生成模型与LLM推理优化

Sparse matrix kernels are fundamental to scientific computing, graph analytics, and machine learning. Their GPU performance depends strongly on the input sparsity pattern and execution strategy. For the same SpMM on the same matrix, cuSPARSE exhibits a 350x performance gap between CSR and Blocked-ELL. Our study of multiple data formats, specialized systems, and sparse compilers shows that no single implementation consistently dominates across sparsity patterns and operators. This motivates a system that can adapt its representation, execution strategy, and hardware mapping to each workload and target GPU. We present SparseDitto, an LLM-based system that constructs a GPU kernel for each matrix, operator, and target GPU. SparseDitto supports SpMV, SpMM, and SpGEMM within a unified design framework. A lightweight additive model ranks established strategies using structural features of the input matrix. An architecture-aware planner then proposes several candidate designs. Coding and verification agents implement and refine them using measurements from the target GPU. Across three sparse operators and a diverse set of matrices, SparseDitto achieves a geometric-mean speedup of 2.68x over cuSPARSE on an NVIDIA RTX PRO 6000 GPU, with a maximum of 146.61x. On an NVIDIA H200 GPU, it achieves 2.79x, with a maximum of 78.5x. Its generated SpMM kernels also accelerate full-batch GCN training by up to 3.39x.

In partially observable reinforcement learning, agents face a dual bottleneck: they must explore to encounter rewarding states and retain that experience in memory to optimize their policies. Exploration bonuses and memory architectures are traditionally evaluated in isolation, leaving their interaction unmeasured, and standard notions of sparse reward conflate temporal signal density with what the reward actually supervises. We present a controlled study crossing episodic exploration bonuses with diverse neural memory architectures across three environments that vary how the content of memory is acquired. An identical bonus signal yields three distinct interaction patterns: it amplifies architectural capacity differences where memory content must be actively discovered and retained unsupervised; equalizes architectures to a shared ceiling where the content, once sought out, is a single reward-supervised cue; and is null where the observation stream is purely scheduled. Controlled reward manipulations verify that these patterns track reward structure rather than density: a dense reward neutralizes a bonus only if it directly supervises the required latent memory, and a small avoidable penalty on exploratory actions (leaving the optimum unchanged) induces policy convergence to suboptimal stationary states, which either bonus resolves. We then formalize reward sparsity with observation-anchored reward machines, separating structural sparsity (an automaton reproduces the return without the task-required history) from potential sparsity (the one-step reward misprices local exploratory actions); the resulting vocabulary organizes the three regimes by the retention burden each task exposes. Together, these results show exploration and memory are complements, not substitutes: a bonus induces exposure, and only memory converts exposure into return.

Achieving local differential privacy in distributed optimization while maintaining low communication cost remains challenging. Existing vector quantization methods, such as vqSGD, use high-dimensional geometric constructions but incur unfavorable dimension-dependent variance. In this work, we propose Subsampled Stochastic TurboQuant (SSTQ), a framework that combines overcomplete equal-norm tight frames, coordinate subsampling, and privacy-aware one-dimensional quantization. SSTQ includes two variants: a Flat Randomized Response version and a Metric-Aware Laplace version, the latter being better suited to higher codebook bit-width regimes. We show that SSTQ achieves optimal mean squared error scaling while using only $\lceil \log_2 N \rceil + b$ bits per client, where $N = Θ(d)$ is the frame size. We also derive a surrogate privacy-aware codebook objective that reduces the codebook-dependent MSE scaling from $O(4^b)$ to $O(2^b)$. Finally, we empirically evaluate SSTQ against established baselines on federated learning tasks using CIFAR-10 and Fashion-MNIST, demonstrating favorable utility and communication efficiency. Some of the analytical derivations were first obtained using a fully automated Gemini-based agentic system developed internally at Google. The authors have verified those derivations and edited them for clarity of presentation.

Reinforcement learning (RL) post-training improves the reasoning capabilities of large language models, but autoregressive rollout generation remains a major efficiency bottleneck. Speculative decoding can accelerate generation, yet applying it during RL is difficult because the target policy continually evolves: static proposers become stale, while frequent drafter updates add substantial overhead. We introduce SpecRoll, a speculative rollout engine that preserves the target model's sampling distribution while adapting at two timescales. Lightweight future-token heads generate parallel proposals, while our proposed Reflex module uses delayed verifier feedback to perform bounded, trajectory-local hidden-state corrections without backpropagation. A complementary slow path updates the head parameters only when sustained degradation is detected. SpecRoll combines these mechanisms with concurrency-aware sparse-tree verification and exact target verification, leaving the target rollout distribution and GRPO objective unchanged. Across five models ranging from 1.5B to 14B and three mathematical reasoning datasets, SpecRoll achieves 1.26-2.15x generation speedup and 1.21-2.04x end-to-end speedup over vanilla GRPO. It also outperforms FastGRPO in both generation and end-to-end time across all 15 matched settings, with an average pairwise end-to-end gain of 1.18x. Controlled ablations show that the fast and slow adaptation paths provide complementary benefits. Our source code is available at https://anonymous.4open.science/r/SpecRoll-26062006.

