每天知道
你的领域
出了什么

AI-generated video is increasingly used across marketing, product storytelling, and creative workflows, yet automated; high-precision quality control remains a major constraint to scaling production. We present HALLELUAI, an end-to-end system that moderates and regenerates image-to-video outputs to meet expert-level creative standards and deliver ultra-realistic videos with consistent end-user quality of experience (QoE) at scale. The system integrates a video moderation module that evaluates frame-level aesthetics, temporal motion fidelity, and fine-grained hallucination risks relative to the source image, with an agentic regeneration module that iteratively fixes failures through prompt refinement, controlled camera adjustments, targeted model or image switching, and structured retry strategies. The moderation logic is aligned with domain-specific creative guidelines and produces granular, machine-actionable feedback that directly drives regeneration. In human-in-the-loop evaluations with creative experts, HALLELUAI shows strong alignment and reliably outputs ultra-realistic, production-grade videos suitable for product and marketing placements at scale. This framework advances trustworthy AI generated video content by enforcing visual realism, brand safety, and strict input-image fidelity while enabling image-to-video generation at scale.

Synthetic aperture radar (SAR) image generation can mitigate data scarcity, but controllablegeneration under sparse observation angles remains difficult. Recent SAR generative studies im-prove texture realism, yet explicit geometry-aware control is still limited. This paper studiesthe focused and verifiable setting of intermediate-azimuth completion: 3D-model-derived geo-metric priors guide a diffusion model to synthesize the views missing from sparse-angle trainingdata. GeoDiff-SAR constructs a lightweight multi-bounce ray-tracing prior, encodes the result-ing point cloud, and fuses it with text conditioning while adapting Stable Diffusion 3.5 Mediumthrough low-rank adaptation. On a real four-category aircraft dataset, GeoDiff-SAR reaches anSSIM of 0.812 and azimuth consistency of 0.940, compared with 0.738 and 0.782 for the text-conditioned SD3.5 Medium baseline. The same sparse-angle protocol on five MSTAR vehicleclasses yields an SSIM of 0.878 and azimuth consistency of 0.917. These results support theconclusion that a lightweight 3D geometric prior improves viewpoint adherence for controllableSAR generation; it is intended as generation guidance rather than high-fidelity electromagneticreconstruction.

AI-generated content (AIGC) has made significant progress, with 2D generative models becoming ready-to-use tools for the digital fashion industry. However, 3D garment generation remains in its nascent stage, where in the realm of fashion, the semantic information of diverse design elements exhibits intricate coupling relationships in 3D representations, posing substantial challenges for generating diverse 3D garments. In this work, to handle the above problem, We introduce Fashion-3DLR, a novel 3D garment generation framework that utilizes diverse design elements to create high-quality, versatile 3D garment assets. Specifically, to bridge the semantic gaps between different fashion elements, we propose a Garment Feature Fusion Diffusion Transformer (GFF-DiT) module to integrate 2D fashion design elements, e.g., sketch and texture, into latent space. Within the latent space, we then employ a rectified flow transformer to generate geometry latents, which can be decoded into various 3D garment representations, including 3D Gaussians and meshes. Furthermore, we integrate Fashion-3DLR into downstream tasks, achieving the 3D Gaussian Splatting (3DGS)-driven cloth physical simulation and mesh-based virtual try-on. Experimental results indicate that Fashion-3DLR surpass the previous state-of-the-art methods, which verify that the proposed work can generate well-structured, non-watertight garments capable of physical simulation and virtual try-on, underscoring its potential as a versatile 3D garment design tool.

Ambient occlusion (AO) and soft shadows are critical visibility cues for spatial perception in real-time rendering. Hardware ray tracing provides a direct way to evaluate these effects, enabling ray-traced AO and area-light shadows that avoid many limitations of screen-space AO and shadow mapping. However, real-time budgets allow only a few rays per pixel, leaving raw ray-traced estimates noisy and expensive. We present an occlusion-point reuse framework that reuses traced samples in the domain of first-hit occlusion points instead of directly reusing final shading values or light samples. This provides a ray-reuse formulation for AO, rather than merely filtering or reusing completed AO values. The key idea is to transform AO and area-light shadow estimators into occluder-domain integrals, then combine neighboring occluder samples with a multiple-importance-sampling (MIS) formulation. For both AO and shadows, we derive unbiased estimators that validate convergence to the transformed integrals, as well as biased estimators designed for practical real-time execution. The biased variants assume local first-hit occluder consistency; for shadows, this occluder-based assumption better matches local visibility geometry than the visibility-consistency assumption commonly used when reusing light samples. Experiments show higher AO and shadow quality than non-reuse ray-traced baselines, and better shadow quality than light-sample reuse at comparable cost.

