinked to original sources

Manufacturing Differential Sparsity: Planner-Hardware Co-Design for Efficient World Model Planning

JEPA-style latent world models let embodied agents plan toward new goals with model predictive control, but they are still too slow for real-life embodied control: sampling-based MPC repeats predictor inference tens of thousands of times per plan, and even on high-end GPUs one planning step can take tens of seconds. We observe substantial latent locality in this repeated-inference workload: the predictor maps close actions to close latents, and inside a quantized predictor this closeness takes a form that hardware can exploit, differential bit-level sparsity. Unlike an input stream, whose temporal similarity is given, a planner itself decides which candidates it evaluates together, so the planner can choose its own references and, with reference forest sampling, manufacture more sparsity: it couples the sampled action sequences of each group to a shared root, with competitive planning quality. We build the accelerator on the insights of the rich line of bit-level sparsity accelerators and design it and the planner tightly coupled as a single co-design: the planner increases sparsity and exposes structure for hardware scheduling, and the bit-exact hardware effectively uses the sparsity that the planner creates. On three JEPA world models, our co-design is 1.28-1.80x faster than a dense accelerator and uses 38-52% less energy, and it is 1.22-1.58x faster than prior bit-sparse design and uses 16-36% less energy, with the same area. To our knowledge, our planner is the first planner that manufactures differential bit-level sparsity across its candidates and exposes its structure for hardware scheduling, and our co-design is the first algorithm-hardware co-design for JEPA-style world-model planning.

cs.AR↗

BFLA: Block-Filtered Long-Context Attention Mechanism

We propose Block-Filtered Long-Context Attention (BFLA), a training-free sparse prefill method that combines deterministic Lightning-style block indexing, query-tile sharing, and length-adaptive Ratio-$K$ selection with block-sparse execution. BFLA evaluates attention using the original Q/K/V, requires no retraining or per-model parameter retuning, and integrates with vLLM while preserving native decoding. For fixed Q/K/V, we bound local attention-output error in terms of the omitted attention mass, indexer-score distortion, retention budget, and query-tile sharing. Evaluations on Gemma4-E4B-IT, Qwen3.6-27B, Qwen3.8-27B, and Qwen3.8-Flash-Next across LongBench-v2, RULER, InfiniteBench, and Terminal Bench 2.1 show BFLA matches or even outperforms dense baselines. BFLA shows up to approximately $4\times$ time-to-first-token and $3.72\times$ end-to-end request speedups over the corresponding dense-prefill baselines. Comparisons with state-of-the-art FlashPrefill V2 and Qwen Sparse Attention also demonstrate that BFLA has better performance and efficiency trade-offs. These findings highlight the value of BFLA for the acceleration of long-context serving. Project website: https://github.com/Alicewithrabbit/BFLA.

eess.SP↗

PACE: Plan, Attend, Check, and Emit for Training-Free Response Timing in Streaming Video Understanding

Streaming video understanding requires a model to decide not only what to answer but also when to answer. Evidence arrives progressively, and recent benchmarks identify premature responses, issued before the queried event has concluded, as a recurring failure mode. Existing approaches often learn response timing through additional training, while training-free alternatives typically rely on scalar gating signals that do not explicitly represent when an answer becomes final. We organize chunk-wise streaming errors into three patterns: sampling variation can make an answer unstable, a stable answer may lack support in the observed video, or an answer may be supported by the current evidence while the event remains unfinished. We propose PACE (Plan, Attend, Check, and Emit), a training-free framework designed around these three failure modes. Given a question, PACE drafts a plan specifying what to observe and what will signal completion. As the stream unfolds, it invokes an off-the-shelf video LLM at informative moments, evaluates each draft for stability, grounding in its cited frames, and completion, and records verified observations in an event ledger. Experiments on ProactiveVideoQA and OVO-Bench Forward show that PACE improves over the official protocols by up to 5.9 average PAUC points and 6.2 average accuracy points across proprietary and open-source backbones, yielding state-of-the-art performance.