Benchmark scores are reported as properties of a model, yet the inference framework used to produce them, such as HuggingFace, vLLM, or Ollama, are considered non-influential and their names and versions are almost never disclosed. In this work we investigate how much this choice can influence the model output. In a fully-crossed study (three instruction-tuned models x five inference frameworks x six benchmarks x four generation modes) we investigate how different tools (wrappers/backend) influence benchmark scores and how their score changes is influenced by generation hyper-parameters. We find backend to be a non-negligible factor where even under greedy, sampling-noise-free decoding, changing the backend can significantly alter models performance and this effect is structural and strongly model-dependent. Decomposing the variance according to generation mode reveal that considerable portion of the variability (roughly 39%) a practitioner sees out-of-the-box can stem from the backend, while the remaining stems from sampling noise and each framework's default generation parameters, both of which are avoidable by disclosing and matching the generation configuration. These divergences are more pronounced on factual than on social-bias benchmarks. Overall, benchmark numbers are not backend-agnostic therefore, we recommend disclosing the backend, its version, and the full generation configuration, also using deterministic decoding for cross-backend comparison.

Deep neural networks have shown impressive success in NLP tasks owing to their complex structure and huge number of edges. Achieving state-of-the-art performance in natural language processing with a large pre-trained model such as BERT is expensive and time-consuming, carries a large carbon footprint, and is difficult to realize on machines with minimal computational capability. This creates a barrier to training complex models for resource-constrained languages such as Bengali. However, in a complex neural model, not all edges are equally impactful, and the contributions of some of them can be neglected. Pruning promises to reduce the memory footprint of regular networks, shorten the training time of ever-growing networks, and increase inference efficiency without sacrificing comparable performance. In this work, we introduce BnBERT-iPET, a sparse few-shot language modeling approach for Bengali, and experimentally show that a lightweight few-shot-learned language model retaining only 10% of the edges of an initial model such as BERT can perform neck and neck with much larger models on challenging tasks for a resource-constrained language such as Bengali. By learning from few shots through iterative pattern exploiting training and achieving 90% sparsity with the Lottery Ticket Hypothesis pruning technique, our pruned BnBERT-iPET model proves to be a tough competitor to state-of-the-art language models such as Bangla Electra, Indic-BERT, and XLM-RoBERTa on downstream tasks over standard benchmark datasets of the Bengali language.

World Action Models (WAMs) improve robot manipulation by learning how the environment evolves beyond the current observation. However, existing approaches face a fundamental dilemma: Joint-WAMs preserve future-aware representations during inference but incur prohibitive computation costs, while efficient alternatives remove future modeling at inference time and may lose the robustness benefits of temporal reasoning. In this work, we revisit the role of future representations in WAMs and show that inference-time future conditioning is critical for generalization under distribution shifts. This observation motivates Faster-WAM, an efficient future-conditioning WAM that preserves future representations while avoiding expensive video-action interaction. Faster-WAM introduces a sparse future-conditioning framework that computes future representations once and selectively reuses them throughout action denoising. Specifically, we propose SparseMoT to replace ubiquitous layer-wise fusion with selective video-action interaction at a compact subset of network stages, and Interval KV-Fusion to aggregate multi-depth future representations without increasing attention complexity. Experiments demonstrate that Faster-WAM achieves a substantially better performance-efficiency trade-off than existing WAMs. On the out-of-distribution LIBERO-Plus benchmark, Faster-WAM improves success rate from 49.14% to 73.57% compared with Fast-WAM, while running 2.21$\times$ faster than Joint-WAM. It further achieves state-of-the-art performance on LIBERO and RoboTwin 2.0, while demonstrating strong robustness in real-world manipulation.