We develop a data-driven approximation of the projected ambient connection Laplacian acting on differential forms over smooth Riemannian manifolds sampled by point clouds. The proposed construction extends the classical framework of diffusion maps and Vector Diffusion Maps from scalar functions and tangent vector fields to differential forms of arbitrary degree. Our approach is based on a novel representation of differential forms as alternating differential arrays obtained through an extension of the classical musical isomorphism. This representation enables the construction of a matrix-valued diffusion operator that approximates the projected ambient connection Laplacian directly from point cloud data without requiring a mesh or simplicial complex. The proposed discretization admits the asymptotically optimal kernel bandwidth scaling inherited from diffusion maps, leading to sharper convergence guarantees than previous data-driven approximations of the Hodge Laplacian. Building upon this operator, we derive a fully data-driven explicit Euler scheme for the heat equation on differential forms and validate the proposed methodology through numerical experiments on the unit sphere. The experiments confirm the predicted decay of the analytical solution and demonstrate the effectiveness of the proposed discretization. The proposed framework provides a natural generalization of Vector Diffusion Maps to differential forms of arbitrary degree and establishes a practical foundation for the numerical approximation of geometric partial differential equations directly from point cloud data.

In this paper, we study the extension of Neural Jump ODEs to infinite-dimensional function spaces. In particular, the underlying process $X$ now takes values in $L^2(Ξ, \mathbb{R}^{d_X})$ instead of $\mathbb{R}^{d_X}$ and the Operator NJ-ODE approximates the optimal predictor of this process by producing a representative of the conditional expectation. The NJ-ODE model is a framework for online learning the optimal prediction of continuous-time stochastic processes, given discrete, possibly irregular and incomplete past observations. In a series of works, this model has been extended to deal with generic path-dependent processes, with observation noise and dependent observations, with long-term predictions, and with input-output systems. However, throughout all of these works, the underlying processes were restricted to be finite-dimensional. In particular, function-valued problems, like yield curve or volatility surface predictions, could only be handled through discretization, which inherently leads to a loss of information. In this work, we build on ideas from Neural Operator methods that allow us to extend the NJ-ODE framework to an infinite-dimensional output process. To prove convergence of the NJ-ODE to the optimal prediction process, we develop a new approximation strategy that also generalizes previous works in the finite-dimensional setting by considerably weakening the assumptions.

We introduce a new research area that is called Asymptotics Learning Theory (ALT) and combines optimization with asymptotic analysis. In particular, ALT provides a unified approach for computing unknown constants/parameters in proven asymptotic expansions using optimization theory. In this paper, we focus on a general asymptotic form which includes a broad class of asymptotics. Furthermore, we study two powerful numerical methods, namely, sliding Linear Least Squares (sLLSQ) and sliding Tikhonov Linear Least Squares (sT-LLSQ). For these techniques we rigorously prove asymptotic estimates that lead to sufficient conditions for convergence (to the correct values of unknown parameters) and convergence-rate guarantees. Despite their strengths, both methods have also limitations, e.g., slow convergence---or even, counterintuitively, divergence---in some cases. Moreover, we present fundamental applications in analytic combinatorics, a beautiful field of mathematics that deals with asymptotic enumeration of discrete structures using complex analysis. The proposed techniques complement existing approaches, such as the ratio method and its variants. Numerical examples also verify the theoretical results. Finally, we discuss interesting research directions in ALT.

Mixed continuous--categorical data pose a representation problem for continuous generative models. Flow Matching and Gaussian diffusion operate in Euclidean spaces, whereas categorical laws lie on probability simplices and may be highly imbalanced. We study a logit-coordinate framework that encodes categorical variables as smoothed natural parameters and combines them with transformed numerical variables. This yields common formulations of Logit Flow Matching and Logit Diffusion. We introduce a mixed-distribution discrepancy separating categorical marginal error from conditional continuous Wasserstein error, and derive stability bounds and imbalance-aware nonparametric rates linking vector-field or drift error to decoded mixed-distribution error. Controlled simulations show that scaled-logit coordinates improve or match one-hot coordinates, especially under severe rare-cell imbalance. Across four real-data benchmarks and ten splits per dataset, Logit FM improves the primary distributional metrics on three datasets and is comparable on Churn2; Block-Conditional Logit FM consistently improves the flat model; and Logit Diffusion generally improves over or matches One-Hot Diffusion.

