生成模型与LLM推理优化

Most adaptive-inference techniques for foundation models change what the model does - early exit, MoE routing, KV-cache compression, dynamic attention sparsity. The input that hits the backbone, however, remains a fixed-grid tokenisation indifferent to image content. We argue that this is a missed lever. We present SpiralFovea, a parameter-free, input-adaptive tokeniser in which token identity, location, scale, and count are all functions of local visual entropy and selection completes before any backbone parameter is queried. Around content-driven hotspot anchors, multi-scale spiral rings produce <= 78 patches that replace the standard 196-patch ViT grid at the input stage. Across four canonical fine-grained benchmarks, SpiralFovea yields +1.7-2.1 pp accuracy with a 60% reduction in input tokens, an 84% reduction in self-attention FLOPs at every transformer layer, and 18-29% throughput gains over the matched static tokenisation baseline. A controlled ablation on CUB-200-2011 Genus across four backbones reveals a clean diagnostic: the gain magnitude tracks inversely with the strength of the backbone's whole-image positional prior, isolating self-supervised foundation models as the regime where input-adaptive tokenisation is most valuable.

The deployment of Large Language Models (LLMs) with extended context windows is increasingly constrained by the linear growth of Key-Value (KV) cache memory. Vector Quantization (VQ), particularly Residual Quantization (RQ), is a promising approach for pushing KV cache storage toward the sub-1-bit regime by progressively encoding residuals with small codebooks. However, most VQ methods still rely on standard $\ell_2$ $K$-means as the core codebook-learning primitive. We identify a subtle high-dimensional issue of this primitive: Euclidean centroid averaging can induce centroid shrinkage, which weakens the angular alignment term in the $\ell_2$ distortion and makes directional preservation harder. To address this issue, we propose Gain-Shape $K$-means (GSKM), a drop-in replacement for $K$-means that improves directional fidelity while matching, and in some regimes improving, $\ell_2$ distortion. We then build Gain-Shape Residual Quantization (GSRQ) by incorporating a weighted extension of GSKM into an RQ pipeline. On LLaMA-3-8B, GSRQ substantially improves over strong KV cache quantization baselines across bit rates. At 1-bit, it improves the average accuracy across LongBench tasks from 11.34 to 33.54, a gain of 22.20 percentage points over VQLLM.

In prefill-decode (PD) disaggregated LLM serving, each request is assigned to a decode worker after prefill. Existing decode routers balance only load; for mixture-of-experts (MoE) models this is incomplete: equally loaded workers can differ in latency, since each decode step loads the weights of every distinct expert its batch activates. We present ELDR, an expert-locality-aware decode router for PD-disaggregated MoE serving. From a request's prefill expert activations, ELDR builds an expert signature predicting the experts it will activate during generation. Offline, balanced K-means partitions signature space across decode workers; online, locality-band routing sends each request to the least-loaded worker among those best matching its signature. A signature cache, co-indexed with the KV cache at KV-block granularity, keeps signatures exact under prefix caching. Implemented in vLLM and evaluated on deployments of up to 40 GPUs, ELDR reduces median TPOT by 5.9-13.9% over the strongest of four load-balancing baselines across three MoE models and two workloads, with model outputs unchanged.

Long-context LLM services now sustain prompts with hundreds of thousands to millions of tokens, making the key-value (KV) cache a first-order serving cost. Because the cache grows linearly with context length, it can exhaust GPU memory, force smaller batches, and reduce serving throughput. Prior KV cache compression techniques typically target only the sequence dimension or only the channel dimension, which leaves limited headroom as context windows scale. Compressing both dimensions promises higher memory reduction, but applying the two forms of compression directly leads to significant accuracy loss. This paper introduces MosaicKV, a dynamic two-D (dimensional) KV cache compression system for extremely long-context serving. MosaicKV uses dynamic two-D compression to address the accuracy challenge, exploiting the non-uniform importance distribution of elements within the KV cache. Instead of applying one compression pattern globally, MosaicKV identifies important elements for each KV vector and selects compression strategies at the granularity of KV cache segments. To address the performance challenge, where fine-grained sparsity and compression management overhead can offset the gains from compression, MosaicKV introduces compressed KV cache management. This mechanism uses underutilized GPU and CPU resources to maintain compressed KV caches and accelerate attention computation. Evaluation on an H800 GPU with multiple LLMs shows that MosaicKV delivers up to 16x attention speedup, 4.8x lower decode latency, and 7.3x higher throughput than the uncompressed baseline. At the same time, it reduces memory usage by 3x and incurs only 1.76% average accuracy loss on LongBench and RULER.

