生成模型与LLM推理优化

Dense causal attention remains expensive at long context even when implemented with highly optimized exact kernels. We study BF1, a deterministic block-aligned dyadic sparse-attention route that combines a small exact local neighborhood, a global first block, and logarithmically spaced historical blocks. The route is related to prior log-sparse and dilated attention patterns; our contribution is a correctness-gated pretrained-model retrofit, a matched topology-control study, and a systems characterization that connects per-layer sparsity to whole-model latency. For fixed block width, every converted layer uses O(n log n) selected token interactions and has O(log n) graph communication depth. On an NVIDIA RTX PRO 6000 Blackwell GPU, an optimized BF16 implementation crosses dense attention between 2K and 4K tokens and reaches a 10.91x per-layer prefill speedup at 32K. Retrofitting eight of 28 Qwen3-0.6B attention layers lowers warm whole-model time to first token by 7.7%, 11.3%, and 15.3% at 8K, 16K, and 32K, respectively, while the remaining dense layers keep the complete model asymptotically quadratic. Under a matched 1,000-step, 16.384M-token adaptation protocol, BF1 ranks first across three training seeds: mean report perplexity is 1.68639 versus 1.69154 for a matched static-random nonlocal graph, 1.69258 for dense continued training, and 1.81505 for equal-budget local sliding. At seed 1234, the packed-report paired interval places Dense-CT 0.3169-0.4055% above BF1 and static-random graph 17 0.2441-0.3642% above BF1. These results establish BF1 as a reproducible sparse operator and selective retrofit primitive with real long-context systems value. This paper evaluates numerical correctness, selected-interaction scaling, kernel performance, partial-model inference, and matched next-token language modeling.

Recent Visual-Language Models (VLMs) have enhanced the capabilities of pre-trained LLMs by adding vision tokens alongside text, with approaches like LLaVA showing impressive results. However, the computational burden of processing up to 576 or 729 visual tokens makes edge deployment challenging. While various token pruning techniques require retraining, some are training-free and thus can easily adapt to architecture changes. We introduce ClustRS, a two-part, training-free algorithm for robust token pruning. Its first component is an attention-weighted, clustering algorithm that selects representative tokens from each semantic cluster. The second component, Residual Shrinkage, is a one-pass denoising step on the selected tokens. These training-free lightweight steps make LLaVA ready for real-world data, improving robustness to a wide range of image-noise types and intensities. Experimental results on the ScienceQA-IMG and MM-VET benchmarks show our method outperforms attention- and diversity-based methods by up to 20% under extreme noise and token conditions (reducing tokens by 97%, down to 16 tokens) on LLaVA 1.5 7b and achieves exceptional results on LLaVA-OneVision, where we match baseline performance with fewer than one-third of their tokens under mild noise conditions. Our study demonstrates a simple yet powerful alternative to both score-only and diversity-only pruning rules, paving the way for compute-efficient and noise-resilient VLM deployment.

Hybrid Transformer-Mamba large language models (LLMs) enhance long-context efficiency, but their heterogeneous computation and communication patterns complicate efficient hardware acceleration. Chiplet-based architectures offer a scalable solution by integrating specialized compute and memory units. However, the design space spanning static architectural configurations and dynamic runtime policies is prohibitively large to explore exhaustively. To address this challenge, we present HYDRA, a comprehensive design space exploration framework for hybrid LLM serving on heterogeneous chiplet systems. HYDRA jointly explores chiplet composition, placement, inter-chiplet bandwidth provisioning, dynamic batching, and runtime scheduling. It integrates communication-aware placement, dynamic batching, elastic task scheduling, and a fast Markov-based performance estimator that captures multi-tenant runtime dynamics for efficient and accurate exploration. Across all workloads, HYDRA delivers 1.55x the throughput and 43.7 percent lower time-to-first-token on average, with throughput gains reaching up to 2.3x compared to state-of-the-art baselines. These results highlight that co-designing architecture and runtime policies is critical for efficient large-scale LLM serving on heterogeneous chiplet systems.

Lazy container-image pulling promises to eliminate the dominant cost of starting a model-serving pod by mounting the image immediately and fetching content on demand. We evaluate this promise for model delivery on Kubernetes, using KServe with two production lazy-pulling systems — eStargz/stargz-snapshotter and AWS SOCI — against eager baselines, on artifacts from 2 to 140 GB including real fp16 weights. Lazy pulling delivers its headline: cold time-to-first-prediction becomes size-independent (16.9--17.6s, versus 24.5--573.0s eager). But the cost is deferred, not eliminated: a full read of a 14 GB model through the lazy mount takes 105.3s, slower than the 72.4s eager pull it replaced, and the two systems pay at opposite lifecycle ends (SOCI prefetches nearly the full image before Ready; eStargz defers nearly everything to first read). More consequentially, we characterize a failure mode eager pulling structurally cannot exhibit: under sustained legitimate reads with default configuration, the snapshotter's node-level cache exhausts its finite volume and already-running pods begin failing reads of model files. At the earliest stage of exhaustion, an instrumented serving pod passed every Kubernetes-visible and application-level check for 196s while its snapshotter was already logging real failures; under heavier pressure, 67--94% of model files fail, scaling monotonically with residual cache occupancy. A live pod self-heals if cache space is freed under it, but a snapshotter-daemon restart under a live pod leaves permanently stale file handles in a pod still reported Running. We derive placement, monitoring, and cache-sizing guidance for serving platforms and operators.

