生成模型与LLM推理优化

Retrieval-Augmented Generation (RAG) repeatedly prefills identical text chunks across queries, incurring redundant computations. Position-Independent Caching (PIC) mitigates it by reusing precomputed Key-Value (KV) across positions, but its efficiency is constrained by the large volume of text tokens. Rendering text chunks as images can compress the text into fewer visual tokens, but the rendered-image PIC suffers more severe quality degradation than the text PIC. This representation-specific gap primarily arises from contextual mismatches across independently compiled caches and the loss of fine-grained textual evidence during visual compression. Existing PIC repair methods mainly address the former through selective recomputation, but they incur online computation and cannot recover lost textual details. We propose QV-PIC, a query-aware dual-resolution PIC reuse framework guided by model-native templates. Offline, QV-PIC compiles visual caches under the model's native chat-template prefix, improving PIC quality without online recomputation. Online, it preserves global context with low resolution and restores fine-grained textual evidence within a high-resolution budget by cumulative query relevance scores, retaining the efficiency benefit of visual compression. Across six tasks, QV-PIC improves average F1 by 21.6 points over vanilla rendered-image PIC, closes the gap to vanilla text PIC, and surpasses optimized text PIC by 2.58 F1 while reducing TTFT by 17.2%. Relative to full prefill, it cuts TTFT by 83.8%.

Use this plain-text version for the arXiv abstract field: Learned image compression (LIC) models achieve strong rate-distortion performance but are hindered by high computational complexity and encoding-decoding mismatches across heterogeneous hardware platforms. Uniform fixed-precision quantization alleviates these issues but suffers severe quality degradation at low bit widths because it ignores differences in the quantization sensitivities of individual layers. To enable efficient and accurate low-bit deployment of pretrained LIC models, we propose HAMP-LIC, a Hessian-aware mixed-precision post-training quantization (PTQ) framework with a four-stage optimization strategy. First, block-wise sensitivity is estimated from the Hessian trace to capture second-order importance. Second, a task-aware refinement module adjusts these sensitivities by jointly considering quantization distortion and rate-distortion performance. Third, guided by the refined sensitivity profile, bit widths are allocated under a global model-size constraint to balance efficiency and reconstruction quality. Finally, block-wise reconstruction using a small calibration set further suppresses quantization error. Experiments on representative LIC models, including Minnen2018 and Cheng2020, demonstrate that HAMP-LIC achieves up to 4.85x model compression with as little as 0.59% BD-rate loss. It consistently outperforms existing fixed- and mixed-precision PTQ methods across multiple datasets while completely eliminating cross-platform encoding-decoding errors.

NVIDIA's published specifications give the Blackwell Ultra GPU (B300) a dense-compute ratio of roughly 30:1 between FP8 and INT8 tensor-core throughput; its predecessors, H200 and B200, both provide 1:1. We audit what this deprioritization means in practice by tracing INT8 W8A8 support through four layers of the stack: the published specifications, the PTX ISA, NVIDIA's CUTLASS kernel library, and the two major open-source LLM serving engines (vLLM and SGLang). We find a consistent, layered withdrawal: (i) the PTX ISA never exposes the fifth-generation tensor-core integer path (tcgen05.mma with .kind::i8) on sm_103a, even though the same PTX revision extends the FP4 kinds to that target, leaving legacy warp-level IMMA as the only architecturally legal integer tensor-core path on B300; (ii) CUTLASS's kernel generator explicitly skips INT8 UMMA generation for any build targeting 103a, while generating FP8 unconditionally; (iii) vLLM ships no INT8 GEMM for Blackwell and fails with a hard runtime error at the first forward pass, after the model has loaded; and (iv) SGLang's ahead-of-time INT8 GEMM stops at Sm90, while its FP8 tuning configurations already cover B200. We document an escape hatch (rerouting vLLM's INT8 path to a JIT-compiled Triton backend via an environment variable), a false-negative trap in the obvious profiler methodology for detecting "native INT8" on sm_103, and the practical failure semantics that make naive testing expensive. Together, these findings show that a quantization format's availability is a property of the whole stack rather than of the model or the spec sheet. Four distinct layers, three of them NVIDIA's own, withdrew INT8 support in mutually consistent ways, and a format that is nominally present on the datasheet is, by default, undeployable on this hardware.