Mixed-precision quantization (MPQ) has become a key technique for deploying large language models under stringent memory and compute constraints. We first identify a phenomenon that we term the Perplexity Illusion: layers ranked as important by perplexity-based sensitivity show little rank correlation with those that are most influential for complex reasoning performance, with Kendall $τ\approx 0$ in our analysis. We further reveal an Alignment-Diversity Tradeoff: using only target-task calibration data can degrade post-quantization performance, whereas incorporating general-domain data stabilizes sensitivity estimation and improves robustness across tasks. Based on these observations, we propose TASA (Task-Aware Sensitivity Analysis), a two-level framework that jointly optimizes calibration-data composition and mixed-precision bit allocation. Specifically, TASA searches for a calibration-data mixture using a training-free gradient-trace alignment criterion, and then aggregates perplexity and reasoning-oriented sensitivity signals to guide both inter-layer and intra-layer bit allocation. Experiments on LLaMA-3-8B and Qwen2.5-7B reveal a precision inversion: appropriately allocated 3.5-bit models can match or surpass less task-aware 4-bit baselines. At an average precision of 3.5 bits, TASA matches or outperforms several competitive 4-bit uniform baselines in aggregate accuracy, and improves over the strongest W3 baseline on GSM8K by more than 20 absolute points on LLaMA-3-8B. These results show that calibration-data composition substantially affects task-sensitive quantization, a factor underexplored in prior work.

Object-centric models inspired by DETR have become the dominant paradigm for open-vocabulary video instance segmentation (OV-VIS). While recent efforts have reduced the computational cost of pixel decoding, textual modality fusion, and object decoding to make these architectures more suitable for mobile devices, real-time on-device inference at high frame rates remains an open challenge. In this paper, we introduce SegFS, a dual-stream fast-slow framework that significantly improves efficiency without sacrificing accuracy. On sparse keyframes, an open-vocabulary object-based model predicts instance-level representations. These representations are then projected back into the backbone feature space to condition a lightweight fast network, which efficiently relocalizes and segments the instances in subsequent frames. By shifting instance propagation from object decoding to feature-space conditioning, our approach decouples multimodal semantic understanding from dense mask prediction and enables efficient temporal propagation. The proposed fast branch achieves up to 14x lower latency than the mobile-oriented MOBIUS model, while maintaining competitive segmentation performance on standard OV-VIS benchmarks.

Federated Learning (FL) is a distributed machine learning (ML) paradigm with collaboration among multiple clients without sharing data. FL is challenging under data heterogeneity and partial client participation. Learning sparse models is useful for communication and computational efficiency in FL, but it is especially difficult in the small-sample high-dimensional regime (d >> N) where optimization can yield parameter configurations that fail to generalize to unseen test data. While magnitude-based pruning doesn't account for uncertainty exploration in the parameter space, a formulation with probabilistic gates and an L0 constraint allows sampling from competing sparse configurations during training. In this work, we study entropy regularization of gate distributions as a mechanism to maintain uncertainty in sparse federated optimization by preventing early commitment to sparse support. We examine its impact under data heterogeneity, client participation heterogeneity, and sparsity. Experiments on synthetic and real-world benchmarks show consistent improvements over federated iterative hard thresholding (Fed-IHT) and pruning after dense federated averaging (FedAvg) training, both in statistical performance on test data and in sparsity recovery accuracy.

LLM-based agents execute multi-turn workflows with continuously growing contexts, where LLM calls are interleaved with tool invocations and environment feedback. To maintain model quality, modern agent frameworks rely on context engineering strategies such as offloading, reduction, and isolation to control the context length. However, these strategies introduce significant context transformation overhead: each transformation invalidates existing KV caches and triggers re-prefill, leading to increased time-to-first-token (TTFT). In this paper, we identify that context transformations are segment-decomposable, where the transformation of a prefix is independent of future tokens. This property enables transformations to be executed ahead of time. Based on this insight, we propose a lookahead programming model that allows agent frameworks to express context transformations as asynchronous operations without modifying their execution logic. The runtime proactively executes these transformations and prepares transformed KV caches in advance, enabling direct context replacement without blocking. We further design a lookahead-aware scheduler in LLM serving systems to support these asynchronous requests alongside latency-critical workloads with controlled interference. We implement our approach to support representative context engineering strategies and integrate it into existing agent frameworks and LLM serving systems. Experiments show that our approach effectively eliminates transformation overhead and reduces TTFT by up to 11.9x.