Graph neural networks (GNNs) are widely used, but how parameter sparsity affects the expressivity of relational (RGNNs) and temporal (TGNNs) variants is poorly understood. The Strong Expressive Lottery Ticket Hypothesis (SELTH) posits the existence of sparse GNNs that preserve Weisfeiler-Leman (WL) expressivity on static graphs. We generalize this existence result to a probabilistic statement for multi-relational and temporal domains via the relational WL (RWL). We prove that sufficiently parameterized RGNNs contain sparse subnetworks that maintain 1-RWL expressivity and derive a lower bound on the probability that a random pruning yields such a subnetwork. We show that common TGNNs and cross-graph message passing schemes admit RGNN reformulations such that they inherit these guarantees and, moreover, that the expressivity of a sparse RGNN is connected to its optimization behavior under common update regimes. Experiments instantiate the bound, compare it to empirical probabilities on synthetic data, and study how pre-training expressivity relates to optimization and prediction quality metrics on temporal and molecular benchmarks.

Large vision-language models (LVLMs) often hallucinate, generating content that the input image does not support. Preventing such content during decoding calls for a candidate-specific measure of how strongly the image supports the token under consideration. The model's visual-token states offer a natural source of this evidence because projecting each state through the output head reveals which vocabulary items that position favors. These position-wise readouts cannot be pooled directly because their probability magnitudes are not comparable across visual positions. Vocabulary ranks provide a scale-invariant basis for pooling, but tokens still differ systematically in their typical rank-based evidence. We propose ReWEIGH, a training-free decoding intervention that aggregates these ranks across visual positions and compares each candidate with a token-specific reference estimated from unlabeled images. At inference, ReWEIGH caches the image evidence during prefill and applies a bounded penalty only to candidates that fall below their reference. On four 7B backbones, ReWEIGH reduces hallucinated object mentions by up to 21.3% while largely preserving or improving descriptive and general performance. With evidence cached, the average added latency is 1.33% per token, and the reductions extend across six architecture families to 32B parameters.

Training-free block-sparse attention can accelerate video transformers, but row-wise attention concentration does not by itself specify an executable sparse operator. Queries sharing a block route may have poorly overlapping supports, while retained attention mass alone does not determine the post-softmax error from skipped interactions. We show that partition geometry affects both pooled support and the predictability of the remaining residual from the sparse output. We introduce SparsePR, which combines Response-Coupled Partitioning with Probe-Fitted Residual Reconstruction. Sampled-query key responses form paired K/V groups, whose centroids induce query-response coordinates for shared routing. A small set of exact query rows then calibrates a call-specific affine correction from the sparse output within the output subspace observed in the probe residuals. Across four heterogeneous video generation and world models, SparsePR consistently reduces attention-reconstruction error. Ablations show that probe fitting accounts for most of this reduction, while response-coupled partitioning lowers hard-drop error and improves reconstruction under a finite probe budget. SparsePR preserves generation quality at 22.0-26.0% realized executed-pair density while achieving 1.48x-2.61x end-to-end speedups. Project page: https://pardistaghavi.github.io/SparsePR-website/

We present Flama, an open-source Python framework for developing and deploying production-ready web APIs, machine learning services, and large-language-model (LLM) applications. Built on the Asynchronous Server Gateway Interface (ASGI), Flama offers a type-driven, async-first programming model that unifies REST API development, predictive model serving, and generative AI inference in one architecture. It is organised around seven subsystems: a component-based dependency injection system resolving handler parameters from type annotations at startup; a pluggable schema layer supporting Pydantic, Marshmallow and Typesystem behind a single adapter; an automatic CRUD generator turning a SQLAlchemy table and a schema class into REST endpoints backed by the Repository and Unit of Work patterns; a portable binary format (.flm) packaging models from scikit-learn, TensorFlow, PyTorch and Hugging Face Transformers with their metadata for zero-code deployment; a multi-backend LLM server running vLLM (Linux/CUDA) or MLX (Apple Silicon) and exposing four wire protocols (OpenAI, Anthropic, Ollama, and a native streaming dialect) through a shared codec; a Rust-accelerated core compiled via Maturin for routing, JSON encoding, compression and parsing; and a Model Context Protocol module turning any application into an MCP server over JSON-RPC 2.0. Built-in capabilities include JWT authentication, two pagination strategies, background tasks in threads or processes, WebSocket endpoints, Server-Sent Event and NDJSON streaming, OpenAPI 3.2.0 generation from handler signatures, and a command-line interface for running applications and for serving, packaging and inspecting models. We describe the architecture, present the programming model through worked examples, and compare Flama with existing frameworks, model serving platforms and LLM inference engines.