Diffusion Large Language Models (dLLMs) have emerged as a competitive alternative to autoregressive language models, offering the potential for substantially faster inference through parallel decoding. Existing parallel decoding schedulers typically commit positions only after they meet a per-position criterion, overlooking how early commitments may benefit subsequent decoding. We identify a ripple effect in dLLM decoding: proactively committing a mid-entropy pivot position can induce a pronounced reduction in uncertainty across the remaining masked positions. This uncertainty reduction allows subsequent steps to unmask more tokens in parallel, thereby accelerating the overall decoding process. To exploit the ripple effect, we propose Ripple-Pivot Search (RPS), a novel training-free decoding method that seeks mid-entropy positions as promising candidate pivots (where to decode), and determines their token assignment that yields the greatest downstream benefit via lookahead evaluation (what to decode). Across 3 dLLMs and 4 reasoning and code-generation benchmarks, RPS achieves 4-10$\times$ wall-clock speedup over the standard decoder while preserving generation quality, and improves accuracy over the previous lookahead baseline by up to 5.49% while delivering higher throughput in most settings. When integrated with KV caching, RPS further achieves up to 18$\times$ wall-clock speedup over the standard decoder.

Tokenizer vocabulary size is a foundational design choice in large language model (LLM) infrastructure, yet it is typically fixed at training time based on convention rather than deployment analysis. We show that the cost-optimal vocabulary is not a constant but a function of the serving regime. We formalize total deployment cost as $C_{lifecycle}(V) = C_{train}(V) + λ\cdot C_{infer}(V, B)$, where $λ$ is inference volume and $B$ is the serving batch size. Through controlled experiments on two GPU families spanning the memory-bound to compute-bound regimes (A10G, ridge $\approx$ 117 FLOP/byte; A100, ridge $\approx$ 183 FLOP/byte), we demonstrate: (1) the inference-optimal vocabulary shifts 16x with serving batch, from 32k at $B=1$ to 524k at $B=64+$, driven by amortization of the $V \times d$ unembedding matrix read; (2) at 1.3-2.3B model scale, quality (bits per byte, BPB) is optimized at $V=65$k, confirming scale-dependent vocabulary preference; (3) the lifecycle-optimal vocabulary diverges from training-optimal by up to 16x for production deployments. Quality is approximately invariant across the optimal range ($<$2% BPB spread), making vocabulary a pure systems optimization with no quality penalty in the measured range. Our results provide actionable capacity planning guidance: on-device deployments ($B=1$) should use $V \approx 32$k; datacenter serving ($B \geq 64$, $λ\geq 10$) should use $V \approx 131$-262k.

Despite the practical relevance of sparse decision-based black-box threats, they have received limited attention in semantic segmentation. To bridge this gap, we adapt the most representative decision-based black-box sparse attacks from the classification domain to serve as baselines, establishing a rigorous benchmark for this underexplored setting. In this context, we demonstrate that one of the existing methods suffers from severe query inefficiency due to its image-centric pixel accumulation, which rapidly exhausts query budgets across the vast image space. To overcome this, we propose SegPAR, a novel decision-based framework that shifts to a class-centric exploration paradigm. Furthermore, to eliminate the misleading feedback generated by standard decision rewards during pixel accumulation, we introduce a novel discrepancy reward. Extensive experiments show that SegPAR significantly outperforms black-box baselines in sparsity efficiency and MIoU reduction, while remaining competitive with white-box sparse attacks. Code is available at \href{https://github.com/KAU-QuantumAILab/SegPAR}{https://github.com/KAU-QuantumAILab/SegPAR}.

Branch mispredictions and data-cache misses are usually evaluated as separate bottlenecks: studies report perfect-branch or perfect-cache speedups as isolated upper bounds and often treat their product as the joint ceiling. In irregular workloads, however, hard-to-predict branches and cache-missing loads often occur in the same hot loops. Removing one penalty can expose the other: faster memory reaches mispredicted branches sooner, while better branch prediction leaves more long-latency loads in the out-of-order window. We call this interaction symbiotic stall latency (SSL). This paper quantifies when isolated ceilings fail using joint speedup synergy (JSS), the observed joint perfect-branch/perfect-cache speedup divided by the product of the isolated speedups. Values above one mean independent-ceiling analysis understates attainable gain. Across 53 simulated workloads, 70% show measurable coupling (JSS > 1), though many are near unity, especially in lower-pressure cases. With a conservative threshold, 40% exceed the independence product by more than 6%, and kernels with SSO > 20 show JSS from 1.23 to 3.29. We introduce symbiotic stall opportunity (SSO), a lightweight MPKI-based screen for workloads that merit full joint simulation. We map high-SSO workloads to four recurring software patterns: neighbor access, hash lookup, linked-structure traversal, and data-dependent modification. We connect SSL to reorder-buffer occupancy, squash rate, and commit starvation under isolated perfect modes. The resulting methodology is simple: use SSO to screen, JSS to validate, and report conditional branch-after-cache and cache-after-branch gains when evaluating branch predictors, prefetchers, caches, or coupled branch/memory mechanisms. Our contribution is a measurement framework showing when isolated perfect modes are adequate and when they understate joint performance headroom.