We present LuckyStar 111B, a 111B-parameter hybrid reasoning model developed through a collaboration between Cohere and LG CNS for Korean-English enterprise agents under practical memory and serving constraints. The model trains from Cohere's fully post-trained Command A model rather than a new pretraining run, and uses preamble conditioning to switch between concise non-reasoning behavior and longer tool-oriented reasoning. We study four choices for scaling tool-using agents efficiently: multilingual supervised fine-tuning, reinforcement learning with verifiable rewards for multi-step tool-use tasks, language-consistency rewards for Korean user-facing responses, and 4-bit quantization for single-GPU serving. The adapted model improves mathematical reasoning, function calling, and agentic natural-language-to-SQL (NL2SQL) performance while preserving general Korean and English instruction-following quality. These results provide a practical recipe and failure-mode analysis for adapting post-trained multilingual models to verifiable agentic workflows under memory-constrained deployment.

Semantic navigation is a fundamental task for embodied agents operating in unseen environments, requiring both semantic understanding and long-term decision-making. Recent foundation models have empowered agents with rich semantic priors for this task. However, without structured global representations, decision-making often falls back on local observations and greedy strategies, resulting in inefficient exploration and myopic behaviors, especially in long-distance navigation. To address these challenges, we propose a zero-shot semantic navigation framework. Our method incrementally maintains an online Hierarchical 3D Scene Graph (HSG) to form a multi-granular semantic topology over objects, zones, and regions, serving as a compact state abstraction for global planning. Building on this memory, we introduce a hierarchical belief-based planning framework that fuses semantic priors with exploration evidence on the HSG, and performs finite-horizon rollouts on an HSG-based simulator to explicitly estimate the long-term expected returns of candidate macro-actions. This enables globally consistent decisions and reduces redundant backtracking. Extensive experiments in high-fidelity simulation environments across multiple tasks and datasets demonstrate that our method outperforms existing state-of-the-art methods, particularly in long-distance scenarios, where our approach improves SR and SPL by an average of 9.4% and 5.0%, respectively.

VLA models have emerged as a powerful paradigm for transferring semantic knowledge from web-scale data to physical robotic control. However, current single-frame architectures suffer from intrinsic limitations: temporal myopia that discards historical dynamics, reasoning gaps between high-level instructions and low-level motor commands, and inference inefficiency due to autoregressive scalar decoding. In this work, we propose MIRTH, a unified framework designed to address these challenges. MIRTH augments a pretrained VLA backbone with three key innovations: (1) dual-scale temporal memory hubs that compress long-term scene evolution and short-term motion trends into compact embeddings; (2) latent reasoning tokens optimized via a mutual-information objective carving out a semantic plan space to align multimodal context with action trajectories; and (3) a parallel action decoding scheme that replaces autoregressive generation with vector-wise prediction to maximize control throughput. Extensive evaluations on the LIBERO simulation benchmark and a real-world LeRobot platform demonstrate that MIRTH achieves state-of-the-art performance and exhibiting emergent error recovery capabilities. The codes and collected datasets are released at http://github.com/kiva12138/mirth.

Quantifying the marginal impact of individual optimization passes underpins phase ordering, pass selection, optimization design, and analysis of pass/hardware interactions. In LLVM — the standard backend for C/C++, Rust, and ML stacks via MLIR — interactions among optimization passes, measurement noise, and pipeline scale make this difficult. We present a systematic, empirical study of the LLVM -O3 optimization pipeline. We decompose the pipeline into cumulative per-pass prefixes. We then measure execution time, compile time, binary size, hardware counters, and RAPL energy across 84,750 measurements covering 113 cumulative prefixes of the -O3 pipeline evaluated on 30 PolyBench/C kernels under rigorous noise mitigation. On these compute-bound affine kernels, the pipeline is non-monotone (6.6-9.7% of transitions regress) and strongly back-loaded (the median non-regressing kernel needs 84.8% of the pipeline for 80% of its speedup). Most gains are driven by a small Pareto-dominant core of passes, while the final -O3 configuration is Pareto-dominated on (size, speedup) for 29 of 30 kernels. We further show that IR instruction count is an unreliable predictor of runtime, that runtime-targeted passes are de facto energy-targeted (30-60% savings), and that the search-free idealized-additive upper bound on losses due to phase interference is 46.35%. These findings enable more informed pass pruning, cost-model calibration, and autotuning.