Vision-Language-Action (VLA) models process long multimodal token sequences, making inference expensive in both memory and computation. Existing efficiency methods mainly reduce visual tokens, but aggressive token pruning becomes fragile because removing a token discards its entire representation. Sub-token compression provides a complementary alternative by retaining more tokens while reducing their value width. However, directly applying sub-token compression to VLA policies is less effective because information important for perception, language understanding, and control is distributed differently across the multimodal representation. We introduce Role-Conditioned Sub-Token Routing (RoleSub), which learns how to compress the value representations of retained tokens. After visual token reduction, RoleSub partitions each retained value representation into groups in an orthogonal space and uses a lightweight router to determine which groups should be preserved. The routing decision is conditioned on the token representation, a learned latent role representation, and language context. The same mechanism can also be applied to language values, allowing visual and language representations to be compressed without removing additional tokens. We evaluate RoleSub on OpenVLA-OFT-7B across the four LIBERO suites. At matched visual-KV budgets, RoleSub outperforms a trained token-only control in 33 of 36 settings, with the largest gains under aggressive compression. Combining visual and language compression reduces total KV to 9.2--11.3% of the original while retaining strong control performance on most tasks. These results show that reducing the representation within retained tokens provides an effective complement to token pruning for aggressive VLA compression.

Proactive interference (PI) is a documented failure mode in large language models in which retrieval of a repeatedly overwritten value degrades as prior overwrites accumulate, mirroring a classical phenomenon in human working memory. Post-training quantization (PTQ) is now the default deployment path for open-weight models, yet its effect on this failure mode has not been tested. We evaluate three precision levels (FP16, INT8, INT4/NF4, via bitsandbytes) across three architecturally distinct instruction-tuned models (Qwen2.5-7B-Instruct, Mistral-7B-Instruct-v0.3, Phi-3.5-mini-instruct), holding the retrieval task fixed. INT4 quantization significantly reduces accuracy under high interference in every model (e.g., from 81.0% to 68.3% for Qwen), confirmed by paired McNemar's tests ($p \le 2.6 \times 10^{-6}$) and a mixed-effects regression spanning all interference levels; INT8, often assumed safe, also carries a smaller but real penalty in two of three models. The effect is specific to semantically similar (word-type) distractors and reverses sign under a numeric control condition, and is mechanistically linked to a rise in same-key intrusion errors under INT4 (from 21.5% to 24.6% of trials, $p = 4.8 \times 10^{-7}$). A follow-up ablation shows the effect originates in the quantized transformer backbone rather than the output projection layer. These results suggest that bitsandbytes 4-bit quantization can impose an additional cost on applications relying on long, updatable, semantically dense contexts, even when aggregate benchmark accuracy appears largely unaffected. We release our code and tokenizer-verified vocabulary construction method at https://github.com/ShayanShahrabi/compress-and-forget

Spiking Neural Networks (SNNs) have emerged as an energy-efficient alternative to Artificial Neural Networks (ANNs), leveraging sparse accumulate operations in the place of power-hungry multiply-and-accumulate operations. ANN-SNN conversion is a widely adopted approach to realize deep SNNs with accuracy comparable to that of ANNs. The Quantization-Clip-Floor-Shift (QCFS) activation minimizes conversion error, yet requires a large number of inference timesteps to match the source ANN accuracy on real-world vision datasets. PASCAL addresses this by proposing the Precise ANN-SNN Conversion Integrate-and-Fire (PASC-IF) neuron, which guarantees mathematical equivalence between the converted SNN and the source ANN, thereby achieving ANN-equivalent accuracy at significantly reduced timesteps. Despite this algorithmic advancement, the hardware implications of deploying the PASC-IF neuron remain unexplored. In this work, we present APEX, a dual-sparsity SNN inference accelerator that integrates the PASC-IF neuron into the LoAS hardware framework. The three-stage PASC-IF datapath is realized as a fully combinational circuit with no additional latency cost. APEX exploits dual sparsity in both input spikes and weights through a fully temporal-parallel dataflow, enabling efficient sparse computation and reduced memory traffic. Across all evaluated models, the PASC-IF neuron on average achieves up to 3% higher accuracy than the standard IF neuron, with a power overhead of only 1.3%-5.4%, an area overhead of 2.1%-2.7%, and 40% energy reduction for best accuracy configurations.