Mean-field games (MFGs) offer a unifying lens on continuous-time generative modeling: a cost tuple recovering twelve prominent models---Continuous Normalizing Flows, OT-Flow, Score-based Models, Schrödinger Bridges, and more---as special cases of one variational problem. Yet two dimensions of this space remain entirely unexplored: the interaction term $\mathcal{I}$ is set to zero in many existing models, and the rich family of MFG solvers has never been applied to generative modeling. We address both gaps with MFGLab an open-source PyTorch library whose primary API is the cost tuple: all twelve models are specified by four composable cost functions, and the training loop, log-Jacobian, and reverse-ODE sampler are shared automatically. We additionally propose DI-Flow, a novel cost design that uses a differentiable entropy functional to encourage mode coverage, and provide learning-based MFG solvers that substantially outperform neural training on stochastic-dynamics rows. Experiments on two 2-D benchmarks confirm that the unified API is lossless relative to hand-coded implementations.

Despite the empirical success of score-based diffusion models, a complete theoretical understanding of how finite-sample learning, network parameterization, and numerical discretization jointly dictate generative quality remains underdeveloped. Existing sampling analyses often evaluate the generative performance conditional on an oracle score or a pre-specified error threshold. In this work, we establish a unified convergence and generalization framework for score-based diffusion models parameterized by practical ResNet-type architectures. We analyze the generalization and convergence properties from the practical finite-sample, discrete-time learning problem of the score function to the ideal continuous-time, population-level objective. Based on the generalization result of the learning problem of score function, we analyze the sampling process induced by the learned score function and provide an end-to-end total variation distance estimate for the generated terminal distribution. This estimate explicitly decomposes the overall generative error into four interpretable components: the truncation error of the forward process, the reverse-time discretization error, the generalization error incorporating both finite data and forward-time discretization, and the training optimization gap. Our results quantitatively characterize how the training sample size, temporal discretization grids, and optimization accuracy jointly control the final fidelity of samples generated by diffusion models.

Large language models now power autonomous agents capable of complex, multi-step tasks in different environments. Accurate and reliable execution of these tasks requires the agent to predict the results of its actions. Recent research proposes to enhance predictive capabilities via specially trained environment simulators-world models. While world models can improve performance, they can also mislead agents into executing harmful actions, creating significant security and privacy risks. In this paper, we raise security concerns regarding the usage of world models in agentic systems. We discover a range of world model specific vulnerabilities, which can be exploited in terminal-based agents to execute malicious code or extract sensitive data. To facilitate future development, we introduce a security benchmark dataset designed for text-based world models. We argue that some risks are intrinsic to approximate world modeling, and show that attackers can induce mispredictions in agentic pipelines with up to 95% success rate, possibly resulting in unintended command execution, denial of service, drainage of wallet and private information extraction. Finally, we provide practical recommendations for practitioners to mitigate the discovered harms and harden agentic systems.

Language agents can now interact fluently with users in software, but robots still struggle to bring comparable interaction to physical tasks. Current robot-control paradigms, including vision-language-action policies and world-model-based planners, are mainly optimized for instruction execution, leaving users with little visibility into why an action is chosen and few mechanisms to redirect, correct, or teach the robot through interaction. To solve this problem, we present the World-Cognition Model (WCM), a human-centered embodied agent built on the SLAK architecture (Sensing, Logic, Action, and Knowledge) and an asynchronous runtime. SLAK separates perception, reasoning, control, and memory, while the runtime allows reasoning, dialogue, and execution to proceed concurrently. WCM further introduces a human-in-the-loop teaching mode that enables users to interactively teach the robot difficult or long-horizon tasks. Teaching episodes and autonomous task rollouts are refined into chain-of-thought supervision to continually improve the model. WCM achieves a 73.8% average success rate across nine real-world human-robot interaction tasks, including tasks held out from CoT fine-tuning and a long-horizon task learned through teaching.