Continuous physiological monitoring using consumer-grade wearables offers a transformative opportunity for clinical care and research, yet integration remains hindered by device heterogeneity, proprietary data formats, and strict regulatory requirements. We present an event-driven, cloud-native system designed to ingest, normalize, and analyze high-frequency vital signs from wearables at scale and without vendor lock-in. The system design proposes a multi-layered microservice architecture using cluster orchestration. Data acquisition is handled via a cross-platform mobile application that leverages native health frameworks, ensuring compatibility across fragmented device ecosystems. To address interoperability, we implement an event-driven transformation pipeline using stream processing engines and specialized services to map raw measurements to the FHIR standard for medical interoperability. Our novel dependency-aware FHIR minimization scheme reduces storage overhead while maintaining lossless resource reconstruction. Furthermore, the platform integrates a modular data analytics and machine learning layer based on a medallion lakehouse architecture, supporting the full machine learning lifecycle from real-time stream processing to model serving. Performance evaluation demonstrates that the ingestion pipeline sustains 50 full ingestion requests per second with median response times under 8 ms, satisfying the low-latency requirements for real-time patient monitoring. Our open-source implementation adheres to regulatory compliance standards through role-based access control and secure service-to-service communication, providing a robust foundation for deploying wearable-based monitoring in institutional healthcare settings for clinical decision support and research workloads.

Mixture-of-experts (MoE) models have recently moved beyond routing a fixed number of complete experts. Shared-expert designs preserve reusable knowledge, fine-grained methods vary computation within experts, and dynamic routers adapt the number of active experts. Yet these decisions are usually made independently, overlooking a basic dependency: extracting reusable computation changes both what remains and how much expert capacity the remainder needs. We study this dependency by decomposing sparsely upcycled feed-forward experts into key-value channels. Co-activated experts align at a subset of value positions; removing these positions changes expert preference; and greater shared coverage is associated with lower residual expert demand. These observations lead to one principle: share first, then route what remains. We instantiate it in UniF-MoE, a unified framework for token-adaptive MoE computation. Each expert is partitioned into aligned blocks. A shared-demand score sets the shared block count and pathway weight, key prototypes select the shared content, and the complementary demand determines the residual expert count through cumulative routing mass. A Gram regularizer separates and normalizes router embeddings, promoting diverse routing directions, sparse expert overlap, and a simple routing geometry. Experiments on DomainBed and GLUE show that this unified design improves predictive performance over representative static and dynamic MoEs while reducing activated computation, inference latency, and memory. Code is available at https://github.com/existence0420/UniF-MoE.

Accurate pre-deployment estimation of CNN inference cost--energy, latency, and peak memory--is increasingly critical as models are deployed on resource-constrained GPU platforms. Existing approaches rely on FLOPs, latency measurements, or single-device profiling as energy proxies, overlooking the non-linear interactions between architectural design and hardware load. We present a workload characterization study of 13 419 CNN configurations on two GPU platforms (RTX 5090 and RTX 3080) under GPU telemetry, revealing that energy, latency, and memory exhibit fundamentally distinct scaling behaviors: energy and latency diverge by 3x under high computational demand, and cross-GPU transferability differs by target--energy and latency require platform-specific models while memory transfers well across the two tested platforms. Building on these characterization findings, we develop CARB, a cascade-blended ensemble that jointly predicts all three targets with R2 ~0.99, and a two-stage deployment screening workflow that eliminates over 90% of candidates in seconds, reducing large design spaces to a Pareto-prioritized shortlist validated against real hardware.

Historical context integration presents a fundamental challenge for Vision-Language Models (VLMs) in sequential decision-making tasks. Current VLMs process visual inputs independently, which creates critical limitations for downstream applications that require temporal understanding. Direct incorporation of historical frames into Transformer inputs produces quadratic attention complexity and excessive memory consumption. Existing approaches suffer from significant drawbacks: computational inflation or substantial information loss through temporal compression. To address these challenges, we introduce Dynamic Context Adapter (DCA), a novel context injection approach for pretrained VLMs. Our method employs fixed-size, dynamically compressed memory to preserve historical semantics without frame concatenation. DCA bridges static VLMs and recurrent policies and enables memory capabilities in pretrained models while maintaining computational efficiency. DCA achieves over $25%$ reduction in attention FLOPs and $13%$ memory savings while improving performance on long-horizon tasks.