Inference with transformer models on CPUs is increasingly important, especially for Small Language Models (SLMs), where vector architectures are emerging as a promising execution substrate. The attention module is a major bottleneck due to high memory bandwidth requirements; FlashAttention mitigates this by fusing operations to improve data locality and reduce intermediate memory traffic. In this paper, we present FlashAttention-V, a blocked FlashAttention for scalable vector architectures that adapts efficiently from short to very long vectors by exploiting parallelism across attention heads, inter-head packing to enable efficient utilization of vector lengths beyond the head dimension, and improving vector register utilization and memory access locality. We integrate FlashAttention-V into ggml within llama.cpp and evaluate it on TinyLlama, Llama 3.2, Qwen2.5, and Pythia-410M using gem5 and a Banana Pi BPI-F3. On the Banana Pi BPI-F3, we confirm that loop reordering and loop unrolling across attention heads are effective optimization principles, scaling performance gains with larger models and most pronounced with short contexts and during decoding. Simulation-based analysis shows that FlashAttention-V achieves 22x-42x speedup over scalar FlashAttention at 512-bit VL in prefill, with an additional 2x-2.5x gain scaling to 64 lanes and 4096-bit VL. During decode, FlashAttention-V achieves 8x-11x speedup using 512-bit vector lengths over scalar FlashAttention, with performance showing diminishing sensitivity to vector width and lane count due to single-token, memory-bound execution. We further identify structural bottlenecks in Q8_0 quantized linear layers that limit arithmetic amortization under long-vector execution, consistent across RVV and Arm SVE, indicating that current quantization formats pose a fundamental challenge to long-vector scalability.

Modern Intel AI PCs ship capable integrated GPUs and NPUs with 16+ GB of unified memory, and they spend considerable time idle. That is not enough memory to fit a large model such as a 70B-parameter LLM. We show that a handful of AIPCs, working together over an ordinary network, can serve models beyond the capability of any single one. We use pipeline parallelism: a model is split by layer into per-stage shards, each pre-compiled into an OpenVINO graph, so that every machine runs one shard and passes activations to the next. Three techniques make this fast enough to be useful. First, we recover the speed of the unsplit model: a naive per-stage export runs well below monolithic inference because it misses an OpenVINO GPU optimization, and injecting a beam_idx Gather into each shard triggers that optimization (the IndirectKVCache fusion) and brings the shards to parity. Second, we leverage speculative decoding on stateful OpenVINO models. Third, the pipeline serves several users at once by interleaving their requests across the stages, each request carrying its own cache (micro-batching). Together, a two-node Llama 3.1 8B INT4 pipeline serves two concurrent users at 1.79x the single-user throughput of the unsplit model on the same hardware, and the gap widens under simulated wide-area latency. The same design scales to a 70B model that no single fleet member can hold: a four-node deployment of Lunar Lake AI PCs on Intel Tiber Cloud serves a single user at interactive speed, with output token-for-token identical to the same four-node pipeline decoding without speculation. Code, raw benchmark logs, and reproduction scripts ship as a self-contained package at https://github.com/labscommunity/pipeline-sharded-inference-paper (in the top-level reproduction/ directory).

Small NLP models, especially BERT-family encoders, remain important in industrial workloads such as classification, ranking, and retrieval even in the era of large language models. On server CPUs, INT8 quantization offers an attractive latency-throughput-cost trade-off, but users increasingly expect such acceleration to be available directly in the native PyTorch stack. We integrate SmoothQuant into TorchAO and optimize the resulting inference path for Intel Xeon CPUs through graph-level fusion in TorchInductor and efficient INT8 GEMM kernel selection across oneDNN-, AVX512_VNNI-, and AMX-based implementations. Across BERT, DistilBERT, and XLM-RoBERTa benchmarks, the approach delivers up to 5.8x end-to-end throughput speedup with negligible---and in some cases no measurable---accuracy loss relative to the FP32 baseline. We also validated our work by detailed performance analysis with roofline models. The implementation has been upstreamed to PyTorch and TorchAO, enabling out-of-the-box deployment with native PyTorch tooling

Serving a 235B-parameter Mixture-of-Experts (MoE) model on a single 8 GB GPU is bottlenecked not by compute but by memory bandwidth: decode must stream each token's active experts from whichever tier holds them, and on consumer hardware most experts sit on an SSD far slower than RAM. We quantify this bandwidth wall on Qwen3-235B (Q4_K_M, 134 GB): measured decode is 0.44 tok/s warm, matching a bytes-per-token / bandwidth model, while a batching scheme that should amortize one disk sweep instead collapses at batch 32 from paging thrash. We build llama-moe-trace, a zero-surgery router-telemetry tool, and measure routing on Qwen3-30B: adjacent-token expert reuse is 2.0x chance, 95% of traffic uses 52.5% of experts, and an LRU cache of 13.4% of experts serves 66% of requests. We then ask whether cacheability is trainable: we pre-register training of 137M MoE language models with auxiliary locality and domain router losses, under joint criteria on cache-miss reduction and perplexity. The mechanism works (misses down up to 60%; a 99% static-pin hit rate) but every configuration fails the pre-registered <=1% perplexity gate — miss reduction and quality are tightly coupled. Concurrent StickyMoE reports the same loss as near-free on single-domain sub-25M models; on multi-domain 137M we find the tax real. Our contribution is this pre-registered, stricter-criterion, multi-domain evaluation plus edge-serving measurements. A 340M rung shows the tax does not shrink with scale (it rises slightly). We further show training-free cache-aware rerouting stacks with trained locality — together ~80% miss reduction at <=3.4% perplexity at both sizes, far cheaper than either alone — while domain-primed prefetching does not help. All code, traces, and the pre-registration are released.