cs.CV↗

Auto-Director: Leveraging Multi-Agent Collaboration for Autonomous Video Editing

Autonomous video editing from open-world resources remains a challenging problem in multimedia content creation. Retrieval-based methods are constrained by fixed source material and lack creative flexibility, while generative models struggle with long-horizon coherence and often diverge from user intent. Both paradigms fail to replicate the structured, multi-stage reasoning that underlies professional post-production. To address these limitations, we present Auto-Director, a tool-augmented multi-agent framework that mirrors professional editorial workflows: a Planning Agent decomposes abstract topics into fact-grounded scripts via retrieval-augmented planning; a Material Agent indexes open-world footage through shot segmentation and visual tagging; an Editor Agent assembles semantically aligned clips with MLLM-driven spatial adaptation; and a Critic Agent governs an iterative Reflect-and-Revise loop that checks scene-level alignment and pacing. To rigorously evaluate this new paradigm, we introduce VEdit-Bench, a benchmark of 100 expert-curated topics with human-annotated Golden Scripts and a five-dimensional metric suite covering Narrative Logic, Factual Accuracy, Semantic Alignment, Compositional Quality, and Stylistic Cohesion. Extensive experiments demonstrate that Auto-Director significantly outperforms baseline methods in both automated metrics and human evaluation, effectively bridging the gap between amateurs and professional-grade video production.

cs.CV↗

BiMoGen: Bidirectional Motion-Text Generation via Unified Masked Discrete Diffusion

Text-to-motion generation and motion-to-text captioning are two fundamental tasks in human motion modeling, both grounded in the same underlying motion-text correspondence. Existing unified approaches mostly rely on autoregressive modeling, which imposes a fixed generation order and is therefore poorly suited to the bidirectional dependencies between language and motion, allowing early prediction errors to persist as fixed context and degrade both temporal coherence and cross-modal consistency. Masked discrete diffusion, which models sequences through iterative bidirectional prediction, allows prediction from bidirectional context without a fixed generation order. We therefore propose BiMoGen (Bidirectional Motion-text Generation), a unified masked discrete diffusion framework for bidirectional motion-text modeling. To stabilize training, we design Decoupled Uni- and Cross-Modal Training, in which masked pretraining first establishes cross-modal correspondence on paired motion-text sequences, after which supervised fine-tuning specializes the model for bidirectional generation. Masked diffusion nonetheless faces a train-test mismatch, as the model is trained on clean ground-truth context yet encounters self-generated and potentially erroneous context at inference, with errors committed under heavily masked states propagating through subsequent steps. We further introduce Generation-Aware Self-Correction that exposes the model to its own predictions during training and uses correction passes at early sampling steps for further refinement of previously unmasked tokens. Extensive experiments on HumanML3D and KIT-ML demonstrate competitive performance on both tasks, validating the effectiveness of the proposed two-stage training and self-correction designs. The project page is available at https://wengwanjiang.github.io/BiMoGen-Page.

cs.CV↗

Predictive Rolling-Horizon Optimization for Commitment-Aware Model-Parallel Inference under Spatio-Temporal Edge Dynamics

Model-parallel inference over dynamic edge systems requires scheduling decisions that account for not only instantaneous resources but also future resource contention and reliable service commitments. Existing edge-inference designs, however, predominantly optimize performance metrics based on current or short-term system states, without explicitly coupling current assignments with future commitment fulfillment. To address this issue, we propose PROMISE, a predictive rolling-horizon framework for commitment-aware model-parallel inference under spatio-temporal edge dynamics. PROMISE jointly models stochastic task generation, mobility-induced communication variations, privacy-aware model partitioning, and load-dependent edge computing capability. We introduce committed completion time (CCT) as an endogenous service decision and formulate joint SD--ES mapping and CCT determination to balance commitment fulfillment and service reward. To address cross-timeslot coupling, PROMISE estimates future task arrivals and computational workloads over an adaptive horizon and embeds them into certainty-equivalent ES-state rollout. A progress-aware surrogate then evaluates feasible current-stage mappings, while the corresponding CCTs are analytically recovered from predicted completion times. Only the first-stage decisions are committed, and the optimization is repeated with newly observed states in a receding-horizon manner. Numerical experiments demonstrate robust scheduling performance under diverse system scales and workload dynamics, while Raspberry-Pi-based experiments validate the practical feasibility and key system characteristics of model-parallel edge inference.