Deciding whether a trajectory actually fulfills its instruction governs how we measure computer-use agents on long-horizon graphical-user-interface tasks and how we train them with reinforcement learning. This judgment has long relied on rule-based evaluation, which struggles to align with human intention and goes stale when an app updates or its online content drifts. Existing model-based judges attempt to address these problems but still leave a performance gap to the rule-based evaluation. We propose the SeekJudge framework, in which four role-specialized agents, a Condense, a Ground, a Seek and an Analyze agent, reach a verdict through a Seek--Analyze loop over the trajectory. A seed-calibrated distillation pipeline trains one specialized $9$B model to serve as the shared backbone for all four agents. Measured by downstream success rate on held-out RL test goals, SeekJudge is the first practical model-based reward to match or surpass native rule-based supervision in online RL. Beyond accuracy, SeekJudge provides step-level judgments, runs far cheaper than a closed-source large model, and keeps a small per-call context that scales to much longer trajectories. We further contribute a general architectural improvement to the reward server that speeds up judging in RL. Together these make model-based reward a practical drop-in for rule-based supervision in CUA reinforcement learning.

Rare diseases collectively affect an estimated 3.5% to 5.9% of the population, yet more than 70% of patients are misdiagnosed and many endure years of evaluation before a diagnosis is reached, because early presentations are nonspecific and relevant expertise is scarce and unevenly distributed. Artificial intelligence could provide support, but existing systems address isolated stages of care, overwhelmingly diagnosis. They typically depend on the results of downstream investigations, and they treat the variability between models as noise to be eliminated. Here we present RareLens, a system that supports clinical decision-making across the entire rare disease trajectory by exploiting this variability. When heterogeneous large language models evaluate the same case, they generate divergent but complementary reasoning, which RareLens aligns and calibrates into a single convergent, actionable decision at each stage. Four coordinated modules perform primary-visit risk screening, diagnosis, treatment planning and prognosis. Developed and evaluated on RareBench, a real-world dataset of 157,525 cases spanning all 33 Orphanet categories and more than 7,000 conditions, RareLens outperformed every frontier model tested, including GPT-5, DeepSeek-R1, Claude-3.7-Sonnet and Gemini-2.5-Pro, at each stage. It achieved an area under the curve of 0.917 for screening and top-1 accuracies of 65.5% and 89.8% for diagnosis and treatment. In an external study spanning 1,287 cases and 23 physicians, autonomous RareLens and physicians assisted by RareLens both substantially outperformed unaided physicians. These findings indicate that aligning divergent model reasoning, rather than scaling a single model, offers a generalizable strategy for high-uncertainty clinical decision-making.

Creating a reusable tool is an investment: an agent pays a fixed cost now in exchange for the potential of future reuse. Therefore, a user should prefer an agent that creates a small number of highly reusable tools, rather than many one-offs. We introduce a paired benchmark that tests whether LLM agents exhibit conscious allocation behavior under a fixed budget in two contexts: an abstract text-based formulation and a code-construction task. We find that every frontier model we test---Claude Haiku, Claude Opus, GPT-5.4-mini, and GPT-5.6 Sol---acts near-optimally in the abstract framing but fails to transfer this ability to script-writing. Through further experiments, we identify the particular failure modes for each model. Notably, the first three models fail even when the scripts are not evaluated, while GPT-5.6 Sol stays selective under that weaker manipulation and collapses only at full construction. Furthermore, an open-source Qwen model policy-trained for abstract allocation generalizes this ability across held-out lexical variations, but sees no improvement at script allocation. Together, these results establish online tool allocation as a significant capability boundary, even for modern frontier models.

Federated learning enables collaborative model training across distributed clients without centralising their data, yet privacy remains a persistent concern because the shared model updates can leak information about local datasets. Existing privacy-preserving methods either inject calibrated noise into client updates, limiting their composition guarantees, or formulate client privacy choices as a multi-agent game whose Nash equilibrium becomes intractable as the number of clients grows. We bridge these two lines of work by formulating privacy-preserving federated learning as a mean-field privacy game: each client strategically chooses its own privacy budget while interacting with the population only through a single mean-field statistic. The mean-field limit yields a tractable equilibrium for arbitrarily many clients, accommodates heterogeneous client preferences, and inherits an exponentially decaying privacy guarantee through a log-Sobolev contraction. The framework recovers the entropic privacy baseline as the homogeneous special case and the multi-agent privacy game as the finite-population case. Experiments on quadratic regression, logistic regression, and MNIST demonstrate that the proposed framework attains the privacy-utility trade-off of the entropic baseline while delivering a personalized privacy guarantee that the homogeneous baseline cannot express.