Processing-in-Memory (PIM) is a proven paradigm for overcoming the ``memory wall". However, while data indexing is severely bottlenecked by this same wall, it remains unclear how indexing can effectively benefit from PIM's unique capabilities. We present PATH, an in-situ indexing architecture that bridges this gap by leveraging the massive parallelism and inherent data-movement of PIMs. Specifically, we first reformulate the fundamental indexing operations, namely Insert, Search, Update, and Delete, into highly parallel in-situ content-addressable memory operations executed directly within memory arrays. Taking hash indexes as a typical case, we elaborate how PATH breaks the inherent trade-off among memory accesses, load factor, and process latency in conventional hashing schemes. By adopting ultra-large logical buckets and in-memory moving, PATH virtually eliminates the cost of hash collision resolution and significantly reduces resizing overhead. Compared with state-of-the-art schemes, PATH achieves $4.7-7.8\times$ higher throughput, $>14.5\times$ lower tail latency, and $>61.4%$ fewer memory accesses under insertions, laying a scalable foundation for next-generation data-centric computing.

Vision-language models (VLMs) combining reinforcement learning (RL) ignite remarkable progress in multimodal reasoning, yet still struggle with medical images, which typically exhibit extremely sparse visual evidence to inform clinical decision-making. We recognize that pruning visual tokens outside the grounding region greatly enhances medical reasoning. However, a united RL framework for active visual token pruning (VTP) and medical multimodal reasoning remains unestablished. Here, we propose a dual-stream RL framework, ViToS, to fulfill token pruning and question answering. ViToS trains one policy model with two task branches, where one focuses on grounding while the other conducts token-sparse reasoning after VTP. Furthermore, we solve the coupled policy learning problem by introducing the cross-feedback sequential optimization, avoiding gradient conflict and facilitating convergence of the shared policy model. Evaluated on seven medical benchmarks, our method reduces visual tokens to 77% of the original sequence length while achieving a 108.27% relative performance on Lingshu-7B and 104.16% relative performance on HuatuoGPT-Vision-7B. Overall, ViToS delivers superior performance and inference speedup, establishing an efficient paradigm for medical multimodal reasoning.

Deploying Vision Transformer (ViT) models on edge platforms remains challenging due to their high computational demands and the architectural heterogeneity of modern hybrid ViT models, which incorporate both fully connected and convolutional layers. This heterogeneity leads to significant variation in tensor shapes, requiring flexible and efficient FPGA-based acceleration. In this paper, we present FlexViT, a reconfigurable FPGA accelerator for efficient ViT inference on resource-constrained edge devices. Built on the SECDA-TFLite framework, FlexViT employs a hardware-software co-design approach that maps both fully connected and convolutional layers onto a unified high-throughput INT8 GEMM engine using a runtime im2col transformation. To efficiently support diverse layer configurations, we propose a dual-mode dataflow that dynamically switches between input and weight reuse by reconfiguring the compute array at runtime. We further introduce a depth-first tiling strategy that completes accumulation in a single pass, eliminating off-chip partial-sum transfers and reducing memory bandwidth requirements. We implement FlexViT on a PYNQ-Z2 FPGA and evaluate it across a representative set of ViT models. FlexViT achieves up to 2.74x speedup on accelerator-executed layers, translating into up to 1.40x end-to-end speedup compared to CPU-only execution. The code is available at: https://github.com/gicLAB/FlexViT

Radar sensors provide reliable perception under adverse weather and lighting conditions, but their sparse, noisy, and weakly semantic measurements make dense semantic segmentation challenging. Most existing radar segmentation methods rely on grid-based encodings and pairwise interactions, which struggle to capture the higher-order relational structure formed by multiple radar returns from the same physical object. We introduce a unified higher-order structural alignment framework for multi-view radar segmentation. The proposed method refines radar feature representations using learnable hypergraphs to capture higher-order dependencies among spatially related responses. To ensure consistency across heterogeneous radar projections, we further align view-specific features using Unbalanced Optimal Transport (UOT), enabling correspondence-free alignment under varying measurement densities and partial observations. An adaptive attention mechanism then fuses complementary radar views while emphasising structurally informative responses under sparsity and noise. The resulting architecture learns structurally consistent representations across Range Angle (RA), Range Doppler (RD), and Angle Doppler (AD) views and is trained using supervised segmentation together with cross-view consistency regularisation. Experiments on the CARRADA and RADIal benchmarks demonstrate consistent improvements over strong radar-specific baselines, achieving 63.8% mIoU on CARRADA and 83.4% mIoU on RADIal, improving the previous best methods by +1.7 and +2.3 mIoU, respectively. These results highlight the importance of higher-order relational modelling for robust radar perception.

