arXiv:2608.18234v3 Announce Type: replace-cross Abstract: Whole-body motion tracking policies turn a humanoid into a robust control interface: the teleoperator---or an upstream model---only supplies a coarse movement intent, while the low-level policy keeps the robot balanced and physically feasible. Existing trackers deliver this interface only on flat ground: trained in empty scenes, they never learn how contact with terrain and objects reshapes their dynamics, and they attempt to teach the policy to balance under any command by continually enlarging the reference-motion corpus, which stops working once feasible behaviors become environment-dependent. We present GigaBrain-WBC-0.5, the first Behavior World Model (BWM) for humanoid whole-body control. Rather than a purely reactive tracker, we train a causal Transformer to jointly predict its next action, next state, and the distribution over its next latent behavior command, so the network that acts also models how the environment shapes what it can do next. An automatic terrain-annotation pipeline recovers full 3D contact geometry from retargeted motion, enabling terrain annotation at the scale of existing motion datasets. The predicted distribution is reused at deployment to detect implausible commands online and retract them onto learned behaviors, so the robot attempts tasks in a "best-effort" manner. The result is a unified policy that takes real-time command, interacts with environment, and stays robust to implausible commands, falls, and disturbances. GigaBrain-WBC-0.5 achieves the highest success rate across all four regimes among three large-scale tracker baselines: 81.3% on terrain interaction (4.3x the strongest baseline), 83.1% under implausible commands, and 99.3% fall recovery (16.8x the strongest baseline). Hardware trials show robust interaction under missing supports and disturbances; the Unitree G1 checkpoint transfers to the Maker L01 robot with simple fine-tuning.

arXiv:2609.09187v1 Announce Type: cross Abstract: Advertising video generation is not only a video synthesis task, but also a product-conditioned reasoning problem whose success is measured by online business metrics. Recent video foundation models can generate realistic clips from multimodal conditions, yet they do not optimize how a product should be transformed into an effective advertisement or how future generation should be improved from online business feedback. To close this loop, we propose AgenticGen, a reward-guided agentic framework that decomposes advertising video generation into two trainable reasoning stages, strategy selection and draft generation, thereby exposing optimization targets that online business feedback can supervise. AgenticGen learns a performance-based reward from accumulated online feedback and a complementary rubric-based reward aligned with human quality standards, then uses them to supervise policy optimization. DPO first moves the agentic policies toward online preferences, and GRPO further refines both stages with process and outcome rewards. Offline experiments validate the reward models and successive policy optimization. Online A/B experiments in the TikTok advertising system show that AgenticGen after DPO and GRPO improves CTR by 2.72%, CVR by 2.63%, and Advv by 9.61% over the SFT baseline.

arXiv:2503.07482v2 Announce Type: replace-cross Abstract: Membership Inference Attacks (MIAs) aim to estimate whether a specific data point was used in the training of a given model. Existing state-of-the-art attacks typically rely on training multiple reference models to approximate the conditional score distribution for individual data points, which leads to significant computational overhead and limits their practical applicability. In this work, we propose a novel approach -- Bayesian Membership Inference Attack (BMIA), which performs conditional attack through Bayesian sampling. Specifically, we apply Laplace approximation to a single reference model to obtain a posterior over model parameters, enabling direct estimation of the conditional score distribution. Theoretically, we demonstrate that Bayesian sampling reduces intra-model variance, thereby improving attack power. This insight naturally motivates the multi-reference variant that further enhances performance when additional reference models are available. Extensive experiments across image, text, and tabular datasets indicate that our method achieves state-of-the-art performance in both effectiveness and efficiency.