Successful mutation strategies in evolutionary code search may contain reusable knowledge that is useful beyond a single run, and in some cases may transfer across related tasks and domains. However, existing LLM-driven evolutionary frameworks largely discard such knowledge, repeatedly rediscovering similar ideas and limiting opportunities for cross-run and cross-task learning. We introduce EvoMem, a persistent memory architecture for LLM-based evolutionary program search that captures and reuses candidate mutation knowledge. EvoMem converts successful mutation events into structured, task-aware advice for future runs. It operates in two phases: after each run, it extracts and stores promising ideas with provenance, and during subsequent evolution, it retrieves a small set of relevant instructions based on the current task and program context to guide mutation. Across geometric optimization, multi-hop question answering, GPU kernel optimization, and related benchmarks, our experiments show positive average improvements in target metrics or search speed for most evaluated settings, while also revealing variability across tasks. Overall, EvoMem provides evidence that persistent memory can reduce some redundant exploration and improve the reuse and adaptation of successful strategies in LLM-driven evolutionary search.

The attention mechanism forms the foundation of many modern AI models such as the Transformer. In one subclass of problems where attention is used, inputs and outputs are bound to the probability simplex so that all outputs sum to one. In this setting, softmax attention admits an exact, component-by-component quantum realization. Attention scores are Hadamard-test statistics on block-encoded projections of amplitude-encoded inputs. The exponential softmax is the interior of a cosine-squared family generated by Born-rule measurement under an exact bijection, whose boundary expresses sparse attention with exact zeros at finite parameter values. The softmax temperature is a repetition count where post-selected measurement rounds realize discretized inverse temperature exactly. Value aggregation is a deterministic column-loading channel that dilates the column-stochastic value matrix. The gated residual is the preparation angle of a single ancilla, with the additive identity at a mixing angle of π/2. Every learnable parameter is a rotation-gate angle. The composed layer is exact in the infinite-shot limit with one measure-and-reload step per attention score; a fully-coherent variant is ε-approximate via quantum singular value transformation in the infinite depth limit. The algebraic core is machine-checked in Lean 4.

We introduce Amulet, a rendering method that transforms a scene into a sparse, tiled and layered intermediate scene representation (cache) for high-frequency frame extrapolation. In contrast to reprojection-based techniques, Amulet explicitly rasterizes and stores potentially visible geometry in its layered image-space cache, allowing accurate shading and inpainting of newly disoccluded regions without hallucination. Our key contribution is a cache that is predictively filled with shading information for future views, amortized over multiple current frames. Novel views are synthesized by hierarchically traversing the cache front to back and refining stale or missing shading on the fly. Using a predictive, gradient-based scheduler that assigns lifetimes for each tile, we enable adaptive shading updates under motion and dynamic lighting. Amulet decouples the rasterization and shading rate from the refresh rate of the display. In many scenarios, our cache can use a single shaded frame to synthesize multiple extrapolated frames with only a few localized updates. In a typical application, we extrapolate a 60 Hz shading rate to a 240 Hz display. Amulet achieves up to 250 Hz at 4K resolution and is competitive with state-of-the-art frame generation methods, including DLSS and neural-flow approaches, in multiple metrics. Amulet explores the design space of sparse layered image-space representation. It enables accurate, non-neural multi frame extrapolation with explicit handling of disocclusions. Our findings show that Amulet can extrapolate many more frames than contemporary methods with high quality, rivaling latency-bound frame interpolation methods with similar quality in many scenes.

Quantization-Aware Training (QAT) enables the deployment of quantized models with minimal accuracy degradation. However, in practical scenarios, training labels are often unavailable due to privacy, copyright, or cost constraints. Knowledge Distillation (KD) is a common approach to address this challenge, but we observe that prior work combining QAT with KD suffers from a fundamental limitation: during distillation, the range mismatch between the teacher and the quantized student model induces an unattainable residual, resulting in an irreducible lower bound on the distillation loss. Motivated by this observation, we propose SQuaT (Student-Aware Quantized Teacher Features), a label-free QAT framework with KD that theoretically eliminates this lower bound by applying the student's quantization parameters to quantize the teacher's features during distillation. Through comprehensive experiments across diverse settings, we demonstrate that SQuaT consistently outperforms strong baselines, with particularly pronounced gains in extreme low-bit (e.g., 1- and 2-bit) settings. Furthermore, extensive evaluations across various model design choices show that our approach does not rely on specific architectural assumptions, making it broadly applicable across diverse architectures and quantization settings. The source code is available at https://github.com/lcdbsa522/SQuaT.