AI is evaluated through knowledge, reasoning and plan generation, yet scientific agency requires reliable physical action and adaptation to evidence. Here, we use a robotic chemistry laboratory as a physical-world testbed to make scientific agency measurable. Its 45 modular workstations exposed as machine-readable skills enabled 4,608 trials. Only 3.3% of trials produced expert-assessed executable workflows under laboratory constraints; even the best system achieved 28.1%. Long-horizon planning remained a challenge: only three executable workflows exceeded 30 operations, although the longest contained 44. Across five rounds, experimental feedback prompted local adjustments but no workflow-level replanning or analytical-method redesign. By making physical executability and evidence-driven replanning measurable, our study provides an evidence-based assessment of deployment readiness and a diagnostic framework to guide closed-loop improvements towards physically grounded autonomous research.

LLM agents that operate over many turns accumulate facts in an external memory store and reuse them as premises for downstream reasoning. A hallucinated fact written at one step therefore persists as a false premise for every subsequent step, a failure mode we call memory contamination. Existing memory management addresses retrieval and capacity but not write-time correctness; this admission problem cannot be solved by utility- or recency-based criteria, and uncontrolled contamination compounds across long trajectories. We propose ConsistencyGate, a write-time admission gate that, before committing a candidate fact m extracted from context c, queries the LLM K times for a soft support score and admits m only when the average exceeds a threshold. The mechanism is model-agnostic, requires no fine-tuning, and reduces to a single forward pass in a log-probability variant for latency-sensitive deployments. To measure the effect on natural data, we construct two real-conversation benchmarks (LoCoMo-Contam and MSC-Contam) by planting controlled single-detail corruptions in long-term conversations from LoCoMo and MSC, and complement them with a structured synthetic corpus (MemContam) that isolates a near-oracle upper bound. Across four LLM backbones, ConsistencyGate reduces contamination on every benchmark relative to a write-everything baseline, with the cost concentrated on facts that are stated only implicitly in the source context. We release all three benchmarks together with the gate implementation.

Existing evaluations of large language models cover knowledge, reasoning, coding, and tool use, but they rarely treat a verifiable deliverable produced within a constrained workflow as the unit of evaluation. We introduce SQBench, a benchmark for evaluating production-oriented task delivery by language-model agents. SQBench v1.0 contains 220 standardized tasks organized into L1 atomic capabilities, L2 composite skills, and L3 business scenarios. Each task requires an agent to process input assets, use available tools, and produce an explicitly specified deliverable. The evaluation first computes functional Completion and then derives Risk Penalty and Performance from independently evidenced triggers in a 10D Risk Matrix. A Strict Pass requires Completion = 1 and Risk Penalty = 0. We evaluate 27 model configurations under a common protocol, with one run per configuration-task pair. The highest prespecified Weighted Pass@1 is 60.5%. Mean Strict Pass@1 on L3 is 18.5%, and every configuration performs worse on L3 than on both L1 and L2, indicating that delivery under domain constraints is a shared weakness within the current task set. Of 2,348 results with Completion = 1, 113 (4.8%) fail the Strict Pass criterion because of risks such as unverifiable citations, inappropriate resource use, or format violations. These results show that functional completion alone does not fully characterize delivery quality and that risk determinations should be reported separately.

Test-time search lets small video diffusion models rival larger ones, but costs 2-10x more. All candidates are fully denoised, although most are discarded. Training-free caching makes each rollout 2-3x faster at near-lossless quality. Composition is safe only if lossy caching preserves verifier rankings. We present the first study of whether caching corrupts candidate ranking in video test-time search. On Wan2.1-T2V-1.3B with an adaptive caching wrapper (~2x per-candidate speedup), ImageReward scores seed-matched cached and full rollouts. Median per-prompt Spearman rank correlation is 0.905, with 72% top-1 agreement on the VBench suite. VBench-2.0 replicates this result on a harder suite. Recomputing the cached winner at full compute retains 90-94% of the full-search gain. Errors cluster among near-tied candidates, making corruption self-limiting. This finding leads to CachedSearch. It explores every candidate with aggressive caching, then re-generates only the winner at full compute. At N=8, it captures 94.7% of best-of-N's gain at 63% of the cost. Capture rises with width. At matched budget, it searches twice as wide for 38% more gain. The result holds from 1.3B-14B across six models and four families: Wan, LTX, CogVideoX, and Hunyuan. Wan2.1-14B matches the 1.3B model's fidelity. Mid-trajectory pruning multiplies the exploration saving to 3.11x at 88.6% capture. Ports to other model families require recalibrating a single parameter, showing that fidelity tracks architecture rather than parameter count. CachedSearch is training-free, verifier-agnostic, and orthogonal to the search algorithm, making it a plug-in multiplier for test-time scaling.