arXiv:2605.08063v5 Announce Type: replace-cross Abstract: Existing Flow Matching (FM) text-to-image models suffer from two critical bottlenecks under multi-task alignment: the reward sparsity induced by scalar-valued rewards, and the gradient interference arising from jointly optimizing heterogeneous objectives, which together give rise to a 'seesaw effect' of competing metrics and pervasive reward hacking. Inspired by the success of On-Policy Distillation (OPD) in the large language model community, we propose Flow-OPD, the first unified post-training framework that integrates on-policy distillation into Flow Matching models. Flow-OPD adopts a two-stage alignment strategy: it first cultivates domain-specialized teacher models via single-reward GRPO fine-tuning, allowing each expert to reach its performance ceiling in isolation; it then establishes a robust initial policy through a Flow-based Cold-Start scheme and seamlessly consolidates heterogeneous expertise into a single student via a three-step orchestration of on-policy sampling, task-routing labeling, and dense trajectory-level supervision. We further introduce Manifold Anchor Regularization (MAR), which leverages a task-agnostic teacher to provide full-data supervision that anchors generation to a high-quality manifold, effectively mitigating the aesthetic degradation commonly observed in purely RL-driven alignment. Built upon Stable Diffusion 3.5 Medium, Flow-OPD raises the GenEval score from 63 to 92 and the OCR accuracy from 59 to 94, yielding an overall improvement of roughly 10 points over vanilla GRPO, while preserving image fidelity and human-preference alignment and exhibiting an emergent 'teacher-surpassing' effect. These results establish Flow-OPD as a scalable alignment paradigm for building generalist text-to-image models. The codes and weights will be released in: https://github.com/CostaliyA/Flow-OPD .

arXiv:2604.04339v1 Announce Type: new Abstract: Modeling spatial heterogeneity and associated critical transitions remains a fundamental challenge in geography and environmental science. While conventional Geographically Weighted Regression (GWR) and deep learning models have improved predictive skill, they often fail to elucidate state-dependent nonlinearities where the functional roles of drivers represent opposing effects across heterogeneous domains. We introduce a thermodynamics-inspired explainable geospatial AI framework that integrates statistical mechanics with graph neural networks. By conceptualizing spatial variability as a thermodynamic competition between system Burden (E) and Capacity (S), our model disentangles the latent mechanisms driving spatial processes. Using three simulation datasets and three real-word datasets across distinct domains (housing markets, mental health prevalence, and wildfire-induced PM2.5 anomalies), we show that the new framework successfully identifies regime-dependent role reversals of predictors that standard baselines miss. Notably, the framework explicitly diagnoses the phase transition into a Burden-dominated regime during the 2023 Canadian wildfire event, distinguishing physical mechanism shifts from statistical outliers. These findings demonstrate that thermodynamic constraints can improve the interpretability of GeoAI while preserving strong predictive performance in complex spatial systems.

arXiv:2604.04202v1 Announce Type: cross Abstract: AI agents deployed as persistent assistants must maintain correct beliefs as their information environment evolves. In practice, evidence is scattered across heterogeneous sources that often contradict one another, new information can invalidate earlier conclusions, and user preferences surface through corrections rather than explicit instructions. Existing benchmarks largely assume static, single-authority settings and do not evaluate whether agents can keep up with this complexity. We introduce ClawArena, a benchmark for evaluating AI agents in evolving information environments. Each scenario maintains a complete hidden ground truth while exposing the agent only to noisy, partial, and sometimes contradictory traces across multi-channel sessions, workspace files, and staged updates. Evaluation is organized around three coupled challenges: multi-source conflict reasoning, dynamic belief revision, and implicit personalization, whose interactions yield a 14-category question taxonomy. Two question formats, multi-choice (set-selection) and shell-based executable checks, test both reasoning and workspace grounding. The current release contains 64 scenarios across 8 professional domains, totaling 1{,}879 evaluation rounds and 365 dynamic updates. Experiments on five agent frameworks and five language models show that both model capability (15.4% range) and framework design (9.2%) substantially affect performance, that self-evolving skill frameworks can partially close model-capability gaps, and that belief revision difficulty is determined by update design strategy rather than the mere presence of updates. Code is available at https://github.com/aiming-lab/ClawArena.

arXiv:2604.01878v1 Announce Type: cross Abstract: Spectral graph contrastive learning has emerged as a unified paradigm for handling both homophilic and heterophilic graphs by leveraging high-frequency components. However, we identify a fundamental spectral dilemma: while high-frequency signals are indispensable for encoding heterophily, our theoretical analysis proves they exhibit significantly higher variance under spectrally concentrated perturbations. We derive a regret lower bound showing that existing global (node-agnostic) spectral fusion is provably sub-optimal: on mixed graphs with separated node-wise frequency preferences, any global fusion strategy incurs non-vanishing regret relative to a node-wise oracle. To escape this bound, we propose ASPECT, a framework that resolves this dilemma through a reliability-aware spectral gating mechanism. Formulated as a minimax game, ASPECT employs a node-wise gate that dynamically re-weights frequency channels based on their stability against a purpose-built adversary, which explicitly targets spectral energy distributions via a Rayleigh quotient penalty. This design forces the encoder to learn representations that are both structurally discriminative and spectrally robust. Empirical results show that ASPECT achieves new state-of-the-art performance on 8 out of 9 benchmarks, effectively decoupling meaningful structural heterophily from incidental noise.