Reinforcement learning for coding agents increasingly relies on long-running agent harnesses to manage tool integration, repository contexts, and execution feedback. However, the native execution environments of these harnesses are inherently misaligned with policy-gradient training: environmental crashes and reward hacking corrupt outcome signals, while train-inference discrepancies decouple rollout behavior from policy updates. To address this, we present LEGO-RL, a framework that bridges native coding-agent harnesses with scalable policy-gradient optimization without modifying their internal control flow. LEGO-RL is built upon three pillars: (1) faithful optimization via in-process LLM proxying that captures raw generation streams for token-level alignment and robust trainer-side log-probability recomputation, even under harness-side compaction or re-serialization; (2) reliable execution via scalable sandbox orchestration featuring image caching and stage-wise defenses to mitigate reward hacking; and (3) observable training through an integrated plugin that automates validation and monitoring, paired with a Live UI for granular trajectory diagnostics. We evaluate LEGO-RL by training the sparse MoE model Qwen3.5-35B-A3B with GSPO across three native coding-agent harnesses. LEGO-RL improves Qwen3.5-35B-A3B across OpenHands SDK (64.0% to 70.4%), Claude Code (62.4% to 68.2%), and OpenCode (57.2% to 66.6%) on SWE-bench Verified, while maintaining a rollout-training probability correlation above 0.99.

Regularized sparse regression has been extensively studied in the offline setting, but online formulation remains relatively under-explored. This gap stems from four key challenges: (i) the infeasibility of dynamically updating the regularization parameter in every online round, (ii) managing storage and memory complexity, (iii) enabling real-time computation via closed-form updates rather than solving full optimization problems at each round, and (iv) achieving optimal statistical guarantees under realistic assumptions. In this paper, we propose an online generalized-sparsity-constrained regression framework, focusing on online cardinality-constrained linear regression and low-rank matrix sensing. Unlike online regularized regression, our constrained formulation eliminates the need for dynamic parameter tuning. We introduce an efficient online hard-thresholding algorithm that performs closed-form updates and requires storing only summary statistics, making it computationally, memory, and storage efficient. Despite the inherent nonconvexity and combinatorial nature of the formulation, our algorithm achieves global convergence at the optimal statistical rate under realistic assumptions, provided that the projection set is properly overparameterized. Numerical experiments demonstrate that our method consistently outperforms state-of-the-art alternatives.

Real-time novel view synthesis is dominated by rasterized explicit primitives. These projection-based pipelines provide high throughput but require specialized extensions for non-pinhole effects such as distortion, rolling shutter, and depth of field. Ray-based rendering expresses these effects naturally but is generally assumed too slow for competitive real-time rendering. We analyze the factors governing throughput in differentiable Voronoi ray tracing and identify traversal length, per-cell work, and memory locality as principal determinants. Guided by this, we introduce VoroTracing, which co-designs the scene representation, optimization, and GPU execution to reduce these costs. Compact octahedral appearance textures reduce memory traffic, while surface-concentrated opacity promotes early termination. The fixed-budget representation is optimized without pruning or densification and rendered with a GPU implementation designed for coherent traversal. On Mip-NeRF 360, VoroTracing renders at 623 FPS on an RTX 5090, providing $3.2\times$ the throughput of the fastest prior ray-based method and $2.8\times$ that of 3D Gaussian Splatting, while maintaining competitive reconstruction quality. Our renderer supports fisheye, rolling-shutter, motion-blur, and depth-of-field effects through ray generation and sampling, requiring no specialized rasterization. These results show that real-time throughput can be achieved with the flexibility of ray-based rendering. We release our source code, see https://research.zenseact.com/publications/vorotracing