We present a unified fixed-point construction of the elementary transcendental functions, encompassing the real exponential, the complex exponential (sine and cosine), and the natural logarithm. Each function is characterized as the unique solution of a duplication identity established through the Banach contraction principle. These foundational identities are $e(2x)=e^2(x)$ for the exponentials, and $\log(x^{2})=2\log x$ for the logarithm. Because a direct iteration of these identities is numerically unstable, owing to local expansiveness at the target, the central idea transfers the analysis to a residual function, on which the operator becomes a strict contraction with an explicit convergence rate. Beyond its theoretical economy, which dispenses with differential equations and power series, this characterization translates into efficient algorithms for the machine evaluation of elementary functions: the underlying framework yields floating-point kernels whose accuracy and iteration depth are governed by the theoretical contraction rate. We also present a computational study showing that, in a throughput-bound vectorized regime, these kernels are competitive with standard production libraries, and in favorable configurations exceed them, with the sine--cosine kernel faster at every tested iteration depth. These implementations operate without lookup tables or memory traffic, an architectural advantage for modern high-performance and energy-efficient computing.

Fine-grained, device-initiated communication lets persistent GPU kernels in distributed diffusion transformer (DiT) inference issue remote stores and overlap data movement with Tensor Core computation. Existing systems schedule when communication is issued and when received data becomes consumable, but omit post-issue progress before remote-visible completion, making sender backpressure hard to predict. We identify X-Stage, a software-visible post-issue pipeline stage. Measurements on an eight-GPU node with a recent NVIDIA architecture show that short remote-store bursts drain as the issuer resumes work, whereas sustained injection exhausts finite outstanding capacity and delays later issues. A lightweight Burst-Gap model parameterized by backpressure-free issue time, effective drain rate, and outstanding capacity predicts issue overhead, recovery between bursts, and the onset of backpressure. Guided by the model, we redesign two communication-computation fused kernels. For DeepGEMM MegaMoE, interleaving Linear-1 and Linear-2 work across expert waves places computation between concentrated remote-store bursts, yielding a 1.18x geometric-mean and 1.62x maximum kernel speedup over the Expert-Wave baseline across 84 configurations. For Ulysses sequence-parallel attention, tile-granular fusion of the post-attention All-to-All with FlashAttention lets an output-tile owner issue remote stores and resume computation without a dedicated communication warp or streaming multiprocessor. FlashAttention-3 and FlashAttention-4 reach maximum sender-visible speedups of 1.43x and 1.42x over serial execution, and at long sequences their steady-state times approach those of FlashAttention alone. These results establish post-issue progress as a measurable scheduling lever for shaping bursts, avoiding backpressure, and hiding sender-side overhead.

Recent high-resolution Multimodal Large Language Models (MLLMs) generate thousands of visual tokens per input, leading to a visual token explosion that introduces severe latency bottlenecks. While token pruning mitigates this issue, state-of-the-art subset-optimization methods typically rely on iterative subset construction to jointly capture visual diversity and instruction relevance. As visual token counts scale, this sequential dependency introduces significant selection overhead, severely limiting the translation of theoretical FLOPs reductions into actual wall-clock speedups. To address this limitation, we propose Single-Forward Pruner (SFPruner), a structural reformulation of visual token pruning that embeds redundancy control directly into the scoring space, bypassing the need for iterative combinatorial optimization. Our non-iterative framework achieves redundancy-aware importance selection in a single forward pass through two complementary mechanisms. First, to attenuate redundancy at the covariance level, we introduce a semantics-guided ridge leverage scheme. By integrating instruction relevance and visual saliency, this mechanism suppresses dominant covariance directions and mitigates representation bias. Second, ranking-based directional masking resolves residual overlap through asymmetric similarity competition, where higher-scoring tokens explicitly suppress redundant lower-scoring alternatives via parallel tensor operations. Extensive evaluations demonstrate that our approach maintains stable selection costs, reducing the token selection process by up to 110 ms, from 112.4 ms to just 2.5 ms at 512 tokens in Qwen2.5-VL. This structural efficiency successfully translates theoretical token reductions into tangible inference speedups while preserving highly competitive performance against state-of-the-art techniques under aggressive compression.