arXiv:2509.23067v2 Announce Type: replace-cross Abstract: The rising cost of acquiring supervised data has driven significant interest in self-improvement for large language models (LLMs). Straightforward unsupervised signals like majority voting have proven effective in generating pseudo-labels for verifiable tasks, while their applicability to unverifiable tasks (e.g., translation) is limited by the open-ended character of responses. As a result, self-evaluation mechanisms (e.g., self-judging and entropy minimization) are predominantly used to derive pseudo-labels. However, self-evaluation relying on LLMs typically incurs high computational overhead and introduces overconfidence issues due to intrinsic biases. To address these challenges, we propose a novel self-evaluation-free approach for unverifiable tasks, designed for lightweight yet effective self-improvement. Inspired by majority voting commonly employed in verifiable tasks, we propose semantic voting as a novel mechanism that relaxes the principle of hard matching (i.e., exact matching) toward soft matching (i.e., semantic similarity). Soft matching is achieved by leveraging a lightweight sentence embedding model to quantify semantic similarity, thereby mitigating excessive computational burden and intrinsic bias-associated limitations of self-evaluation. Comprehensive experiments demonstrate that our method achieves substantial gains in computational efficiency and overall better performance than self-evaluation methods across diverse model architectures and tasks.

arXiv:2603.16790v3 Announce Type: replace-cross Abstract: Recent code large language models have achieved remarkable progress on general programming tasks. Nevertheless, their performance degrades significantly in industrial scenarios that require reasoning about hardware semantics, specialized language constructs, and strict resource constraints. To address these challenges, we introduce InCoder-32B (Industrial-Coder-32B), the first 32B-parameter code foundation model unifying code intelligence across chip design, GPU kernel optimization, embedded systems, compiler optimization, and 3D modeling. By adopting an efficient architecture, we train InCoder-32B from scratch with general code pre-training, curated industrial code annealing, mid-training that progressively extends context from 8K to 128K tokens with synthetic industrial reasoning data, and post-training with execution-grounded verification. We conduct extensive evaluation on 14 mainstream general code benchmarks and 9 industrial benchmarks spanning 4 specialized domains. Results show InCoder-32B achieves highly competitive performance on general tasks while establishing strong open-source baselines across industrial domains.

arXiv:2603.22305v1 Announce Type: cross Abstract: Large Language Models (LLMs) are rapidly transitioning from static Natural Language Processing (NLP) tasks including sentiment analysis and event extraction to acting as dynamic decision-making agents in complex financial environments. However, the evolution of LLMs into autonomous financial agents faces a significant dilemma in evaluation paradigms. Direct live trading is irreproducible and prone to outcome bias by confounding luck with skill, whereas existing static benchmarks are often confined to entity-level stock picking and ignore broader market attention. To facilitate the rigorous analysis of these challenges, we introduce CN-Buzz2Portfolio, a reproducible benchmark grounded in the Chinese market that maps daily trending news to macro and sector asset allocation. Spanning a rolling horizon from 2024 to mid-2025, our dataset simulates a realistic public attention stream, requiring agents to distill investment logic from high-exposure narratives instead of pre-filtered entity news. We propose a Tri-Stage CPA Agent Workflow involving Compression, Perception, and Allocation to evaluate LLMs on broad asset classes such as Exchange Traded Funds (ETFs) rather than individual stocks, thereby reducing idiosyncratic volatility. Extensive experiments on nine LLMs reveal significant disparities in how models translate macro-level narratives into portfolio weights. This work provides new insights into the alignment between general reasoning and financial decision-making, and all data, codes, and experiments are released to promote sustainable financial agent research.