cs.DC↗

Chipstitch: A Scalable and Cost-Efficient Platform for Accelerated Silicon Prototyping

Silicon evidence serves as a critical receipt of credibility across academia and industry, validating innovation and signaling feasibility in ways that simulations alone cannot. Its value is difficult to quantify but widely recognized for driving hardware-software innovation. However, access to this receipt has remained severely constrained by high dollar costs and engineering effort barriers, which lock out innovators unable to afford it. To bridge this gap, we consider a new design paradigm called silicon query architectures that spans EDA, architecture, and VLSI and aims to transform single-die area into a substrate for scalable, tessellated, intra-die silicon receipts, dropping the cost of entry by one to two orders of magnitude without sacrificing research-grade specifications. We concretely propose Chipstitch, a silicon query architecture realized with three contributions: an algorithmic three-layer approach to repeatable and scalable tessellation with a keystone algorithm, a scalable architecture-VLSI system interconnect that fits within interstitial regions of the tessellation with research-grade timing and clocking, and a grid-aligned perimeter-based power shutdown technique that mitigates overheads for research-grade power measurement. Using a vertically integrated methodology extending down to post-layout VLSI implementation, we demonstrate Chipstitch on a 25-site heterogeneous design scenario, achieving 9.9x-14.5x total cost advantage and up to two orders of magnitude reduction in single-receipt entry cost compared to a traditional multi-project wafer baseline in Skywater 130nm and Intel 16nm process nodes.

cs.AR↗

TRACER: Adaptive Multi-Robot Social Navigation via Joint Human-Response Prediction and Interaction-Aware Replanning

Multi-robot navigation in human-shared spaces is inherently interactive: coordinated robot motions influence how nearby entities respond, while those responses provide valuable information for subsequent robot decisions. However, existing methods typically address action-conditioned prediction, multi-robot planning, or online adaptation separately, and therefore lack a unified mechanism for modeling joint robot-entity interactions and adapting future decisions from executed interaction outcomes. To address this gap, we propose TRACER, a bi-directional receding-horizon framework that closes the loop between prediction and adaptation. TRACER evaluates candidate (i.e., alternative feasible future motion plans for the robot team) trajectories using a per-entity probabilistic response model that separates individual-robot effects from non-additive pairwise interactions; after executing the selected trajectory prefix, it updates persistent identity-bound beliefs over latent response modes using the synchronized observed responses. These updated beliefs then guide subsequent candidate evaluation under probabilistic safety and response-aware cost criteria. Experiments show that (i) TRACER more accurately captures non-additive multi-robot interaction effects than a capacity-matched additive predictor, (ii) persistent identity-consistent evidence improves response prediction and downstream replanning, and (iii) the complete TRACER framework improves collision-free completion over an independent-robot baseline on the SocialGym2 multi-robot social-navigation benchmark.

cs.RO↗

Flying over The Uncertain Nature (FORTUNE): Intelligent and Humanistic 3D Path Planning for Low-Altitude Collaboration