3D vision-language models (3D VLMs) enable spatial reasoning over multi-view scenes but suffer from substantial token redundancy due to duplicated observations and large uninformative regions, leading to high computational cost. Although visual token compression has shown promise in accelerating 2D VLMs, it fails to capture the structured nature of 3D scenes and leads to incomplete spatial coverage and loss of fine-grained details. In this paper, we propose HiSC, a training-free framework for hierarchical spatial clustering token compression in 3D VLMs. HiSC lifts token compression from token-level selection to cluster-level processing by organizing tokens into spatially grounded clusters using joint geometric and semantic cues. Specifically, we first introduce a spatial graph-based merging (SGraM) strategy that models cross-view redundancy as spatial connectivity and consolidates physically consistent regions, effectively merging extremely similar redundant tokens prior to LLM inference. We then propose a spatial clustering-based pruning (SCluP) paradigm within LLM inference, which performs hierarchical compression across clusters and within clusters, preserving object instance completeness while retaining fine-grained details for important regions. Extensive experiments on diverse 3D reasoning benchmarks show validate the effectiveness of HiSC, particularly under high visual token pruning ratios. Besides, HiSC achieves over 90% token reduction with minimal performance degradation. Code is accessible at https://github.com/elecreak/HiSC.

Large language model (LLM) agents repeatedly process long, privacy-sensitive contexts, while cloud-only deployment exposes user data beyond the trusted endpoint and fully local deployment often requires costly hardware. Split inference offers a middle ground by executing the model head, tail, and tools locally and the middle layers in the cloud, but its local-cloud-local path transfers boundary hidden states at every invocation and creates a critical communication bottleneck. We present \system, a reference-aware codec that retrieves exact-token historical spans for prefill uplinks, reuses the reconstructed uplink state for same-round prefill downlinks, and generates boundary-specific decode references with lightweight causal predictors. RAC applies grouped affine alignment and calibrated residual quantization with optional prefill outliers, while sender-side wire-format reconstruction synchronizes subsequent references and offline calibration accounts for quality and packed representation costs. Across three models and nine evaluated model-link pairs, Raw-to-RAC mean time to first token (TTFT) and time per output token (TPOT) ratios are 1.24-2.72$\times$ and 1.01-2.79$\times$, while the 12 non-perplexity task-score changes range from $-0.40$ to $+2.50$ points.

Multi-agent LLM systems relay key--value caches instead of text and credit their gains to exchanged ``latent thoughts''. That credit is a claim about which example's cache is relayed, not merely that one is. We audit it causally in released systems. The cache is replaced with deranged (mismatched-example), zeroed, and moment-matched random counterparts, under two regimes defined by whether the receiver needs the sender's private information. Where it does, the battery reads ceiling: 100% against 23--25% for answer-irrelevant relays on the primary backbone, a contrast replicated across three families, five checkpoints, and a prose document-QA surface. Where it does not, a pre-registered five-seed protocol establishes equivalence within 2.8 points, a margin anchored to the audited system's reported gain, under Holm-corrected TOST on GSM8K and ARC-Challenge across three Qwen3 scales and on MedQA at 8B (one cell shows a small detected advantage inside the margin); a second family shows no detected advantage. A large cache effect need not be a pairing effect. In one natural cell, zeroing the relay costs 14.7 points; a mismatched cache, 0.4. Nor is need sufficient: under the same test, delivered channels span ceiling (LatentMAS's native relay), partial (KVComm's layer subset), and no detected example-specific transfer (C2C's released projector). Benchmark deltas do not by themselves establish latent-thought transmission; establishing it takes a mismatched-cache audit, which we release.

Long-context large language models (LLMs) are increasingly deployed in real-world applications, yet self-attention remains a major efficiency bottleneck — especially during decoding — due to the necessity of repeatedly processing ever-growing key-value (KV) caches. Existing sparse attention reduce computation by attending to fewer KV pairs, but often suffer from substantial accuracy degradation, require additional training, or rely on expensive hashing. In this work, we present BinaryPC, a training-free, data-aware hashing-based sparse attention for long-context LLMs. BinaryPC constructs compact binary hash codes and corresponding hash function by computing binary principal components of data. Unlike Locality-Sensitive Hashing (LSH) with data-independent random projections or learned non-linear hashing methods, BinaryPC constructs binary codes that explicitly preserve the structural information of data without requiring gradient-based training. Comprehensive experiments across multiple model families and long-context benchmarks show that BinaryPC preserves accuracy relative to full attention while achieving superior performance among sparse and hashing-based baselines. On modern GPUs, BinaryPC improves end-to-end decoding throughput by 3.56$\times$ over the FlashAttention kernel. Our code is available at https://github.com/yudaohai666/BPC.