Heterogeneous graph neural networks (HGNNs) are highly effective in processing heterogeneous graph data and have been widely adopted in critical domains. As real-world graph data continues to scale, performing direct inference on entire graphs becomes increasingly infeasible, making mini-batch methods the standard approach. However, in end-to-end HGNN inference, metapath-based mini-batch sampling constitutes a significant performance bottleneck due to the extensive random memory accesses induced by the irregular traversal of graph structures. Existing sampling paradigms suffer from excessive redundant traversals caused by inherent semantic redundancy, severely degrading sampling efficiency and, consequently, leading to suboptimal mini-batch inference performance. In this work, we propose a redundancy-aware HGNN sampling paradigm that leverages a metapath trie to reuse traversal paths, effectively eliminating redundant memory accesses. We then map it onto a multi-channel hardware sampling unit denominated ESR-HGNN. Furthermore, we introduce a reusability-driven metapath grouping technique that optimally clusters metapaths to maximize reusable traversal paths within hardware channels, enhancing efficiency in scenarios with semantic parallelism. Extensive experimental results demonstrate that ESR-HGNN achieves an average sampling performance improvement of one order of magnitude over CPU and GPU, accompanied by significant energy savings. Additionally, it delivers substantial speedup in end-to-end mini-batch inference when integrated with GPU and state-of-the-art HGNN inference accelerator.

Recurrent models process long contexts efficiently by compressing their history into a fixed-size state, but modern architectures typically do so in a single causal pass over the sequence. Each input must therefore be compressed before the model knows how it will later be used, forcing a limited state to compromise across possible future demands. We introduce dynamic compression, which allows a recurrent model to selectively revisit past tokens and revise its fixed-size state through additional recurrent updates. The model need not preserve every part of the history at uniformly high fidelity in its recurrent state, because lower-fidelity information can be revisited from the retained raw sequence when it becomes relevant. We study this in a controlled setting where the model first learns multiple functions in-context and, later in the same sequence, encounters a series of few-shot tasks that each require it to identify and reuse one of those functions. A single-pass model must preserve every function at sufficient fidelity for any future task, whereas selective re-scanning allows the model to revisit and refine only the function currently needed. We find that dynamic compression substantially reduces the recurrent state required for accurate reuse and scales more favorably as the number of stored functions grows. These results demonstrate a computation--memory tradeoff in which recurrent models can spend more computation revisiting their history to make more effective use of a fixed-size state.

LLMs increasingly rely on external contexts, such as pre-defined system prompts or retrieved documents, to improve generation quality. However, processing these contexts alongside user queries creates an attack surface: adversarial inputs can induce models to disclose them. Prior probing studies suggest that leakage-related signals emerge in hidden states, yet the need to extract these states poses additional deployment challenges. In this paper, we explore whether this internal signal leaves a more accessible ``tell'' before decoding. We propose LeakGauge, which probes this response by appending a suffix that gauges leakage behavior and mapping its prefill token probabilities to an attack-risk score. While a direct gauge uses the initial tokens of confidential content, we find that a content-agnostic one that verbalizes leakage behavior yields more robust signals. Across 11 LLMs, including GLM-5.2 (753B) and Kimi-K3 (2.8T), LeakGauge reaches an AUROC range of 0.944--0.996 on unseen attacks. The signal remains stable when the content changes language or the attack shifts from verbatim to semantic disclosure. By activation-steering interventions, we further show that the risk score is sensitive to an internal leakage-related direction, relating the observable signal to the model's internal representation. In addition, LeakGauge enables an input detector with fewer than 0.5K extra parameters and added latency of 10.34 ms. Code: \href{https://github.com/yeasen-z/LeakGauge}.

Sparse attention mechanisms, which score all token pairs but propagate only the strongest, now underpin the most efficient Transformers for lightweight image super-resolution. This paper observes that sparsification changes what it means to improve such a network. A dense attention layer has one place where representation quality matters: the aggregation of attended features. A sparse layer has two, because the top-k operator first decides which tokens survive and only then decides what to do with them, and a token discarded at the selection stage cannot be recovered downstream. Selection quality and aggregation quality are therefore separable targets, addressed by modules placed before and after the attention respectively. We test this by pairing a dual-branch spatial enhancement on the input of a progressive focused attention with a wavelet-domain modulation on its output, forming SFMformer. Measuring each module alone and jointly over all fifteen benchmark-scale pairs, we find their gains are not additive: the joint gain exceeds the sum of the individual gains on nine pairs, and the sign of the discrepancy is predicted by how much the weaker module contributes on its own (r = -0.72), so the two compound when they relieve different constraints and overlap when they relieve the same one. Enabling spectral modulation once per block rather than once per layer retains the effect at roughly one-sixth of its cost, keeping the model below one million parameters at every scale. SFMformer ranks first on 28 of 30 PSNR/SSIM entries across five benchmarks and three upscaling factors. We report the cases where the pairing does not help, and deploy the model on a Raspberry Pi 5 to confirm the design is practical under tight resource budgets.