Analog Computing-in-Memory (ACiM) accelerates deep neural networks by keeping weights inside memory arrays and executing dot products in the analog domain. However, modern ACiM accelerators are often limited by the "ADC wall": analog-to-digital converters consume a large fraction of energy and area, while bit-sliced execution repeatedly invokes these converters. Existing designs reduce this cost with low-resolution readout or time multiplexing, but they either lose output fidelity or introduce serialization overhead. Charge-CIM addresses this bottleneck by using switched-capacitor charge redistribution as a unified computing and conversion substrate. The same capacitor fabric performs input conversion, analog MAC, weighted shift-and-add, and readout quantization, reducing both standalone converter overhead and intermediate ADC invocations. A differential readout path further combines paired partial sums during ADC quantization, providing a highly compact and energy-efficient solution for array integration. With dataflow architecture support, we evaluated Charge-CIM on a suite of DNN benchmarks, from CNNs to Transformer models, and experimental results show that Charge-CIM reduces ADC energy by 91.7% under our evaluation setup and improves energy efficiency by 2.7x and throughput by 2.0x compared to the state-of-the-art charge-domain CIM accelerator.

Arabic maqam music microtonal, modal, and built on ornamented call and response is among the traditions most underserved by generative music models, whose training frameworks remain predominantly Western and equaltempered. Real time accompaniment sharpens this gap: an AI partner must listen, adapt dynamically, and respect idiomatic microtonal structures. Streaming text to music models provide strong generative capabilities but lack precise control interfaces. We present MazzikaAI, a knowledge based system that uses natural language as the actuator of a realtime control loop. By compiling live MIDI, gesture, and inferred harmony into continuously updated text prompts, MazzikaAI steers an unmodified streaming generator, Google Lyria RealTime, without requiring model finetuning. The system embeds expert knowledge of six core maqamat, characteristic ornaments, and ensemble dynamics, maintaining realtime responsiveness with subsecond keytoaudibleupdate latency. Empirical evaluations demonstrate that dynamic prompt compilation reliably grounds generation in microtonal scales, significantly increasing offgrid quartertone content over baseline generation. Beyond its core implementation, MazzikaAI illustrates how deterministic knowledgebased rules can effectively bridge expert, nonWestern musical traditions and unfinetuned foundation models. This architecture establishes a scalable paradigm for realtime humanAI cocreation, offering a generalizable blueprint for interactive accompaniment, adaptive music education, and culturally inclusive generative audio across diverse global idioms.

We present a theoretical foundation for inverse-distance attention, from its Euclidean prototype (Resolver) to its non-Euclidean realization (Riemann GeoResolver). The Euclidean part establishes three core theorems: (1) circuit separation---IDA achieves exact retrieval with $\mathcal{O}(1)$ resources while softmax requires $Ω((\log n)^2)$ width; (2) a Polyak--Lojasiewicz inequality with $Ω(e^{Δ^2/\sqrt{d}}/Δ^2)$ stronger constant than softmax, implying linear convergence, $\mathcal{O}(\log n)$ Lipschitz scaling under a low-rank/clustering assumption, $Θ(1)$ Hessian spread, and absence of spurious local minima; (3) a width-independent effective rank bound that limits noise memorization---softmax memorizes arbitrary labels when $d_h\ge n$, while IDA limits test error to $\mathcal{O}(η^2)$. The non-Euclidean extension then builds upon this prototype, replacing Euclidean distance with hyperbolic geodesic distance for storage and spherical geodesic distance for routing. The Riemann GeoResolver framework comprises ten integrated modules: four HIDA operators spanning $Θ(n^2)$ to $Θ(1)$ per token; Hyperbolic Curvature Compression (HCC) with provable error bounds; HyperGate with gradient lower-bound theorem; Spherical Inverse Distance Attention (SIDA) with sphere-analog PL inequalities; Dynamic Memory Genesis (DMG) with $\mathcal{O}(\log T)$ regret bounds; and Geodesic Sparse Routing (GSR) with quality and communication bounds. The Euclidean theorems are proved in full; the non-Euclidean extension theorems are proved with analogous arguments. This work establishes a theoretical arc: from Euclidean attention as a special case, to hyperbolic memory, to spherical retrieval.