arXiv:2510.05318v3 Announce Type: replace Abstract: Large language models (LLMs) have demonstrated remarkable performance on single-turn text-to-SQL tasks, but real-world database applications predominantly require multi-turn interactions to handle ambiguous queries, execution errors, and evolving user requirements. Existing multi-turn benchmarks fall short by treating conversation histories as static context or limiting evaluation to read-only operations, failing to reflect production-grade database assistant challenges. We introduce BIRD-INTERACT, a benchmark that restores this realism through: (1) a comprehensive interaction environment coupling each database with a hierarchical knowledge base, metadata files, and a function-driven user simulator, enabling models to solicit clarifications, retrieve knowledge, and recover from errors without human supervision; (2) two evaluation settings consisting of a pre-defined conversational protocol (c-Interact) and an open-ended agentic setting (a-Interact) where models autonomously decide when to query the user simulator or explore the environment; (3) a challenging task suite covering the full CRUD spectrum for business-intelligence and operational use cases, guarded by executable test cases. Each task features ambiguous and follow-up sub-tasks requiring dynamic interaction. The suite comprises BIRD-INTERACT-FULL (600 tasks, up to 11,796 interactions) for comprehensive performance assessment, and BIRD-INTERACT-LITE (300 tasks with simplified databases) for detailed behavioral analysis and rapid method development. Our empirical results highlight BIRD-INTERACT's difficulty: GPT-5 completes only 8.67% of tasks in c-Interact and 17.00% in a-Interact. Analysis via memory grafting and Interaction Test-time Scaling validates the importance of effective interaction for complex, dynamic text-to-SQL tasks.

arXiv:2603.20670v2 Announce Type: replace Abstract: The rapid growth in the volume, variety, and velocity of geospatial data has created data ecosystems that are highly distributed, heterogeneous, and semantically inconsistent. Existing data catalogs, portals, and infrastructures still rely largely on keyword-based search with limited semantic support, which often fails to capture user intent and leads to weak retrieval performance. To address these challenges, this study proposes a knowledge graph-driven multi-agent framework for intelligent geospatial data discovery, powered by large language models. The framework introduces a unified geospatial metadata ontology as a semantic mediation layer to align heterogeneous metadata standards across platforms and constructs a geospatial metadata knowledge graph to explicitly model datasets and their multidimensional relationships. Building on the structured representation, the framework adopts a multi-agent collaborative architecture to perform intent parsing, knowledge graph retrieval, and answer synthesis, forming an interpretable and closed-loop discovery process from user queries to results. Results from representative use cases and performance evaluation show that the framework substantially improves intent matching accuracy, ranking quality, recall, and discovery transparency compared with traditional systems. This study advances geospatial data discovery toward a more semantic, intent-aware, and intelligent paradigm, providing a practical foundation for next-generation intelligent and autonomous spatial data infrastructures and contributing to the broader vision of Autonomous GIS.

arXiv:2508.09537v2 Announce Type: replace-cross Abstract: The growing capabilities of Large Language Models (LLMs) have led to their widespread adoption for function completion within code repositories. Recent studies on such tasks show promising results when explicit instructions, often in the form of docstrings, are available to guide the completion. However, in real-world scenarios, clear docstrings are frequently absent. Under such conditions, LLMs typically fail to produce accurate completions. To enable more automated and accurate function completion in such settings, we aim to enable LLMs to accurately infer the developer's intent prior to code completion. Our key insight is that the preceding code, namely the code context before the function to be completed, often contains valuable cues that help the model understand the intended functionality. However, inferring intent from such implicit context is non-trivial and constitutes a core challenge in function-level code completion. To tackle this challenge, inspired by how humans interpret context, we propose a reasoning-based prompting framework that guides LLMs to utilize these contextual cues to infer intent step by step. To incentivize LLMs to reason through the preceding code and infer intent, we further curate a dataset of 40k examples, each annotated with intermediate reasoning traces and corresponding docstrings. Extensive experiments on DevEval and ComplexCodeEval demonstrate consistent performance improvements across multiple models, achieving over 25% relative gains in pass@1 for both DeepSeekCoder and CodeLLaMA families. Building upon our framework, we further develop an intent-interactive platform that supports lightweight human feedback. This platform allows developers to select from a set of candidate intentions or edit the intent to better guide the model. Our experiments show that this interactive approach leads to further performance improvements.