Multimodal large language models (MLLMs) achieve strong performance across diverse vision-language tasks, but their efficiency is limited by the cost of processing numerous visual tokens. Visual token pruning can reduce this cost, but requires accurate token importance estimates. Recent studies have demonstrated that text-to-vision attention from middle language model layers can effectively guide visual token pruning, typically using attention from a predefined middle layer to select the visual tokens to retain. Two problems therefore remain. First, our analysis shows that the layer whose attention is most responsive to the question varies substantially across samples, making a fixed layer suboptimal. Second, obtaining attention from the appropriate middle layer requires processing numerous visual tokens through several language model layers, by which point considerable computation has already been spent. To address both problems, we propose Middle-layer Attention Prediction (MAP), which uses Question Contrastive Teacher Selection to identify a sample-specific teacher layer by contrasting attention under the original and reference questions, and distills attention from the selected layer into a lightweight predictor that estimates visual token importance from multi-modal input features. During inference, MAP combines the predicted importance scores with a diversity criterion to prune visual tokens before the first language model layer. Thus, MAP requires no attention maps for pruning and remains compatible with existing inference acceleration techniques. Across ten benchmarks on LLaVA-NeXT-7B, MAP retains 97.5% of the unpruned model performance with only 5.56% of the visual tokens, yielding a 3.09x end-to-end speedup.

Edge platforms used for aerial observation must interpret aircraft imagery under limited memory, limited compute, and intermittent connectivity. This setting is difficult for standard RGB-only recognition models and general-purpose vision-language models, especially when calibrated thermal and LiDAR aircraft data are unavailable. We present TriCLE, an application-oriented tri-modal vision-language system for aircraft taxonomic grouping under edge constraints. From a single RGB aircraft image, TriCLE generates a structure-preserving FLIR-style thermal view and a pseudo-LiDAR depth projection, then fuses the aligned views with task instructions in a compact Qwen3-VL backbone. The model is aligned to an expert aircraft taxonomy based on propulsion, airframe family, size, design era, and configuration, so its outputs reflect engineering-relevant similarity rather than only surface appearance. We evaluate supervised fine-tuning, rotation-preserving SFT, and three policy-alignment strategies: GRPO, GSPO, and DAPO. Sequence-level GSPO gives the strongest validation performance, reaching 88.33% validation accuracy and 0.91 weighted F1 on valid aircraft outputs. On a held-out aircraft test partition, GSPO achieves 78.00% accuracy and 0.793 weighted F1 while preserving 94.00% parseable output formatting. After 4-bit quantization and attention-memory optimization, the aligned 4B model fits an 8GB deployment target and processes each tri-modal triplet in 1.48 seconds. These results support TriCLE as a practical prototype for interpretable, edge-feasible aircraft grouping, while emphasizing the need for further validation on real aligned thermal and LiDAR sensor streams.

Retrieval-augmented generation (RAG), which augments large language model (LLM) generation with information retrieved from databases, has become a widely used approach for knowledge-intensive applications. Modern RAG systems, however, expose many configuration choices, such as retrieval indexes, model selections, and how models invoke retrieval. Each configuration yields a different trade-off between answer quality and serving performance, making it challenging to choose the optimal setting for a specific application deployment. We present RAG-Stack, a framework for efficiently discovering quality-performance Pareto frontiers across diverse RAG applications and serving systems. RAG-Stack consists of RAG-PE, an iterative design-space exploration algorithm that selects the next RAG configuration to evaluate; RAG-IR, a workload abstraction for diverse RAG algorithms; and RAG-CM, a performance model that predicts the optimal deployment and serving performance on the given hardware. Together, these components allow RAG-Stack to search the joint algorithm-system configuration space without deploying every candidate and to transfer an existing Pareto frontier to a new serving system. Given the same number of optimization iterations across diverse datasets, the Pareto frontiers found by RAG-Stack cover 52.5% to 153.2% more of the normalized quality-performance space than those found by state-of-the-art configuration-search methods evaluated over the same RAG design space.

Deep neural networks (DNNs) deployed on resource-constrained neuromorphic hardware face three concurrent challenges: the need for model compression through pruning, vulnerability to adversarial input perturbations, and susceptibility to hardware-induced weight faults such as stuck-at-zero errors. While each of these factors has been studied in isolation, their combined effects on model reliability have received little attention. This paper presents an empirical investigation of how pruning, adversarial training, and hardware fault injection interact to affect the robustness of convolutional neural networks. Using a compact three-layer CNN trained on MNIST, we conduct three experiments: (1) comparing the fault tolerance of naturally and adversarially trained models under simultaneous hardware faults and adversarial attacks, (2) evaluating how pruning affects adversarial robustness, and (3) characterizing the joint accuracy surface across fault rates, adversarial perturbation magnitudes, and pruning levels. Our results show that adversarial training improves robustness against input perturbations but increases sensitivity to stuck-at-zero weight faults. Contrary to intuition, pruning did not significantly increase fault sensitivity, and varying the pruning level had little effect across fault rates and attack strengths. These results highlight the need to jointly consider adversarial robustness and hardware reliability.