Large-scale, high-dimensional tabular regression remains challenging: tree-based models are robust but lack end-to-end representation learning, while deep models enable flexible feature learning but often incur costly interaction modeling and sensitivity to noisy or redundant features. We propose TabNSM, a scalable regression framework that extends our earlier sparse-attention and mixer architectures. At its core, the Adaptive Sparse Interaction Module (ASIM) integrates foreground feature discovery, sparse local interaction encoding, and Feature-Token Mixing, providing near-linear complexity under fixed sparse configurations. For regression, TabNSM introduces three complementary components: a Multi-Stage Regression Head for progressive prediction refinement; GridLoss, an ordinal-aware soft-binning objective that incorporates target structure into representation learning; and RISE (Reweighted Instance Sampling by Error), a difficulty-aware sampling strategy based on loss-quantile bins. Across nine real-world regression benchmarks, TabNSM delivers strong predictive performance and practical scalability, with particularly consistent gains on high-dimensional and heterogeneous datasets. These results demonstrate that selective interaction modeling, structured regression supervision, and difficulty-aware sampling provide an effective and scalable approach to deep tabular regression.

Reconstructing dynamic 4D scenes requires jointly estimating correspondence, geometry, object motion, and camera motion. Existing feed-forward methods typically predict dense task-specific maps or independently process source-target pairs, leading to unnecessary computation for sparse queries and limited feature reuse across different frame pairs. We present UniQuery4R, a query-conditioned framework that encodes a multi-frame clip once and selects the source view, target view, and continuous source-image coordinate only at decoding time via source-to-target cross-attention. Each query jointly predicts target correspondence, target-time 3D position, and scene flow, along with source depth, while camera parameters are estimated per view. This design allows the encoded clip to be reused across arbitrary source-target selections and supports both sparse inference and dense reconstruction through batched queries, without learned temporal embeddings tied to a fixed clip length. We further introduce a direction-magnitude parameterization of scene flow with separate supervision for moving and static points. Among the evaluated methods, UniQuery4R achieves the best macro-average results on WorldTrack for both scene-flow estimation and dynamic-point reconstruction.

High-resolution image editing is increasingly demanded in professional workflows, yet existing diffusion-based models remain constrained to resolutions below 1K due to quadratic attention complexity and prohibitive memory requirements. A prevalent workaround employs a two-stage pipeline: editing at low resolution followed by independent super-resolution. However, this approach suffers from two critical issues: information divergence, where hallucinated details contradict the original high-resolution (HR) source, and texture degradation, manifesting as over-smoothed or over-sharpened artifacts. We propose EditBridge, a diffusion bridge framework for efficient ultra high-resolution editing. Unlike conventional diffusion that regenerates from noise, we formulate refinement as structured data-to-data translation from the low-resolution (LR) edited result to its HR counterpart, explicitly conditioned on the original HR source to preserve authentic details. To efficiently incorporate HR source guidance, we introduce a prior-guided block-wise sparse attention mechanism that exploits semantic correspondence from first-stage editing to constrain cross-image interactions to spatially aligned regions, significantly reducing computational overhead. Extensive experiments demonstrate that EditBridge achieves high-fidelity editing with superior perceptual quality at resolutions up to 4K, delivering 3.6--8.4$\times$ speedup at 2K and enabling practical 4K editing in 61 seconds.

Diffusion Transformers (DiTs) achieve high-quality generation but are costly due to iterative sampling. Dynamic-resolution sampling reduces early-stage cost by denoising at low resolution; however, uniformly upsampling all latent tokens at resolution transitions incurs redundant computation and may degrade fine-detail consistency. Existing partial upsampling strategies typically rely on local latent structure cues or single-step statistics, making it difficult to jointly capture token-text semantic relevance and token-wise representation dynamics across diffusion steps. We propose AViTS, an adaptive spatiotemporal token selection framework for dynamic-resolution DiTs. AViTS models spatial importance via latent-text attention and temporal importance via token-level feature variation across diffusion timesteps, and fuses them to enable spatiotemporal importance-aware selective upsampling: it prioritizes resolution refinement for critical tokens while deferring less important ones, thereby reducing redundant high-resolution computation and improving the quality-efficiency trade-off. AViTS achieves up to 6.34x on FLUX and nearly 9x FLOPs reduction on Qwen-Image-Edit and FLUX.1-Kontext-dev, orthogonal to distillation, quantization, and feature caching, and reaching 14.76x with distilled models. Code: https://github.com/QHR69/AViTS