arXiv:2603.13131v1 Announce Type: new Abstract: Open-world embodied agents must solve long-horizon tasks where the main bottleneck is not single-step planning quality but how interaction experience is organized and evolved. To this end, we present Steve-Evolving, a non-parametric self-evolving framework that tightly couples fine-grained execution diagnosis with dual-track knowledge distillation in a closed loop. The method follows three phases: Experience Anchoring, Experience Distillation, and Knowledge-Driven Closed-Loop Control. In detail, Experience Anchoring solidifies each subgoal attempt into a structured experience tuple with a fixed schema (pre-state, action, diagnosis-result, and post-state) and organizes it in a three-tier experience space with multi-dimensional indices (e.g., condition signatures, spatial hashing, and semantic tags) plus rolling summarization for efficient and auditable recall. To ensure sufficient information density for attribution, the execution layer provides compositional diagnosis signals beyond binary outcomes, including state-difference summaries, enumerated failure causes, continuous indicators, and stagnation/loop detection. Moreover, successful trajectories of Experience Distillation are generalized into reusable skills with explicit preconditions and verification criteria, while failures are distilled into executable guardrails that capture root causes and forbid risky operations at both subgoal and task granularities. Besides, Knowledge-Driven Closed-Loop Control retrieved skills and guardrails are injected into an LLM planner, and diagnosis-triggered local replanning updates the active constraints online, forming a continual evolution process without any model parameter updates. Experiments on the long-horizon suite of Minecraft MCU demonstrate consistent improvements over static-retrieval baselines.

arXiv:2603.12480v1 Announce Type: cross Abstract: Generative flow and diffusion models provide the continuous, multimodal action distributions needed for high-precision robotic policies. However, their reliance on iterative sampling introduces severe inference latency, degrading control frequency and harming performance in time-sensitive manipulation. To address this problem, we propose the One-Step Flow Policy (OFP), a from-scratch self-distillation framework for high-fidelity, single-step action generation without a pre-trained teacher. OFP unifies a self-consistency loss to enforce coherent transport across time intervals, and a self-guided regularization to sharpen predictions toward high-density expert modes. In addition, a warm-start mechanism leverages temporal action correlations to minimize the generative transport distance. Evaluations across 56 diverse simulated manipulation tasks demonstrate that a one-step OFP achieves state-of-the-art results, outperforming 100-step diffusion and flow policies while accelerating action generation by over $100\times$. We further integrate OFP into the $\pi_{0.5}$ model on RoboTwin 2.0, where one-step OFP surpasses the original 10-step policy. These results establish OFP as a practical, scalable solution for highly accurate and low-latency robot control.

arXiv:2603.12912v1 Announce Type: cross Abstract: Federated Domain Generalization for Person Re-Identification (FedDG-ReID) learns domain-invariant representations from decentralized data. While Vision Transformer (ViT) is widely adopted, its global attention often fails to distinguish pedestrians from high similarity backgrounds or diverse viewpoints -- a challenge amplified by cross-client distribution shifts in FedDG-ReID. To address this, we propose Federated Body Distribution Aware Visual Prompt (FedBPrompt), introducing learnable visual prompts to guide Transformer attention toward pedestrian-centric regions. FedBPrompt employs a Body Distribution Aware Visual Prompts Mechanism (BAPM) comprising: Holistic Full Body Prompts to suppress cross-client background noise, and Body Part Alignment Prompts to capture fine-grained details robust to pose and viewpoint variations. To mitigate high communication costs, we design a Prompt-based Fine-Tuning Strategy (PFTS) that freezes the ViT backbone and updates only lightweight prompts, significantly reducing communication overhead while maintaining adaptability. Extensive experiments demonstrate that BAPM effectively enhances feature discrimination and cross-domain generalization, while PFTS achieves notable performance gains within only a few aggregation rounds. Moreover, both BAPM and PFTS can be easily integrated into existing ViT-based FedDG-ReID frameworks, making FedBPrompt a flexible and effective solution for federated person re-identification. The code is available at https://github.com/leavlong/FedBPrompt.

arXiv:2511.18507v3 Announce Type: replace-cross Abstract: Multimodal large language models (MLLMs) deployed on devices must adapt to continuously changing visual scenarios such as variations in background and perspective, to effectively perform complex visual tasks. To investigate catastrophic forgetting under real-world scenario shifts, we construct a multimodal visual understanding dataset (MSVQA), covering four distinct scenarios and perspectives: high-altitude, underwater, low-altitude, and indoor environments. Furthermore, we propose UNIFIER (mUltimodal coNtInual learning with MLLMs From multi-scenarIo pERspectives), a continual learning (CL) framework designed to address visual discrepancies while learning different scenarios. Compared to existing CL methods, UNIFIER enables knowledge accumulation within the same scenario and mutual enhancement across different scenarios via Vision Representation Expansion (VRE) and Vision Consistency Constraint (VCC). Experimental results show that UNIFIER improves the last-step VQA scores by 2.70%~10.62% and the last-step F1 scores by 3.40%~7.69% compared to the state-of-the-art method, QUAD, in 20-step cross-scenario continual learning tasks. MSVQA dataset is available at https://huggingface.co/datasets/Kaij00/MSVQA.