Performance in face and speaker verification is largely driven by margin-penalty softmax losses such as CosFace and ArcFace. Recently introduced $α$-divergence loss functions offer a compelling alternative, particularly due to their ability to induce sparse solutions (when $α>1$). However, standard geometric margins are designed for the softmax function and do not naturally extend to this generalized probabilistic framework. In this paper we propose Q-Margin, a novel $α$-divergence loss that introduces a principled probabilistic margin. Unlike conventional methods that apply geometric penalties to the logits (unnormalized log-likelihoods), Q-Margin encodes the margin penalty directly into the reference measure (prior probabilities). This formulation naturally encourages discriminative embeddings while preserving the beneficial sparsity properties of the $α$-divergence. We demonstrate that Q-Margin achieves competitive or superior performance on the challenging IJB-B and IJB-C face verification benchmarks and similarly strong results in speaker verification on VoxCeleb. Crucially, against ArcFace and CosFace baselines trained under an identical recipe, Q-Margin consistently improves at low False Acceptance Rates (FARs), a capability critical for practical high-security applications. Finally, the extreme sparsity of the Q-Margin posteriors enables exact and memory-efficient training, offering a scalable solution for datasets with millions of identities.

Most token reduction methods for Vision Transformers seek favorable tradeoffs between accuracy and efficiency by pruning, merging, or pooling patch tokens. REDI (Relevance for DINOv3 Token Reduction) studies this question through a controlled supervised reference: how should a fixed token budget be allocated across patches for image classification? REDI quantizes final block DINOv3 patch representations into a visual vocabulary and derives class conditioned corpus scores using supervised TF-IDF over visual words. For each validation image, the ground truth class selects a row of the TF-IDF table, and four transformed views produce a TF-IDF map aligned to a reference center crop. A separate dense pass on the same crop provides an attention map. After independent min max normalization, their elementwise product defines the REDI score. A fixed keep, merge, and compress operator then uses score rank to assign patch roles and score magnitude to weight merging and compression. With precomputed REDI scores, a frozen DINOv3 ViT-B/16 backbone, and the same linear classifier used for dense evaluation, the operator reduces the sequence length from 201 to 107 tokens, a 46.8% sequence reduction. The REDI variant based on incoming attention mass achieves 84.706% Top-1 accuracy on ImageNet-1K, compared with 83.514% for the dense baseline, 82.634% for incoming attention mass alone, and 81.796% for supervised TF-IDF alone. The same corpus term also improves reduced classification for three alternative attention formulations relative to their attention only counterparts. Together, these controlled comparisons indicate that class specific corpus statistics and image specific attention provide complementary signals for patch ranking in this setting.

Real-world detectors for autonomous driving, surveillance, and robotics must handle domain-shifts under strict latency and memory constraints, yet existing source-free object detection (SFOD) methods rely on heavyweight architectures that prioritize accuracy alone. We show this trade-off is unnecessary: building on YOLOv10, an NMS-free dual-head detector, we achieve state-of-the-art adaptation accuracy while being faster and more compact. We observe that directly applying vanilla mean-teacher self-training to dual-head detectors leads to suboptimal adaptation performance due to two key factors. First, simple pseudo-label generation strategies, such as using a single head or directly combining high-confidence predictions from both heads, yield suboptimal supervision under domain-shift. We propose DHF (Dual-Head Pseudo-Label Fusion) which selectively admits one-to-one (O2O) and one-to-many (O2M) head predictions, preserving precision and recovering missed objects. Second, we observe domain-shift collapses multi-scale feature discriminability. We propose the use of our MARD (Multi-scale Adaptive Representation Diversification) loss which mitigates this by enforcing detection-aware variance and covariance constraints on multi-scale feature maps. Both modules are training-time only, leaving inference unchanged. Across domain-shift benchmarks, our method, RT-SFOD yields 1.4 to 3.5% mAP gains, 1.3$\times$ higher throughput, with $\sim$2$\times$ fewer parameters than prior state-of-the-art SFOD methods, thus advancing the Pareto frontier of the speed-accuracy-model size trade-off. We report main results with YOLOv10, and demonstrate generalizability with additional YOLO- and DETR-based dual-head detectors. Code is available here: https://github.com/Sairam13001/RT-SFOD/