Recent advances in large language models (LLMs) have enabled automated kernel generation and optimization, but most existing approaches rely on surface signals such as compilation feedback and profiling metrics. These signals reveal that a kernel is slow, but not why the backend compiler fails to realize a profitable optimization, especially on emerging accelerators such as NPUs. We therefore formulate kernel optimization as a progressive cross-layer diagnosis problem that links runtime symptoms to IR structure and compiler behavior before rewriting source. Based on this insight, we present our system, a compiler-grounded and hierarchical optimization framework for Triton kernels. the system escalates from lightweight pattern triage and profiling diagnosis to IR attribution and compiler-grounded analysis only when deeper evidence is needed, then proposes evidence-backed source-level rewrites. We implement the system on Triton for Ascend NPUs and evaluate it on 37 successfully converted entries from a standardized NPUKernelBench-derived Ascend 950 benchmark. Across these entries, the system attains a geometric-mean speedup of 4.35$\times$ and a median speedup of 2.73$\times$ from the initial to optimized Triton kernel; 22/37 exceed 2$\times$ and 13/37 exceed 5$\times$. The complete distribution ranges from near-baseline entries to large wins, motivating transparent reporting of the current system's scope and limitations.

Multi-head Latent Attention (MLA), introduced in DeepSeek-V2, compresses key-value pairs through a shared low-rank bottleneck (cKV), achieving 81% KV-cache reduction during inference. Despite its adoption in massive production models, no prior work has studied what information this bottleneck preserves or discards, nor how it reshapes internal transformer circuits. We present the first comprehensive mechanistic interpretability study of MLA, training a 114M-parameter transformer (pretrained on a web/code/math mixture, fine-tuned on TinyStories) and analyzing its representations through SVD, attention head taxonomy, linear probing, and a disruption-attribution analysis. Our key findings are: (1) the cKV bottleneck learns a pure content representation, preserving entity identity (98% retention) while discarding positional information, validating MLA's separation of content from position via RoPE; (2) induction heads co-locate at a single layer (Layer 12), unlike their distributed formation in standard MHA; (3) a single "semantic hub" layer (Layer 15) simultaneously exhibits the highest SVD effective rank and strongest disruption-attribution score; and (4) the bottleneck is globally over-provisioned, using only 46% of its capacity on average. These findings suggest MLA does not merely compress attention passively, but reshapes how the model organizes content, position, and circuit structure. We view this as an initial data point and detail scope limitations in Section 5.

This paper aims to enable computation- and communication-efficient GPU sharing across devices within local area networks (LANs), facilitating ubiquitous AI inference on heterogeneous personal devices. We achieve distributed task offloading via CUDA API remoting. However, beyond raw computation, network constraints emerge as the primary bottleneck: limited bandwidth, high-frequency API invocations, and cross-task contention significantly hinder performance. To address these challenges, we propose Gleam, a novel and network-efficient framework for task-generic GPU sharing across local-area CUDA devices, with three key contributions. First, we reduce bandwidth overhead in CUDA API remoting through automatic model weight caching, and mitigate accumulated latency from frequent API calls by asynchronous execution. Second, we design a runtime task scheduler that dynamically determines API remoting pairs between LAN clients and servers, explicitly accounting for both network conditions and GPU resource contention under parallel workloads. Finally, we introduce dedicated mechanisms to ensure CUDA context consistency across distributed executions. Extensive experiments on heterogeneous NVIDIA GPUs and diverse AI workloads show Gleam consistently outperforms state-of-the-art baselines, achieving 1.4-24.2 times improvements in API remoting efficiency and up to 1.79 times higher system throughput.