Heterogeneous DRAM-based processing-in-memory (PIM)-GPU systems promise significant efficiency gains for decode-phase large language model (LLM) inference, particularly in long-output generation, yet current design practices overlook critical factors that determine real-world performance. Through systematic evaluation of diverse architectures and workloads (OPT-7B/70B, Mamba2-2.7B/70B), we reveal three fundamental design principles: (i) static power consumption (DRAM leakage, refresh, and GPU idle power) can dominate the efficiency calculus, causing dynamic-only models to overestimate tokens/s/W by up to 3.85X for realistic deployments (Mamba2-2.7B, batch size 1, 128 input tokens, and 2,048 output tokens); (ii) decoding performance is monotonically non-decreasing with channel count across all evaluated models and workloads, generally plateauing at high channel counts for low-batch workloads; under a fixed-capacity sweep, all models instead share a common near-optimal hierarchy configuration, with substantially larger misconfiguration penalties for attention-based models; (iii) workload mapping strategies provide bounded improvements (up to 14.0%/17.4% kernel-level latency/energy reduction, up to 5.6% end-to-end gain) and are not primary bottlenecks. Significant efficiency gains require system-wide co-optimization. These principles provide design-space guidance for architects designing the next generation of memory-accelerated LLM systems.

Serving large language models (LLMs) under diverse deployment constraints requires flexible trade-offs between accuracy, memory footprint, and throughput. However, conventional quantization methods typically require a separate checkpoint for each target bit-width. We introduce Recurrent Residual Quantization (RRQ), a post-training quantization (PTQ) framework that represents weights as a low-bit quantized base together with a sequence of quantized residual corrections, enabling multiple effective precisions from a single checkpoint. Starting from a 2-bit model obtained via post-training quantization (PTQ) or round-to-nearest (RTN), RRQ progressively adds lightweight 2-bit residuals generated via RTN to construct 4-, 6-, and 8-bit representations. The method is calibration-free and avoids joint multi-bit optimization. In our Qwen3-8B setup, the full all-RTN 2-/4-/6-/8-bit package is constructed in 1,293 seconds, 3.3 times faster than the measured MatGPTQ construction. Experiments on six recent LLMs show competitive accuracy at 6 and 8 bits, with model-dependent behavior at 4 bits. The code will be made publicly available upon publication.

Long-context LLM decoding reads the key-value (KV) cache at every step. Loading it takes longer than computing attention over it, so throughput is bandwidth-bound. Hence, reducing the cache size can raise both decoding speed and serving capacity. The challenge is to reduce cache size while preserving the attention products, keeping reconstruction cheap, and using a fixed per-token bit count. At two bits per element, the most competitive methods rely on orthogonal transforms. However, existing techniques are either data-oblivious or use the query statistics without deriving the transform from a distortion criterion. Moreover, they rely on transforms built on top of random or Hadamard rotations, which equalize variances across entries rather than compacting energy, and fixed-width scalar quantizers, which are suboptimal at low rates. In this paper, we formulate KV cache quantization as a transform coding problem in which distortion is the error in the attention products. We derive closed-form optimal transforms for keys and values from calibration statistics, under a high-resolution model. We show that the optimal key transform is not orthogonal and satisfies a generalized Parseval relation: the attention-aware distortion becomes mean-squared error (MSE) in the transform domain. Thus, we can use MSE-optimal vector quantizers applied directly to the transformed key coefficients. To meet the fixed-width layout requirement, we show that grouping coefficients into equal-volume partitions makes equal-size codebooks attain the variable-rate optimum under the same high-resolution model. At two bits per element, our method, termed NOVA-KV, recovers most of the long-context retrieval accuracy lost by scalar quantization methods at comparable throughput.

Model Predictive Control (MPC) delivers constraint-aware control, but its reliance on online optimization limits its use on systems with fast dynamics, high-dimensional models, or long horizons. Existing GPU implementations typically treat the device as a linear-algebra accelerator, leaving the optimization loop dependent on repeated kernel launches and high-latency memory transfers. This paper introduces CUDA MPC, a GPU-native MPC framework that co-designs the optimization algorithm, execution model, and memory architecture for CUDA hardware. CUDA MPC pairs a parallel-in-horizon alternating direction method of multipliers (ADMM) splitting with a fused CUDA kernel that runs the entire iterative solve on the device. Intermediate optimization variables stay in low-latency, on-chip shared memory, and a localized atomic-flag protocol synchronizes only adjacent horizon blocks, minimizing host intervention, kernel-dispatch overhead, and global-memory traffic. Across six nonlinear robotics benchmarks spanning increasing state dimension and constraint density, CUDA MPC sustains real-time rates at horizons one to two orders of magnitude longer than CPU solvers: it solves an optimization-based collision-avoidance parking problem with 100 s of lookahead within a 0.1 s sampling interval, and is the only solver evaluated that achieves both real-time execution and collision-free coordination for a centralized 10-agent swarm, where acados and CasADi return no feasible solution and require 3.5 s and 4.5 s per solve. Against tensor-framework implementations of the same ADMM splitting, the fused kernel is up to $965\times$ faster.