Accurate prediction of stress and strain fields in hierarchical composite microstructures is critical for physics-informed material design, yet conventional finite element method (FEM) simulations are computationally prohibitive at scale, requiring minutes to days per evaluation. In this work, we propose a hybrid UNet-Transformer architecture that predicts complex mechanical field distributions directly from composite microstructure geometry images, serving as an efficient surrogate for FEM across ten distinct stress and strain field types spanning diverse two-phase composite configurations including square, hexagonal, and triangular tessellations, multiple boundary conditions, and high-resolution geometries. Results demonstrate that the proposed architecture achieves strong predictive performance across the majority of subdatasets, with peak accuracy on periodic tessellation geometries reaching R2=0.9991, SSIM=0.9936, and MAE=0.0050 on the boundary condition subdataset and the triangular tessellation subdataset respectively. Across six of the eight evaluated subdatasets, MAE remains below 0.05 on the normalized [0,1] pixel scale. Encoder attention analysis via Grad-CAM and Grad-CAM++ confirms that the model develops physically meaningful internal representations, localizing attention at mechanically critical regions including phase boundaries, ligament junctions, and indenter contact zones without explicit structural supervision. Performance degrades on irregular square-grid geometries with sparse soft-phase inclusions, with the S11 normal stress subdataset yielding R2=0.7735 and SSIM=0.7126, consistent with the known limitation of smooth-loss image translation models in reproducing sharp stress discontinuities.

Multimodal Large Language Models (MLLMs) incur prohibitive inference costs due to long visual token sequences. Training-free visual token reduction provides an efficient solution. However, existing methods distort attention distributions, giving rise to a phenomenon we term Attention Logit Collapse. To address this issue, we propose ERA, an Entropy-guided visual token pruning framework with Rectified Attention for efficient MLLMs. Specifically, ERA comprises three crucial components: Dual-view Entropy Pruning (DEP), Bias-aware Token Recycling (BTR), and Logit-preserving Attention Rectification (LAR). First, DEP identifies representative anchor tokens by jointly modeling visual diversity and head-wise saliency. BTR then recycles pruned tokens into their corresponding anchors while estimating a cluster-level logit bias. Building upon this, LAR injects the estimated bias into attention logits, effectively rectifying the collapse induced by token reduction. Together, these components preserve visual evidence even under aggressive compression, enabling robust performance across single-image, multi-image, and video settings on a wide range of MLLMs. Beyond delivering practical acceleration, ERA establishes logit-preserving visual token pruning as a principled framework for efficient MLLMs, unifying theoretical foundation, algorithmic design, and practical deployment. The code is at https://github.com/924973292/ERA.

Large reasoning models (LRMs) such as DeepSeek-R1 have achieved strong performance through extended chain-of-thought (CoT) generation. However, deploying them on edge devices raises a conflict between long CoT sequences and constrained resources. Recent confidence-based early exit methods reduce CoT length for individual requests, yet they apply fixed thresholds from a single-request perspective, ignoring multi-request concurrency and load fluctuation in edge serving. To bridge this gap, we propose \underline{L}oad-\underline{A}ware \underline{S}erving with \underline{E}arly-exit for \underline{R}easoning (LASER). LASER couples two complementary designs: (1) a load-aware adaptive exit threshold that adjusts the confidence bar based on real-time system load within an empirically validated robust range, and (2) a difficulty- and load-aware reasoning budget pre-allocation that assigns compute resources by request difficulty and system capacity. We formulate the problem as a joint optimization of reasoning quality and service latency. Experiments on two reasoning models, four benchmarks, and diverse load conditions show that LASER reduces average latency by 17--38% and improves service-level objective (SLO) satisfaction by 3--6% over fixed-threshold baselines, at an average accuracy cost of only 1%.

Speculative decoding accelerates inference by using a lightweight draft model to generate candidate tokens in parallel, and are then verified by the target model, enabling lossless acceleration. Recently, diffusion-based speculative decoding further improves parallelism by generating multiple tokens per forward pass via block-level diffusion, achieving state-of-the-art (SOTA) performance. However, existing methods adopt a fixed inference block size and assume a uniform optimal decoding strategy across all inputs. In this paper, we show that this assumption is suboptimal, as the optimal block size varies across samples and plays a critical role in speculative decoding performance. Moreover, these values exhibit a clear local structure, concentrating around the training block size, which reduces the problem to a low-dimensional and structured decision space. Based on these insights, we propose BlockPilot, a sample-adaptive policy that predicts the optimal block size from the prefilling representation. Specifically, we formulate block size selection as a lightweight policy learning problem and propose an instance-adaptive decision mechanism that predicts the optimal block size based on the representation of the prefilling stage. The prediction is performed only once after prefilling, allowing for seamless integration. Extensive experiments demonstrate that our method is plug-and-play, introduces minimal overhead, and consistently improves efficiency, achieving an acceptance length of 5.92 and a 4.20$\times$ speedup on Qwen3-4B under temperature $T=1$.