The proliferation of low-altitude intelligent agents is increasing the demand for timely and socially responsible collaborative sensing in dynamic urban environments. However, jointly addressing heterogeneous spatiotemporal demands, environmental uncertainty, and human-centered operational constraints remains challenging. This paper studies 3D multi-UAV path planning and task assignment under uncertain ground PoI demands. Unlike existing work assuming static and fully known PoIs, we model persistent, temporally predictable, and emergent demands within a unified framework. We further incorporate altitude-dependent societal and environmental costs, including noise exposure and public safety risks, to balance sensing performance with socially compliant operations. To solve the resulting large-scale mixed-integer nonlinear problem, we propose FORTUNE, a hierarchical offline-online framework. Offline, a Transformer predicts Type-II PoI activation windows, while an enhanced sparrow search algorithm generates coordinated flight plans through priority-aware decoding and danger-aware evolution. Online, a lightweight refinement module accommodates emerging Type-III PoIs while preserving global mission coherence. Experiments on real-world traffic data and synthetic scenarios show that FORTUNE consistently outperforms state-of-the-art methods in effectiveness, scalability, and practical applicability.

cs.RO↗

SAGE: A Socially-Aware Generative Engine for Heterogeneous Multi-Agent Navigation

Safe and socially compliant navigation in open human-robot environments requires robots to reason about heterogeneous participants with different dynamics, autonomy levels, and social roles. Existing trajectory prediction and planning methods often rely on homogeneous interaction assumptions or enforce only geometric collision constraints, making it difficult to jointly model asymmetric interactions, coupled prediction-planning, and soft social norms. This paper proposes SAGE, a socially-aware generative engine for heterogeneous multi-agent navigation. SAGE represents robots and surrounding entities as a directed heterogeneous graph and employs a Heterogeneous Graph Transformer (HGT) to encode type-specific asymmetric interactions. Conditioned on the resulting context, a diffusion-based generative module jointly models future entity trajectories and robot trajectory plans. During inference, a training-free safety-social energy guidance mechanism refines sampled robot trajectories using differentiable collision, kinematic, task-progress, and role-conditioned social-compliance terms. Extensive experiments on real-world (ETH/UCY and SDD) and synthetic datasets verify the effectiveness of SAGE in improving safety and social compliance while maintaining task performance. The proposed guidance mechanism consistently reduces collision and social-violation rates, scales to teams of up to 20 robots, and enables explicit control of the safety-accuracy-task trade-off without retraining. These findings demonstrate the potential of SAGE as a scalable framework for socially-aware multi-agent navigation in complex environments.

cs.RO↗

DRIFT: Risk-Constrained Diffusion with Imitation Priors for Mixed-Autonomy Traffic Generation

Future intelligent transportation systems are envisioned to evolve toward a long-term mixed-autonomy paradigm, where human-driven vehicles (HVs) and autonomous vehicles (AVs) coexist within highly coupled traffic ecosystems. Such coexistence introduces pronounced heterogeneity, amplified uncertainty, and increasingly intricate interaction dynamics. In this context, it remains fundamentally challenging to simultaneously capture the heterogeneous behavioral distribution shifts arising from dynamic AV penetration, generate diverse yet executable trajectories under strong inter-vehicle coupling, and conduct reliable closed-loop safety and stability diagnostics for rare but high-impact events. To this end, we present Diffusion with Risk constraints, Imitation priors, and long-tail Feedback for mixed-autonomy Traffic generation (DRIFT), a mixed-autonomy traffic generation framework that unifies heterogeneity-aware conditional encoding, conditional diffusion-based executable trajectory generation, and progressive adversarial alignment enhanced by risk-aware long-tail feedback, thereby enabling traffic behaviors to be iteratively generated, filtered, selected, and validated within a closed-loop execution pipeline. In addition, a unified evaluation protocol is developed to jointly characterize safety, efficiency, and closed-loop stability across representative traffic scenarios and AV penetration regimes. Experimental results demonstrate that DRIFT achieves a strong safety-efficiency trade-off in closed-loop mixed-autonomy benchmarks, while further revealing the critical influence of candidate executability, online selection, and long-tail feedback on executable traffic evolution.

cs.DC↗

PowerFlow-DNN: Compiler-Directed Fine-Grained Power Orchestration for End-to-End Edge AI Inference