Double descent is commonly studied by scaling an explicit capacity parameter, such as neural-network width. For gradient boosting decision trees (GBDTs), however, an analogous single-axis capacity parameter has not been established. We propose the number of split candidates as an operational capacity parameter for GBDTs. Holding other training controls fixed, increasing the split-candidate budget refines the feature-quantization grid and expands the dictionary of root-to-leaf paths from which boosting selects its updates. To analyze this expansion, we construct an empirical tree-kernel diagnostic that summarizes how candidate-induced paths group the training examples. A regime in which the empirical kernel rank grows toward the sample size and very small positive eigenvalues emerge exposes noise-sensitive directions; in this regime, test error peaks before decreasing again at larger split-candidate budgets. This perspective predicts that deeper trees should reach the regime with fewer split candidates, larger training sets should require finer grids, and label noise should make the peak more pronounced. Experiments support these predictions and show test-error peaks at intermediate split-candidate budgets across XGBoost, LightGBM, and CatBoost, whereas a random-forest control improves monotonically under the same split-candidate sweep. Taken together, our analysis and experiments support split-candidate scaling as a single-axis capacity intervention for studying GBDTs and suggest that the observed double descent arises from an interaction between candidate-induced geometry and boosting dynamics.

Generative retrieval (GR) is a promising paradigm for industrial search advertising, yet its deployment is constrained by strict relevance and latency requirements. Existing systems cascade GR with an independent relevance model, decoupling the generative likelihood objective from query-ad relevance discrimination, which compromises effectiveness and increases serving costs. We propose a Unified Generative-Discriminative framework (UniGD) that integrates retrieval and relevance scoring within a single model. To mitigate gradient interference in joint optimization, UniGD introduces Conflict-Aware Gradient Enhancement (CAGE) to adaptively coordinate the two objectives. UniGD further designs a Codebook-Anchored Representation Module (CAM) that anchors item representations to frozen hierarchical codebooks distilled from a multimodal pretrained model, thereby endowing them with rich and generalizable semantic priors. For heterogeneous short-video, product, and live-stream ads, UniGD proposes Heterogeneous Ad-material Modeling (HAM), which captures cross-type semantic commonality over a shared backbone while preserving type-specific modeling capacity. Online AB tests on Kuaishou search advertising platform show that UniGD raises ad revenue by 5.78%, reduces inference latency by 33%, and improves discriminative relevance estimation. On NQ320K and MS300K, UniGD improves Recall@10 over the strongest reproduced GR baseline by 8.44% and 3.19%, respectively.

Zero-shot visual anomaly detection has achieved remarkable progress, with recent vision-only approaches further improving performance while simplifying the inference pipeline. However, existing methods typically perform dense computation over all images and spatial tokens, despite the fact that normal samples dominate real-world scenarios and anomalies usually occupy only small regions. Token pruning offers a promising solution, but introduces an asymmetric pruning risk in anomaly detection: retaining normal tokens mainly incurs redundant computation, whereas removing anomalous tokens may eliminate the only evidence for detection and localization. This risk is particularly severe in early layers, where pruning provides the greatest computational benefit but anomaly semantics remain unreliable. We propose KeepAD, a defect-preserving token pruning framework that formulates token selection as high-recall, anomaly-aware routing. In shallow layers, KeepAD combines coverage-preserving selection over local $2\times2$ patch neighborhoods with deterministic anomaly rescue to reduce the risk of discarding subtle defects. In deeper layers, frozen normal and abnormal prototypes guide pruning under an image-adaptive token budget, aggressively removing low-risk normal tokens while preserving local anomaly evidence. Dense-to-sparse self-distillation further supervises early token routing without introducing additional inference overhead. Experiments on six industrial and seven medical zero-shot anomaly detection benchmarks show that KeepAD reduces the token retention ratio to below $20%$, while limiting the average degradation in image-level and pixel-level AUROC to within $2.7$ percentage points. At the most aggressive operating point, KeepAD achieves a $7.9\times$ speedup over the strongest CLIP-based baseline.