Modern Vision-Language-Action (VLA) planners for autonomous driving emit a chain-of-causation (CoC) reasoning step before producing a trajectory. The reasoning is autoregressive and dominates inference latency, while the trajectory head is parallel and cheap. Latency is an operational constraint in autonomous driving, so accelerating the reasoning step is the central problem we address. We observe that CoC reasoning has two qualitatively different needs: most tokens continue routine setup that follows naturally from the ego-trajectory history, and a small fraction encode commitments that require fresh visual evidence about an unexpected situation. We split this reasoning into two specialized paths: a routine reasoner that handles the predictable continuation by attending to trajectory history, and a deliberative reasoner (the unmodified VLA target) that handles novel cases by attending to current visual evidence, using the speculative decoding framework as the architectural template for how the two paths cooperate. Unlike standard speculative decoding, our routine reasoner is not a smaller replica of the target; the two reasoners are deliberately specialized to read different parts of the prompt. We propose two techniques to realize this. First, we introduce FlatRoPE, a 1D rotary positional embedding in the draft that breaks the rotational symmetry of the target's 3D M-RoPE, redirecting attention away from visual tokens and onto trajectory-history tokens. Second, we introduce Action-aware RL (AARL), a post-training stage that uses an action-quality reward together with a static-reference KL anchor. Together, our two-reasoner system reduces the reasoning-step running time by approximately $4\times$ relative to the original Alpamayo planner.

Large language models deployed in regulated industries operate under two constraints: compliance enforcement and cost efficiency. Personally identifiable information (PII) in user queries can reach model endpoints before the system determines whether that data should leave its jurisdictional boundary. Serving all queries through a single large model consumes full GPU capacity regardless of query complexity while offering no mechanism for geographic routing. Mixture-of-Experts architectures do not address this routing occurs between expert layers within the model after data has already arrived at the endpoint, with all experts loaded in memory regardless of query complexity. We propose a classifier-gated routing architecture that enforces compliance by design. A trained encoder classifier sits before any decoder inference, evaluating each query for complexity and data sensitivity, then routing it to an appropriately sized dense model in the appropriate geographic location. PII-containing queries route to local endpoints before any LLM computation begins, making data residency violations structurally impossible. Simple queries reach small, fast models at a fraction of the cost. Our evaluation on 600 queries demonstrates 39% median latency reduction, 33-52% cost savings depending on query distribution, and generation throughput of 122-200 tokens/second versus 50-64 for the baseline. The encoder classifier achieves 99.2% accuracy with near-perfect PII recall at 7ms inference overhead, establishing pre-inference classification as a practical path to compliance-by-design LLM deployment.

Large language models increasingly operate over long contexts, where the KV cache becomes a dominant memory bottleneck: its size grows linearly with sequence length and must be retained throughout decoding, making full GPU caching prohibitively expensive without compression. Existing KV cache compression methods struggle to balance efficiency with faithful context preservation. Token eviction discards information, while semantic grouping fixes compression decisions at prefill time; neither can recover token-level detail from a compressed span once it becomes relevant during generation. As a solution, we propose SeKV, a resolution-adaptive semantic KV cache that organizes context into entropy-guided semantic spans and stores them across a GPU-CPU memory hierarchy without discarding information. Each span keeps a lightweight summary vector on GPU for coarse routing and a low-rank SVD basis on CPU for on-demand token-level reconstruction. A trained zoom-in mechanism selectively expands query-relevant spans during decoding, enabling precise retrieval without materializing the full KV cache on GPU. SeKV enables adaptive token-level reconstruction while keeping the base LLM fully frozen and adding fewer than 0.05% trainable parameters. Across four benchmarks, SeKV improves over the strongest semantic compression baseline by 5.9% on average while reducing GPU memory by 53.3% versus full KV caching at 128K context. Code is available on https://github.com/AmirAbaskohi/SeKV.

As large language model (LLM) inference evolves from text-only to multimodal paradigms, inference systems face three challenges: (1) flexible orchestration of multimodal workflows, where heterogeneous computing units exhibit complex dependencies and concurrent control; (2) efficient transmission of massive intermediate data across processes and nodes, with tensors flowing at high speed among heterogeneous roles; and (3) efficient sharing of KV caches and model weights across roles to eliminate redundant GPU memory. Existing solutions deploy LLMs and diffusion models independently, lacking a system-level abstraction for multimodal pipelines; this scatters orchestration logic, tightly couples transmission paths to specific models, and incurs high cost to integrate new models. To address these challenges, we present Omni-Flow, a distributed scheduling framework for multimodal inference through a three-layer abstraction. The Control Flow layer defines workflows via a Python DSL, orchestrating heterogeneous units into a unified dataflow graph that supports static DAGs and dynamic routing, with built-in service discovery and diverse load-balancing strategies. The Data Flow layer provides a distributed KV cache abstraction beyond prefill/decode separation, unifying allocation and enabling direct cross-role transmission across a three-tier paged storage hierarchy (GPU/CPU/SSD) over zero-copy, low-latency channels. The Compute Flow layer supports complex multimodal prefix matching for KV reuse across multi-turn dialogues, and takes over KV cache and sampling logic via a unified SGLang interface, letting diffusion models directly reuse the LLM forward path under unified parallel semantics. We demonstrate that Omni-Flow supports diverse heterogeneous scenarios with a consistent programming model, including omni-modal dialogue (LongCat-Next) and complex image generation pipelines (HunyuanImage-3).