Token-pruning policies are usually designed for a single recognition pipeline, but pretrained Vision Transformers are reused across tasks with different spatial demands. We ask which parts of a pruning policy transfer across image classification, semantic segmentation, and object detection. For each pipeline, controlled probes freeze the no-pruning checkpoint and apply a series of parameter-free reduction criteria at one eligible layer at a time without retraining. The probes reveal three differences: segmentation and detection rank the criteria differently, classification is especially sensitive to attention-based pruning in the earliest layers, and the dense tasks prefer opposite recovery endpoints. These findings motivate Task-Adaptive Pruning (TAP). Existing register tokens serve as task-agnostic storage for feature artifacts. TAP instead introduces one task register per task and activates only the current one. Its evolving state ranks tokens, distributes an exact removal budget over depth, and sets the recovery scale for dense features. At a final keep rate of $ρ=0.5$, our jointly adapted model, TAP-J, reaches $47.0$ mIoU at $1.30\times$ encoder throughput on ADE20K and $53.7$ box AP at $1.32\times$ encoder throughput on COCO while remaining competitive on ImageNet-1K.

Clustering is a fundamental class of data analysis techniques with the most important representatives being centroid-based methods like $k$-means. Such methods are strongly connected to quantization problems, which aim to approximate general probability measures with discrete ones. For example, $k$-means corresponds to quantization with respect to the Wasserstein distance. While Wasserstein quantization clusters points within a fixed space, this paper studies Gromov-Wasserstein (GW) quantization, which additionally aims at clustering the ambient geometry of the space. We show existence of solutions to the GW quantization problem and give a characterization that justifies an analogue to the $k$-means algorithm (Lloyd's algorithm) to approximate them numerically. We further calculate the quantization rate for usual Euclidean geometries that are used in the GW context, and relate it to standard Wasserstein quantization rates. Finally, numerical experiments show that GW quantization opens up many modeling possibilities beyond normal clustering methods (e.g., for geodesic distances of 3D shapes or structured pruning of neural networks) and that the introduced algorithm leads to useful numerical solutions with approximation quality often in line with theoretically optimal rates.

In fault-tolerant quantum computing, fast and accurate decoding is crucial. Snowflake is a decoder for the surface code that runs in a streaming fashion. In this paper, we implement Snowflake on commercial FPGAs and validate them at room and cryogenic temperatures. Our results demonstrate high decoding throughput for small code distances that, when extrapolated, remains within acceptable limits for larger distances. Further, we incorporate the calculation of certain decoder confidence scores with negligible overhead both in terms of latency and physical resource utilisation. We note that implementing a large-scale system would require either a large FPGA beyond today's technology or clusters of FPGAs connected via a high-speed bus. Thus, we discuss an alternative architecture that exploits the locality of Snowflake by processing 2D slices of the 3D decoding window and offloading segments of the 3D structure to a high-speed memory.

InfinityStar extends visual autoregressive generation to video through a sequence of image and clip pyramids. Its changing scale and cross-clip context, however, leave late-scale attention costly and make sparse patterns reused from diffusion or image VAR models unreliable. We introduce SparSTAR, a training-free block-sparse attention method tailored to this setting. At each expensive scale and attention head, SparSTAR scores contiguous key blocks from the current query and key activations, retains required conditioning context, and executes the selected blocks through a forward-only sparse path. We analyze cross-scale consistency within a clip, pattern persistence across clip boundaries, and quality degradation as reuse spans increasingly distant scales. Across these analyses, important key blocks shift, showing that recomputing block selection at each target scale is more reliable than reusing a transferred mask. On 720p text-to-video and image-to-video generation, SparSTAR preserves every token and refinement scale while providing about a 1.6x end-to-end speedup and maintaining VBench and paired-output reconstruction fidelity close to dense InfinityStar.