Edge AI systems operate under stringent energy and volume constraints, demanding extreme efficiency on limited battery capacity, with requirements worsening as intelligent capabilities advance. Prior work suggests fine-grained power orchestration through DVFS and power gating significantly improves efficiency critical to meeting such constraints, but introduces new challenges. We observe that layer-level approaches incur unintended overheads due to inter-layer coupling of power-control decisions, and jointly managing these mechanisms under limited voltage rails and transition overheads leads to a rapidly growing combinatorial schedule space. We propose PowerFlow-DNN, a compiler-directed framework for end-to-end power-state orchestration in ultra-low-power accelerators. By constructing a rigorous problem formulation for deadline-constrained, real-time, periodic inference as a unified inter-layer power-scheduling problem, our framework discovers energy-minimal power-state schedules while accounting for inter-layer impacts. We evaluate the framework on a DNN accelerator VLSI implementation in TSMC 40nm technology. Across representative edge networks, our approach discovers near-optimal solutions and achieves energy within 0.04\% of the exact ILP oracle, reducing energy by up to 48\% compared to an aggressive baseline without power orchestration, while reasoning over a combinatorial schedule space of over $10^{160}$ possible power-state assignments, yet operating on a structured layered state graph that enables efficient optimization, achieving up to 2.14$\times$ solver speedup via lightweight pruning.

cs.AR↗

Dual-Route Top-K Retrieval with 1v1 VLM Reranking for the CoVR-R

We describe \emph{Dual-Route Top-K Retrieval with 1v1 VLM Reranking} for the CoVR-R challenge. The method treats composed video retrieval as two coupled problems: finding a sufficiently complete top-k candidate set, and then safely deciding whether any candidate should replace a strong current top-1. We first improve the reasoning/text seed with a VLM slot selector over existing candidates, without introducing DFN visual retrieval. We then add a visual route from contact-sheet embeddings using DFN-H/DFN-L. The routes are merged into a top-10 candidate set, after which a VLM final reranker performs conservative 1v1 comparisons between the current top-1 and each challenger. On the hidden test split, the final system reaches 95.28 R@1, 97.47 R@5, 98.48 R@10, and 99.66 R@50. The main lesson is that CoVR-R benefits more from recall-selection decoupling than from broad text reranking or direct multi-candidate VLM classification.

cs.CV↗

Adaptive Dense Evidence Refinement for Video Relational Reasoning for VRR-QA Challenge

VRR-QA evaluates whether video-language systems can infer spatial, temporal, viewpoint, depth, and visibility relations that are not always resolved by a single frame. We present an inference-only system built around adaptive test-time computation. The system first answers each question with a direct video-language model pass, then uses multiple lightweight views to find unstable questions. Only these difficult questions are routed to a high-budget dense evidence module that constructs timestamped frame observations, relation-specific probes, candidate verification, and conservative temporal aggregation. This design separates two problems that are often confused in video question answering: finding plausible alternative answers and deciding when a current answer should actually be changed. On the test split, the final system obtains 90.07 average accuracy and 87.81 macro average accuracy. The report focuses on the final test system and the implementation settings required to reproduce the adaptive dense verifier.

cs.CV↗

Temporal Evidence Routing with Structured Visual Evidence for TimeLogicQA

TimeLogicQA evaluates whether video question answering systems can reason over temporal relations such as event existence, ordering, persistence, boundary conditions, and overlap. We address this task with a visual evidence routing pipeline that separates perception from symbolic temporal reasoning. The system first parses each question into event targets, answer mode, candidate options, and temporal operators. It then routes videos according to duration and operator difficulty, using ordered full-frame evidence for short clips and event-focused candidate windows for long videos. A multimodal large language model produces structured visual evidence for the relevant events, while programmatic verifiers recover dense action intervals and a deterministic reducer applies operator-specific temporal rules to produce the final answer. Conservative fusion accepts an answer only when the visual evidence, temporal program, and confidence checks agree, reducing noisy answer flips. On the official test evaluation, our final system achieves an AvgAcc of 81.8.