arXiv:2603.07516v1 Announce Type: cross Abstract: Interaction is one of the core abilities of humanoid robots. However, most existing frameworks focus on non-interactive whole-body control, which limits their practical applicability. In this work, we develop InterReal, a unified physics-based imitation learning framework for Real-world human-object Interaction (HOI) control. InterReal enables humanoid robots to track HOI reference motions, facilitating the learning of fine-grained interactive skills and their deployment in real-world settings. Within this framework, we first introduce a HOI motion data augmentation scheme with hand-object contact constraints, and utilize the augmented motions to improve policy stability under object perturbations. Second, we propose an automatic reward learner to address the challenge of large-scale reward shaping. A meta-policy guided by critical tracking error metrics explores and allocates reward signals to the low-level reinforcement learning objective, which enables more effective learning of interactive policies. Experiments on HOI tasks of box-picking and box-pushing demonstrate that InterReal achieves the best tracking accuracy and the highest task success rate compared to recent baselines. Furthermore, we validate the framework on the real-world robot Unitree G1, which demonstrates its practical effectiveness and robustness beyond simulation.

arXiv:2508.09486v2 Announce Type: replace-cross Abstract: Video Large Language Models (Video-LLMs) have shown strong video understanding, yet their application to long-form videos remains constrained by limited context windows. A common workaround is to compress long videos into a handful of representative frames via retrieval or summarization. However, most existing pipelines score frames in isolation, implicitly assuming that frame-level saliency is sufficient for downstream reasoning. This often yields redundant selections, fragmented temporal evidence, and weakened narrative grounding for long-form video question answering. We present \textbf{Video-EM}, a training-free, event-centric episodic memory framework that reframes long-form VideoQA as \emph{episodic event construction} followed by \emph{memory refinement}. Instead of treating retrieved keyframes as independent visuals, Video-EM employs an LLM as an active memory agent to orchestrate off-the-shelf tools: it first localizes query-relevant moments via multi-grained semantic matching, then groups and segments them into temporally coherent events, and finally encodes each event as a grounded episodic memory with explicit temporal indices and spatio-temporal cues (capturing \emph{when}, \emph{where}, \emph{what}, and involved entities). To further suppress verbosity and noise from imperfect upstream signals, Video-EM integrates a reasoning-driven self-reflection loop that iteratively verifies evidence sufficiency and cross-event consistency, removes redundancy, and adaptively adjusts event granularity. The outcome is a compact yet reliable \emph{event timeline} -- a minimal but sufficient episodic memory set that can be directly consumed by existing Video-LLMs without additional training or architectural changes.

arXiv:2510.05109v5 Announce Type: replace-cross Abstract: Large Multimodal Models (LMMs) are inherently modular, consisting of vision and audio encoders, projectors, and large language models. Yet, they are almost always executed monolithically, which underutilizes the heterogeneous accelerators (NPUs, GPUs, DSPs) in modern SoCs and leads to high end-to-end latency. In this paper, we present NANOMIND, a hardware--software co-design inference framework for Large Multimodal Models (LMMs) that breaks large models into modular ``bricks'' (vision, language, audio, etc.) and maps each to its ideal accelerator. The key insight is that large models can be broken into modular components and scheduled to run on the most appropriate compute units. It performs module-level dynamic offloading across accelerators on unified-memory SoCs. By combining customized hardware design, system-level scheduling, and optimized low-bit computation kernels, we demonstrate our framework with a compact, battery-powered device capable of running LMMs entirely on device. This prototype functions as a self-contained intelligent assistant that requires no network connectivity, while achieving higher throughput and superior power efficiency under strict resource constraints. The design further bypasses CPU bottlenecks and reduces redundant memory usage through token-aware buffer management and module-level coordination. Our system outperforms existing implementations in resource efficiency, cutting energy consumption by 42.3\% and GPU memory usage by 11.2\%. This enables a battery-powered device to run LLaVA-OneVision with a camera for nearly 20.8 hours.