Mixed-precision quantization improves the accuracy of post-training quantization by allocating higher bitwidths to sensitive layers, but existing methods solve the allocation for a single fixed memory budget. In practice the budget varies across deployments and is unknown at calibration time. Adaptive quantization addresses this with one offline calibration that serves any budget, yet current methods score layer sensitivity in a manner that does not consider its dependency on quantization levels of other layers. We show that a layer's sensitivity depends strongly on the bitwidths of its upstream layers and that this dependence shifts the resulting preferred bit allocation. We propose MixQuant, a technique-agnostic adaptive framework that wraps any base quantizer. MixQuant marginalizes each layer's distortion over random quantized upstream configurations to obtain budget-agnostic scores, calibrates the quantizer's parameters on plans the allocator itself produces, and penalizes allocations that leave layers at the lowest bitwidths. A single greedy pass then serves any budget at deployment. Across Llama-3.2-3B, Llama-2-7B, and Mistral-7B under AWQ and GPTQ, MixQuant outperforms adaptive and mixed-precision baselines in every setting, improving average accuracy by up to 8 points and reducing perplexity from 12.43 to 10.70 at the tightest budget, while matching an ILP solver at negligible deployment cost.

Existing token compression methods for omnimodal large language models typically rely on one modality to determine what to retain in the other. We show that this assumption often breaks down: for the same query, audio and video relevance often peaks at different moments. This cross-modal salience mismatch makes unidirectional guidance prone to discarding answer-critical cues under aggressive compression. We propose OmniScope, a training-free token compression framework that uses the query as a shared semantic anchor while estimating relevance separately for audio and video. OmniScope allocates modality-specific token budgets, prunes visual tokens with an anchor-delta strategy that preserves both global context and temporal changes, and merges audio tokens within each second to reduce redundancy while maintaining temporal continuity. Across four audio-video benchmarks and two Qwen2.5-Omni model scales, OmniScope achieves the best average accuracy across all compression settings. At 25% overall token retention, it delivers up to 3.53x prefill speedup and more than 15% GPU memory reduction, with only a 0.35-point drop in average accuracy. These results suggest a simple design principle for OmniLLM inference: share the query across modalities, but not the salience estimates. The code is available at https://github.com/MAC-AutoML/OmniScope.

On x86, kernel dispatch fragments the outputs of the same neural network into many equivalence classes across hardware. We ask whether the same fragmentation governs ARM edge inference, where most edge ML actually runs. Across four Raspberry Pi devices spanning Cortex-A53, A72, and A76 under ONNX Runtime CPU, microarchitecture is not observable in the outputs of a fixed FP32 CNN. Holding hardware constant at Cortex-A76 and switching only the execution provider, FP32 outputs disagree on every CIFAR-10 image with a mean remaining precision of 14.97 of 23 mantissa bits. INT8 QDQ post-training quantization collapses both axes to a single equivalence class. We trace this to a structural property of QDQ graphs that we call H1+H2: discrete-grid inputs make any Conv dispatch-deterministic (H1) and QuantizeLinear at every layer boundary preserves that precondition (H2). H1+H2 predicts that bit-exact agreement should extend to production CNNs under runtimes that confirmably exercise different ARM microkernels. We verify this on MobileNetV2 and ResNet50V2 under TensorFlow Lite with XNNPACK, where timing evidence confirms SDOT dispatch on A76 and NEON multiply-accumulate on A72 yet every intermediate INT32 accumulator and every final output is byte-identical across 500 ImageNet images per model. We then identify the specific x86 mechanism that breaks the same invariant on x86, namely PMADDUBSW saturating INT16 intermediates, which has no ARM analogue. The Schlögl et al. divergence phenomenon is delineated rather than contradicted. For practitioners deploying quantized CNNs across heterogeneous ARM fleets, the operational consequence is direct. INT8 inference is the reproducible mode and the relevant behavioral variation axis is precision, not microarchitecture.

Mixture-of-Experts (MoE) inference consists of sparse expert GEMMs whose shapes vary with the runtime routing distribution. Existing serving systems typically select fused-MoE kernels using static token-count buckets, ignoring the per-expert routing distribution that determines tile padding, memory reuse, and kernel efficiency. We introduce a distribution-aware framework for modeling and benchmarking MoE inference. The framework combines the compact Effective Experts metric with a Dirichlet-based reverse-modeling procedure that generates controllable routing distributions for systematic hardware studies. Using it, we show that the best fused-MoE kernel changes with routing skew and token count. We further present DA-MoE, a GPU-resident kernel-dispatch runtime for NVIDIA GPUs that matches the live routing histogram to offline-tuned distributions and selects a near-optimal fused-MoE kernel without CPU--GPU synchronization. On HumanEval-X serving traces, DA-MoE improves geomean fused-MoE latency by 1.16X on DeepSeek-V3 and 1.29X on Kimi K2, with peak speedups of 1.40X and 1.56X.