cs.CV↗

Look One Step Ahead: Forward-Looking Incentive Design with Strategic Privacy for Proactive Service Provisioning over Air-Ground Integrated Edge Networks

In air-ground integrated networks (AGINs), unmanned aerial vehicles (UAVs) provide on-demand edge services to ground vehicles. Realizing this vision requires carefully designed incentives to coordinate interactions among self-interested participants. This is exacerbated by the dynamic nature of AGINs, where spatio-temporal variations introduce significant uncertainty in matching UAVs and vehicles. Existing real-time service provisioning typically relies on precise trajectory information, raising privacy concerns and incurring decision latency. To address these challenges, we propose look one-step ahead (LOSA), a novel framework for efficient and privacy-aware service provisioning. By exploiting predictable vehicle travel times between intersections, LOSA decomposes the process into two coupled phases: (i) a privacy-aware look-ahead phase and (ii) a lightweight real-time execution phase. The look-ahead phase allows vehicles to adaptively adjust privacy budgets based on historical utility, balancing trajectory exposure and matching accuracy. Leveraging this, a double auction mechanism establishes binding one-step-ahead agreements (OSAAs) through trajectory similarity clustering, while constructing preference lists to hedge against mobility uncertainty. The execution phase then enforces pre-established OSAAs and preference lists, resolving real-time resource conflicts without costly re-negotiations. This design reduces computational overhead and preserves robustness. We analytically corroborate that LOSA guarantees truthfulness, individual rationality, and budget balance. Experiments on real-world datasets (DAIR-V2X, HighD, and RCooper) demonstrate that LOSA achieves superior privacy protection while lowering transaction latency compared to baseline approaches.

cs.NI↗

Adapting Point Cloud Analysis via Multimodal Bayesian Distribution Learning

Multimodal 3D vision-language models show strong generalization across diverse 3D tasks, but their performance still degrades notably under domain shifts. This has motivated recent studies on test-time adaptation (TTA), which enables models to adapt online using test-time data. Among existing TTA methods, cache-based mechanisms are widely adopted for leveraging previously observed samples in online prediction refinement. However, they store only limited historical information, leading to progressive information loss as the test stream evolves. In addition, their prediction logits are fused heuristically, making adaptation unstable. To address these limitations, we propose BayesMM, a Multimodal Bayesian Distribution Learning framework for test-time point cloud analysis. BayesMM models textual priors and streaming visual features of each class as Gaussian distributions: textual parameters are derived from semantic prompts, while visual parameters are updated online with arriving samples. The two modalities are fused via Bayesian model averaging, which automatically adjusts their contributions based on posterior evidence, yielding a unified prediction that adapts continually to evolving test-time data without training. Extensive experiments on multiple point cloud benchmarks demonstrate that BayesMM maintains robustness under distributional shifts, yielding over 4% average improvement.

cs.CV↗

On the Generalization of SFT: A Reinforcement Learning Perspective with Reward Rectification

In this work, we present a simple yet theoretically motivated improvement to Supervised Fine-Tuning (SFT) for the Large Language Model (LLM), addressing its limited generalization compared to reinforcement learning (RL). Through mathematical analysis, we reveal that standard SFT gradients implicitly encode a problematic reward structure that may severely restrict the generalization capabilities of model compared to RL. To rectify this, we propose Dynamic Fine-Tuning (\model), stabilizing gradient updates for each token by dynamically rescaling the objective function with the probability of this token. With just a single-line change, the method outperforms standard SFT on multiple difficult benchmarks and base models, from math reasoning to code generation and multi-modal tasks, demonstrating improved generalization. Additionally, \model~achieves competitive results in offline RL settings, providing an effective yet streamlined alternative. By bridging theoretical insights with practical solutions, this work advances the state of SFT. The source code will be available at https://github.com/yongliang-wu/DFT.

cs.LG↗