Large Multimodal Models (LMMs) have achieved remarkable success on images and short videos, yet scaling them to long videos remains challenging due to frame-centric tokenization and limited context windows. 3D geometry provides a natural compression mechanism for visual streams: depth and camera pose enable observations from multiple views and time steps to be fused into a persistent, world-aligned representation. While recent 3D LMMs leverage geometry-aware representations to improve spatial reasoning, they continue to lag behind specialist 3D perception systems on grounding and segmentation tasks. We argue that a key limitation is geometry-aware decoding: existing methods communicate 3D predictions through language tokens, proposal selection, or lightweight grounding queries, creating a bottleneck between language reasoning and dense geometric prediction. Building on these insights, we introduce Qwen-3D, a geometry-aware LMM that compresses visual information within the Qwen backbone using multi-view geometric cues, enabling efficient long-horizon visual reasoning over static scenes. Qwen-3D augments visual tokens with 3D Rotary Positional Embeddings, allowing attention to operate directly in 3D scene space rather than across independent image frames and thereby facilitating scalable cross-view and temporal reasoning. To bridge language and geometry, Qwen-3D incorporates a query-based segmentation decoder that grounds language directly in the underlying 3D scene representation, unifying referential grounding, instance segmentation, and visual question answering across both images and videos. Across a diverse set of benchmarks, Qwen-3D surpasses existing 3D LMMs and outperforms several large proprietary 2D models. Notably, Qwen-3D achieves these improvements while maintaining strong performance on standard 2D vision-language benchmarks by jointly training on 2D and 3D data.

LLMs are widely used to generate structured output like source code or JSON. Grammar-constrained decoding (GCD) can guarantee the syntactic validity of the generated output, by masking out tokens that violate rules specified by a context-free grammar. However, the online computational overhead of existing GCD methods, with latency typically scaling linearly with vocabulary size, limits the throughput of LLMs, especially for models with large vocabularies. To address this issue, we propose PSC, a novel grammar-constrained decoding method. By combining acceptance conditions of all vocabulary tokens into a single classifier of the parser stack during preprocessing, PSC can compute the complete vocabulary mask by checking the parser stack exactly once per decoding step, with time complexity independent of the vocabulary size. Experiments show that PSC computes masks up to 700$\times$ faster than baselines on complex programming language grammars, and up to 30$\times$ faster for schema-conformant JSON; end-to-end LLM throughput with PSC approaches that of unconstrained decoding. We analyze the preprocessing overhead for preprocessing providers and decoding users, and provide a break-even point analysis to help users decide whether to do preprocessing by themselves.

While Vision-Language Models (VLMs) have demonstrated remarkable performance in processing and understanding both text and images, their large parameter sizes lead to significant computational overhead, limiting their deployment on resource-constrained devices. While pruning has been effective for compressing Large Language Models (LLMs), directly applying it to VLMs leads to significant performance drops, largely due to redundant visual tokens interfering with importance estimation. To this end, we propose SlimVLM, a structured pruning framework designed to compress VLMs while preserving their task performance. We introduce an adaptive visual token selection strategy for VLMs that leverages average text-to-visual attention scores to assess the importance of visual tokens, removing redundant ones during pruning based on a set threshold, thereby optimizing the importance calculation. Recognizing the varying tolerance to sparsity across different modules, we also propose a Sensitivity-aware dynamic pruning mechanism that determines the appropriate pruning ratio for each module by calculating the linear reconstruction error between the outputs of the pruned and unpruned modules, ensuring overall performance stability. Experimental results show that SlimVLM outperforms existing methods across multiple multimodal benchmarks, achieving state-of-the-art performance.

This paper proposes FedCritic-MIMO, a communication-efficient serverless federated multi-agent reinforcement learning framework for AI-native resource control across independently deployable cell-level controllers in open and disaggregated 6G RANs. Controllers share no trainer, retain local actors and personalized critic components, and exchange only compatible shared critic parameters. FedCritic-MIMO targets reuse-$1$ multi-cell massive-MIMO OFDMA deployments, where RAN controllers jointly manage user scheduling, per-stream power allocation, beamforming, interference, and long-term QoS with limited inter-controller signaling. Each base station locally executes its actor without centralized training or actor federation, while critic knowledge is exchanged peer-to-peer over an interference-aware graph. It enables this collaboration through wireless-aware event triggering, adaptive layer-wise top-$k$ sparse critic exchange with error feedback, and balanced interference-aware fusion. We establish conditional finite-time stationarity and consensus guarantees for the balanced, compressed peer-to-peer critic recursion under a fixed-policy, frozen-target critic-regression model. In strongly interference-coupled reuse-$1$ simulations, FedCritic-MIMO achieves the best performance-communication tradeoff among heuristic, independent-learning, centralized-training, and communication-ablation baselines. It achieves the highest held-out throughput, improves user-rate distribution and mean SINR, increases QoS satisfaction, and attains the lowest interference cost per delivered bit among learning baselines. It reduces critic-communication overhead by $76%$ relative to uncompressed distributed critic exchange. These results demonstrate that serverless exchange of compatible shared critic parameters can coordinate RAN controllers without centralized trajectory collection or parameter-server aggregation.