Speculative decoding accelerates autoregressive large language model (LLM) inference by using a lightweight draft model to speculate multiple tokens, reducing expensive target model decoding steps. Its effectiveness depends heavily on draft selection, motivating adaptive methods that exploit variation across inputs and generation stages. On memory-constrained edge devices, however, these methods often fail to improve end-to-end throughput due to the overhead of switching between draft models. We identify a key limitation in this setting: the mismatch between draft selection and draft availability under tight memory budgets. To address this challenge, we present MemSpec, a prediction-guided, memory-aware runtime for adaptive speculative decoding on edge devices. MemSpec decouples draft selection from execution through proactive resident working-set management. A lightweight predictor estimates draft effectiveness from prompt and generation context, while a memory-aware scheduler reduces reactive model loading overhead. Experiments on a Jetson Orin Nano show that MemSpec improves steady-state generation throughput by 40.7% on average over state-of-the-art bandit-based adaptive methods while closely approaching the oracle upper bound.

ReRound (Reconstructive Rounding) is a post-training quantization method that addresses the midpoint ambiguity inherent in standard round-to-nearest (RTN) schemes when quantizing weights near the centers of quantization intervals. Starting from a pretrained LLM, ReRound trains a conditional diffusion model to produce continuous reconstructions of low-bit weights for the LLM. These reconstructed weights act as a guidance signal to disambiguate the rounding direction of weights located close to interval midpoints. To integrate this reconstruction-guided rounding with conventional RTN, ReRound introduces a tolerance metric measuring how far the quantized weight (not the final quantized integer) is away from the midpoint: quantized weights within a tolerance region around midpoints are quantized using diffusion-based reconstructions, whereas weights closer to quantization boundaries are quantized with RTN. By sweeping the tolerance parameter, ReRound generates multiple candidate quantized integer weight matrices and selects the de-quantized weight matrix candidate whose leading singular values most closely match those of the original full-precision weights. This selected candidate determines the tolerance parameter ReRound uses. ReRound is particularly effective for smaller LLMs. Across a range of such models, it consistently outperforms standard RTN for 3-bit and 4-bit weight quantization. ReRound achieves superior accuracy compared to an extensive set of calibration-free methods, remains competitive with calibration-dependent approaches, and operates entirely offline, introducing no additional overhead during low-bit inference. The ReRound strategy represents a new approach for low-bit quantization. The method applies to AI models beyond LLMs. This paper focuses on its applications to small LLMs.

Edge LLMs must preserve inference continuity when a user hands over between edge nodes, requiring key-value (KV) cache transfer to the target node. However, simultaneous handovers saturate the backhaul, preventing full cache delivery within the mobility-imposed transfer window. Rather than allocating bandwidth as if all cache entries were equally valuable, we order each user's KV cache by importance and transmit only its most informative fraction, turning token-level sparsity into communication savings. We cast the transfer as a multi-user backhaul allocation problem that maximizes average accuracy across users. Each user's partial-cache accuracy serves as its utility: a sigmoid that fits measurements on the RULER benchmark with $R^2>0.99$ across models and context lengths. Because importance ordering front-loads the high-value entries, the concave region of the accuracy curve spans nearly the entire cache. Our proposed allocator keeps served users within this region, making each per-slot allocation problem convex. The optimum is derived via a closed-form weighted water-filling solution that generalizes information-theoretic water-filling and enables online scheduling. The proposed allocator attains over 93.7% average accuracy in a 500ms transfer window, within 0.5pp of the full-cache ceiling, and reaches 98.2-99.5% of a clairvoyant upper bound.

Reinforcement learning (RL) for large language models is moving toward multi-turn agentic workloads, where rollout tasks repeatedly pause for external environments, resume with growing contexts, and finish at highly variable times. In this setting, RL training goodput, measured by training throughput, matters more than raw GPU occupancy: GPU waiting and repeated prefill recomputation are pure overhead. We present TideRL, a readiness-aware elastic RL system with Continuous Task Batching, Resource-Aware Ref-Actor Pipelining, and Elastic Resource Scaling. CTB preserves useful rollout state, $\textrm{RA}^2\textrm{P}$ selects between decoupled streaming and colocated aggregation from the ready backlog and arrival interval, and ERS moves ranks between rollout and training using the same readiness signals. Across text-only and multi-modal agentic workloads, TideRL improves RL training goodput by up to 5.6$\times$ over synchronous baselines and over 33% over asynchronous baselines, while reaching similar task performance. It also improves KV cache hit rate by 1.58$\times$, reduces per-step training time by up to 44.3%, and cuts total waiting time by up to 77.6%.