Hierarchical KV-cache systems can retain long-context LLM execution state beyond GPU memory, but retention capacity does not determine the local memory required to make that state executable again. We isolate this second resource as the restoration working set (RWS): the peak local staging state whose lifetimes overlap during restoration. In the pinned upstream LMCache whole-plan path, measured full-reuse points for 1.956, 7.823, and 15.646 GiB/rank states first succeed at 2, 8, and 16 GiB L1 rungs, with successful L1 peaks of 1.956, 7.824, and 15.648 GiB/rank. We introduce Bounded-State Restoration (BSR), which separates complete discovery from local residency. BSR probes the complete reusable prefix without materializing the whole hit in L1, then installs confirmed state through a reusable window of at most $W$ chunks. Under bounded auxiliary state, peak restoration capacity is $O(W)$ while total transfer and installation work remains $Θ(|S|)$. Because reusable state spans heterogeneous allocator groups and tensor-parallel ranks, BSR uses a request-level commit rule: partial installation is never exposed as a valid reusable prefix; failures invalidate the advertised prefix and fall back to a lower valid tier or deterministic recomputation. On DeepSeek-V4-Flash with TP=2 across two DGX Spark nodes, a clean no-resume sweep grows external state from 1.956 to 31.277 GiB/rank while measured L1 RWS remains exactly 500.75 MiB/rank at $W=32$, a 63.959x largest-state external-to-live-staging ratio. A second fresh 524K-token run repeats the largest-state acceptance result. Evaluated tier and rank-asymmetric failures expose either complete reuse or zero external reuse before fallback. A matched SSD optimization reduces 512K restore TTFT from 43.1 to 17.6 seconds without changing RWS.

Long-context prefill in large language models (LLMs) incurs substantial computation and memory traffic because dense self-attention computes quadratic query-key scores. Existing methods either use a uniform low-precision path or select token interactions, leaving spatial precision routing over hardware-aligned score tiles outside fused dense attention. We introduce TileMix, a tile-centric precision-routing kernel that makes numerical precision an executable spatial decision over score-tile groups within fused dense attention. TileMix partitions the attention matrix into hardware-aligned score tiles, packs routing decisions into compact bitmasks, and dispatches each tile group through FP16 or INT8 score computation while both paths update a shared online-softmax state. Scalable precision grouping lets each routing bit govern multiple adjacent key tiles, preserving hardware-aligned compute tiles and compact metadata at long contexts. By routing all legal tile groups, TileMix preserves dense token connectivity, requires no training, and supports grouped-query attention, variable-length batches, and INT8 key/value caches. Across LongEval, LV-Eval, and A100 prefill benchmarks on LLaMA, Qwen, and Vicuna, TileMix recovers long-context quality lost under uniform INT8 and improves prefill throughput over FP16, yielding a controllable accuracy-efficiency frontier across model families. The implementation is available at https://github.com/HanzhiZhang-Ulrica/TileMix.

Multi-agent AI workflows are limited not only by model quality but by token cost, latency, and context-window quality. This paper presents a practitioner framework for token optimization and context-window management, grounded in an internal production dashboard that extracts structured work items from meetings, email, and chat with LLMs and routes summaries across workstreams. Six patterns are described: context stratification, fetch-once/process-locally architecture, schema-contracted prompts, token-aware fallback chains, semantic caching, and inter-agent communication compression. In production they cut measured cold-load latency to 61-116 seconds (six timed runs) from an operational baseline of roughly 3.5-10.5 minutes, with an estimated 60-70% token reduction. It also reports a controlled context-composition study: 2,420 confirmatory trials across 11 model configurations, using 661 anonymized workplace items scored for relevance. Holding the prompt at a fixed ten items, replacing some high-relevance items with same-domain low-relevance items improves the model's relevance-score concordance on the target items, versus high-relevance items only; we call this relevance-contrast context. In the all-11 paired analysis, the 50:50 signal/noise condition improved relevance accuracy by +0.077 over the 100% condition (naive 95% CI [+0.056, +0.098], Cohen's d = 0.49, Holm-adjusted p < .001, n = 220). These cells are not independent; by the nine model families the effect is +0.084 (95% interval [+0.064, +0.103]), reported as a within-corpus descriptive comparison, not a population inference. A Fusion-of-N follow-up found that learned synthesis did not beat the mechanical set union of item IDs. The contribution is a measured engineering layer between model research and production agent practice: repeatable patterns and evaluation methods for faster, cheaper, more reliable workflows.

We present KernelArc, a multi-agent framework for autonomous GPU kernel optimization across heterogeneous workloads. Strategy-specialized agents run in parallel and coordinate through conclusions-only shared memory, a deterministic benchmark guard, and read-only cross-agent state with plateau-triggered drafting. We evaluate \kernelarc{} on NVIDIA H100 and B200 GPUs using category-representative SOL-ExecBench workloads. The resulting implementations span custom BF16 GEMM, static cuBLASLt Expert-API configuration tables, fused mixture-of-experts backward, shape-gated decoder-layer fusion, native NVFP4 grouped-query attention, and paged prefill attention. At the public SOL-ExecBench leaderboard snapshot recorded on July~30, 2026, these submissions ranked first on representative L1, L2, Quantization, and FlashInfer tasks. The trajectories support the paper's central motivation: shared multi-agent search can broaden exploration and reach stronger incumbents within a fixed candidate budget, while the value of individual coordination features depends on the kernel and optimization stage.