Query-based mask transformers assemble segmentation outputs through pixel-wise competition among query predictions of the final layer, yet this inference process is not explicitly optimized during training. We identify two key mismatches: the query with the highest probability-mask score does not necessarily produce the most accurate mask, and final-layer decoding may discard superior predictions from intermediate layers. To address these issues, we propose Inference-Aware Learning (iFAN), a general training framework for plain mask transformers. iFAN introduces Adjusted Probability-Mask Ranking (APMR), which aligns query competition with predicted mask quality and suppresses high-confidence but inaccurate competitors. We further employ Cross-Layer Self-Distillation (CLSD) to transfer stronger intermediate predictions to the final layer. The ranking and distillation objectives are training-only, while inference retains efficient final-layer decoding. Experiments on COCO, ADE20K, and Cityscapes demonstrate consistent improvements across panoptic, instance, and semantic segmentation, as well as across different architectures, backbone scales, and input resolutions. Overall, iFAN improves performance by an average of 1.20 PQ, 1.30 AP, and 0.63 mIoU, with negligible additional parameters, FLOPs and inference latency.

Large language models have transformed artificial intelligence from isolated prediction services into components of long-running, distributed systems that reason, invoke tools, retrieve external state, delegate tasks, and act on behalf of users and organizations. The surrounding ecosystem has responded with agent frameworks, workflow engines, model-serving platforms, memory systems, communication protocols, and observability tools. These technologies improve execution, but they do not provide a stable, implementation-independent operating architecture for governing intent, selecting capabilities, preserving authority across delegation, controlling uncertainty, coordinating runtime behavior, and reconstructing why consequential actions occurred. This paper proposes the Agent Operating System (AOS), a vendor-neutral reference operating architecture for distributed agentic systems. AOS contains two internal planes: a Control & Governance Plane responsible for intent, policy, trust, authority, confidence, auditability, observability, and human oversight; and a Runtime & Coordination Plane responsible for agent lifecycle, workflow coordination, model and tool routing, context and memory coordination, scheduling, traffic management, and runtime assurance. Platform services, Linux or Windows, container runtimes, and physical infrastructure remain outside the AOS boundary and are integrated through explicit interfaces. The paper specifies AOS concepts, invariants, interface objects, optimization objectives, deployment profiles, and reliability responsibilities. It also identifies tradeoffs and unresolved research questions. AOS is not presented as a replacement for existing frameworks or infrastructure; it is proposed as the operating architecture through which heterogeneous components can be composed into governable, reliable, observable, and interoperable agentic systems.

Omni-modal large language models (Omni-LLMs) have achieved remarkable performance on audio-visual understanding tasks, but processing long and highly redundant visual and audio token sequences incurs substantial computational overhead, demanding aggressive token compression for efficient deployment. Existing methods often degrade at low token budgets: pre-LLM compression may discard structurally important and globally distributed evidence, whereas inner-LLM compression often underexploits query-conditioned audio-visual collaboration. To address these limitations, we propose OmniPack, a training-free framework that coordinates structural compression before the LLM with task-relevant semantic refinement within the LLM. Before the LLM, OmniPack removes structural redundancy through modality-specific importance, global coverage, and similarity-aware merging. After sufficient multimodal interaction, it further consolidates diverse, task-relevant representations through textual guidance and audio-visual collaboration. Extensive experiments on five benchmarks with three Omni-LLM backbones demonstrate that OmniPack consistently achieves the best performance-efficiency trade-off across diverse retention ratios, outperforming all existing methods. Notably, on Qwen2.5-Omni-7B, OmniPack preserves 98.0% of the original performance while reducing FLOPs to 16.7%, and still retains 92.9% of the original performance with only 6.8% of the original FLOPs.

Training-free speculative decoding drafts by matching an exact suffix of the context against a pool of earlier context. That lookup misses correct drafts already in the pool, most visibly on tool-calling traffic, where a request repeats almost everything but the few values minted for it, and where one rejected token discards the correct continuation behind it. We diagnose the failure position by position across ten benchmarks and find it to be a problem of addressing rather than of coverage: on our densest tool-calling benchmark, about half of what the strongest exact-match drafter misses is present in the pool yet unreachable by exact matching. We therefore propose a second, semantic draft source: the same pool, re-keyed by the hidden state the verifier has already computed at each committed token, together with a merge that lets it ride inside an existing lexical drafter's tree. In three published drafters, at matched pool and budget, it lifts accepted length by 24-29%. Oilbird reaches 4.4x autoregressive decoding speed on API-Bank, against 3.9x for the strongest training-free baseline in our harness and 2.0x for EAGLE-3.