Vision-Language-Action(VLA) models map camera images and language instructions directly to motor commands through a single autoregressive transformer. In real-time control, they still spend substantial compute recomputing key-value(KV) representations for visual tokens that barely change across neighboring frames. Recent work such as VLA-Cache reduces that cost by reusing KV states for visually static patches, but its policy relies only on observation-space heuristics and does not account for the model's own uncertainty. We propose Gated VLA-Cache, a lightweight, training-free extension that augments visual-similarity caching with neural introspection. The method monitors the logit margin between the top two predicted action tokens, a zero-cost confidence signal available during decoding. When the margin drops below a threshold, the cache is invalidated and a full recompute is triggered. Evaluated on four LIBERO benchmark suites with both OpenVLA and OpenVLA-OFT, Gated VLA-Cache improves reliability when blind caching hurts. On LIBERO-Goal and LIBERO-Long, it recovers over 100% of the lost accuracy while retaining 80% of the compute savings.

Modern reinforcement learning (RL) post-training pipelines for large language models (LLMs) increasingly combine rollout workloads across multiple domains and feedback paradigms. Prefix-aware routing improves inference efficiency through cache reuse and load balancing, but it does not control how heterogeneous rollout sessions compete for KV-cache capacity. When reinforcement learning with verifiable rewards (RLVR), reinforcement learning from human feedback (RLHF), and agentic rollouts share an asynchronous inference service, their distinct sequence structures, interaction patterns, and KV-residency times create substantially different serving demands. Rollout scheduling must account for this heterogeneity without distorting the workload mixture specified by the trainer. We present MISA-T, a routing-layer admission policy for mixed rollout serving. MISA-T combines adaptive session admission, workload-aware KV-capacity allocation, and residency-time-aware KV accounting. In rollout-only ablations on Step3.7 and Qwen3.6-35B-A3B, MISA-T improves rollout throughput over a sweep-tuned cache-aware vLLM Router by 53.3% and 43.6%, respectively, while maintaining high prefix-cache hit rates. In a matched 50-iteration Step3.7 experiment, it increases rollout throughput by 35.6% and reduces mean iteration time by 22.8%, while keeping the consumed workload mixture close to the trainer target and achieving comparable task scores.

Abundant visual information strengthens vision-language model (VLM) perception, yet massive visual tokens raise inference costs. Existing visual token pruning methods rely on similarity-based guidance, which exploits pairwise text-vision and vision-vision token correlations for compression. However, such methods only capture local layer-level signals and overlook the whole inference process in VLM. In this paper, we revisit VLM inference and present a new efficient guidance scheme that complements similarity-based guidance. In particular, we identify a key observation: as LLM layers deepen, text tokens continuously aggregate visual information via self-attention and progressively absorb partial visual content into textual representations. To quantify this phenomenon, we propose Cross Modal Absorption (CMA) from a geometric representation perspective to measure how much visual information is absorbed by text, revealing that more visual tokens in deeper layers can be approximately explained by the text subspace. We accordingly propose Cross Modal Residual (CMR). It projects visual tokens onto the text subspace via Tikhonov regularized least squares and exploits reconstruction residuals to quantify visual information that cannot be explained by text. Finally, based on CMR, we present SIEVE, a training-free visual token compression method that combines CMR, text-attention relevance, and residual-space diversity to retain task-relevant and complementary tokens. Experiments on diverse VLM architectures verify the effectiveness of SIEVE. For instance, on LLaVA-NeXT-7B, SIEVE keeps only $11.1%$ of visual tokens while preserving $97.5%$ of the original average performance, achieving $3.62\times$ prefill speedup, $2.49\times$ end-to-end speedup, and a $6.02\times$ KV-cache reduction.

Streaming video understanding requires multimodal large language models (MLLMs) to preserve relevant evidence from continuously evolving streams under strict causality and bounded memory. Yet existing paradigms remain limited: model-based methods require intrusive backbone updates, while memory-based methods expend substantial visual-encoding computation on temporally redundant content and rely on rigid access to visual history. To address these limitations, we introduce StreamFlow, an efficient visual memory framework that enables dynamic, on-demand access to historical visual information. StreamFlow combines a lightweight, dynamics-aware mid-term memory that filters temporal redundancy before visual encoding with a latent long-term memory that consolidates historical video content into visual latents accessible to subsequent reasoning. During generation, an attention-guided retrieval mechanism injects relevant visual latents when the model's reliance on visual evidence weakens. StreamFlow achieves state-of-the-art streaming video understanding performance, reaching 67.73% overall accuracy on StreamingBench, while also delivering strong performance on offline long-video benchmarks. Relative to the vanilla setting, it improves the visual attention score (VAS) by 59.1% while reducing end-to-end latency and peak memory by 50.4% and 21.1%, respectively, enabling more visually grounded and efficient reasoning.