Long-context Large Language Model inference is severely bottlenecked by the massive Key-Value (KV) cache, yet existing sparse attention methods often suffer from static fixed-budget (Top-k) retrieval or rely on proxy scores that are computationally expensive and biased. To address these limitations, we propose RaBitQCache, a novel sparse attention framework that utilizes randomized rotated binary quantization and high-throughput binary-INT4 arithmetic to efficiently estimate attention weights. Our proxy score serves as an unbiased estimator with a proven error bound, enabling adaptive Top-p retrieval that dynamically adjusts the token budget based on actual attention sparsity. We further implement a hardware-aware system with asynchronous pipelining and lazy updates to mask overhead. Evaluations demonstrate that RaBitQCache significantly accelerates inference and reduces memory I/O while preserving generation quality compared to state-of-the-art baselines. Code is available at https://github.com/Sakuraaa0/RaBitQCache.git.

Spiking neural networks (SNNs) have emerged as a promising paradigm for energy-efficient brain-inspired computing. However, existing online unsupervised SNN learning suffers from low training accuracy and poor scalability. Although current online supervised learning algorithms perform well on large-scale datasets and networks, the non-hardware-friendly operations hinder efficient edge deployment. In this work, we propose SpikON, the first algorithm-hardware co-design framework for efficient and scalable end-to-end online supervised SNN learning. We first propose the learnable threshold through time and scaled weight centralization through time techniques to address the inefficiency of traditional algorithms. Moreover, to reduce latency and energy consumption, we introduce the novel training dataflow and cascade computation reuse scheme for SNNs that allows concurrent forward-backward computation and temporal reuse across timesteps. We further design the dedicated SNN accelerator with a dual-parallel engine and customized SIMD-based SNN core for efficient end-to-end online learning. Experiments show that the SpikON algorithm achieves 32.2% and 35.0% reductions in training latency and energy consumption over the baseline, without sacrificing accuracy. Moreover, the SpikON co-design achieves 7.2x (11.5x) and 26.8x (15.8x) training throughput (energy efficiency) compared with the edge Apple M4 GPU and TPU-like accelerator, respectively. The code is available at https://github.com/peilin-chen/SpikON.

Micro-ultrasound ($μ$US) is a new, emerging, and promising imaging modality for prostate cancer (PCa) detection, but accurate identification of suspicious tissue remains highly dependent on clinical experience, leading to substantial inter-observer variability. Machine-learning assistance can reduce this variability; however, training reliable deep models is challenging because supervision is sparse and noisy — typically limited to core-level histopathology outcomes (e.g., cancer grade and its percentage in a biopsy core) without pixel-level lesion annotations and under severe class imbalance. We introduce Prost-RL, which reframes $μ$US PCa detection as a spatially aware, policy-driven inference problem by learning where to look before decoding. Prost-RL integrates a lightweight reinforcement-learning policy into a foundation-model encoder-decoder to generate interpretable spatial attention maps that act as soft prompts for both cancer-likelihood heatmap prediction and image-level classification. We further propose Adaptive Policy Optimization (APO) to stabilize hybrid supervised-RL training and a noise-robust objective combining symmetric cross-entropy with negative-entropy regularization to mitigate weak-label noise and encourage sharp localization. On a cohort of 6,607 biopsy cores from 693 patients across five clinical sites, Prost-RL achieves $79.0\pm3.5$ AUROC with $64.6\pm6.3$% sensitivity at 80% specificity for core-level detection (+2.1 AUROC and +4.5 sensitivity points over the strongest baseline), and $79.3\pm5.8$ AUROC for clinically significant cancer classification. The learned policy highlights biopsy-aligned regions, providing transparent, spatially grounded evidence alongside quantitative risk predictions. Code is available at: https://github.com/DeepRCL/Prost-RL.