生成模型与LLM推理优化

Large language model agents increasingly act through stateful tools, yet model generation and environment execution remain serialized at every step. As decoding accelerates, tool execution becomes a growing bottleneck. Existing action- or observation-only speculation leaves much of this latency exposed: value is concentrated in a few slow calls, some outcomes emerge only through execution, and longer lookahead typically requires an increasingly unlikely chain of action predictions. We present AOSpec, a lossless framework that co-speculates actions and observations across the full agent-environment loop. Expected Value Decoding (EVD) directs observation speculation toward outcomes with the greatest expected latency benefit, optimizing expected time hidden rather than hit rate. For outcomes only execution can reveal, AOSpec launches latency-critical target actions in isolated forks that contain their effects, while Joint Action-State Verification (JASV) verifies both the action and its origin state against committed execution before reuse. JASV recasts long-horizon action dependency from full-chain prediction into target action-state verification, breaking the lookahead--accuracy tradeoff and unlocking long-range overlap without sacrificing serial semantics. Across Terminal-Bench serving settings spanning four harnesses, five actor models, and five serving speeds, AOSpec outperforms every practical baseline, reducing mean end-to-end latency by 11.8-32.5% and p99 latency by up to 42.8%. Its gains increase as decoding accelerates, and its observation model transfers from Terminal-Bench to SWE-bench Verified without retraining.

Kolmogorov--Arnold Networks (KANs) replace scalar edge weights with learnable univariate functions parameterized by multiple basis coefficients. This introduces a source of redundancy that conventional neural-network compression does not directly expose. We present SparseKAN, a unified approach that compresses KANs along three complementary axes: basis functions, neurons/channels, and numerical precision. SparseKAN equips the base branch, nonlinear basis branch, and individual basis terms with hierarchical learnable gates trained under a differentiable active-cost objective. The learned importance structure is subsequently hardened under explicit basis and width budgets, recovered in full or low precision, and physically compacted into smaller dense tensors rather than retained as sparse masks. Experiments on MNIST, CIFAR-10, and CIFAR-100 across spline, polynomial, RBF, wavelet, and convolutional KAN variants show that the structural axes compose predictably in cost. We also find strong basis-dependent differences in term importance: coefficient-based selection outperforms matched low-order truncation by up to 15.25 accuracy points in the evaluated Gram-polynomial settings. Eight-bit quantization is broadly robust, whereas 4-bit convolutional KANs require quantization-aware adaptation. Physical compaction removes up to 73.0% of parameters without accuracy loss on MNIST and reduces large-batch CUDA latency to as little as $0.51\times$ dense execution. On a ZCU104 FPGA, the resulting sparse low-bit models achieve up to $23.63\times$ lower inference latency, demonstrating that SparseKAN converts functional redundancy into measurable software and hardware efficiency. The SparseKAN implementation is available at https://github.com/OSU-STARLAB/SparseKAN.

Mixture-of-Experts (MoE) models have become a dominant architecture for large-scale AI services, yet deploying them over geo-distributed heterogeneous edge servers remains challenging. When the Top-k activated experts of a token are spread across multiple servers, the optimal routing depends jointly on cross-server link bandwidth, heterogeneous GPU computing capability, GPU-CPU expert loading delay, instantaneous queueing backlog, and replica-level quantization quality loss. Existing distributed inference and MoE serving methods address these factors separately and do not provide a unified framework for online multi-server collaborative routing. In this paper, we propose HetRoute, a heterogeneous-cost-aware collaborative routing framework for distributed edge MoE inference. HetRoute introduces a unified per-assignment cost model that explicitly captures four cost components: cross-server transmission, GPU-CPU offloading, GPU computation with queueing, and quantization-induced quality penalty. Guided by this model, the offline stage determines expert server placement, GPU-CPU residency, and replica precision through a routing-cost-coupled deployment algorithm, while the online stage routes the Top-k activated expert set as a whole by minimizing the bottleneck layer cost via exact enumeration or beam search. Theoretical analysis establishes fallback feasibility, a bound on the number of participating servers, per-layer optimality for small candidate domains, and online computational complexity. Trace-driven evaluation on three MoE models over a heterogeneous 10-server edge testbed shows that HetRoute reduces average inference latency by up to 59.0% and P99 latency by up to 58.0%, cuts cross-server traffic by up to 72.1%, and achieves 2.13x throughput improvement compared with representative baselines, while keeping quality degradation within the configured budget.

Production LLM serving generates millions of diverse requests, making full-trace replay across serving configurations increasingly expensive. Existing trace reduction methods mainly preserve workload distributions or representative requests, but bottleneck-revealing workloads may be rare and non-representative. Moreover, evidence for one component cannot compensate for missing evidence in another, while using predicted bottlenecks as target truth creates circular evaluation. These limitations make it necessary to preserve evidence for every bottleneck component rather than rely on workload representativeness alone. We propose Bottleneck-Preserving Witnessing (BPW), a quality-constrained framework for compact and diagnostically reliable LLM serving replay suites. BPW first performs Workload Candidate Nomination using response-blind workload features and closed source-side measurements. This stage identifies workloads that may expose scheduler, prefill, decode, or KV-cache bottlenecks. Coverage-Priority Sequence Construction then organizes multi-component proposals as reusable hyperedges and prioritizes weak and uncovered dimensions. Finally, Bottleneck Truth Verification derives prediction-independent labels solely from direct target-system measurements. The verified results determine the earliest prefix satisfying the direct two-witness requirement for every component. Experiments on BurstGPT, ServeGen, and Mooncake show that BPW reaches the verified gate with a compact workload set and outperforms 16 policies, achieving relative improvements of 2.3% and 16.3% in Mean prefix Macro-F1 and WBRC-AUC, respectively. Stage-resolved and sensitivity analyses confirm the distinct contributions and local stability of its three stages. Our code is publicly available at https://github.com/llmllmllm/BPW

The growth of context window lengths in Large Language Models (LLMs) significantly enhances their long-context capabilities but incurs prohibitive memory costs due to the Key-Value (KV) cache. Although low-rank compression of KV cache is a promising remedy, existing methods face a dilemma: offline approaches depend on external calibration data, whereas online approaches incur substantial compute for full-prompt decomposition and reconstruction. In this paper, we propose S$^4$R, which builds low-rank subspaces from selectively sampled tokens and computes attention over a sparsely reconstructed KV representation. S$^4$R uses prompt-aware initialization to build initial key/value bases from a representative prompt subset, trading off calibration-data dependence against prefilling cost. Because fully reconstructing the cache at every decoding step is prohibitively expensive and hurts throughput, we further adopt sparse reconstruction to retain only informative positions during decoding. Extensive experiments on LongBench and RULER with Llama and Qwen model families show that S$^4$R achieves up to 5$\times$ KV compression with near full-cache accuracy, combining the efficiency of fixed compression with the adaptability of prompt-dependent methods.

Vision foundation models, such as DINOv2, learn highly expressive representations but rely on massive, opaque architectures that demand substantial computational power and memory. To provide an interpretable-guided and efficient solution to this issue, we first propose a spectral analysis and new visualization technique for individual attention heads based on the Laplacian eigenvectors of their attention maps. Building upon recent observations regarding the block structure of Vision Transformers, we perform semantic clustering of attention heads and identify functional redundancies. Leveraging these insights, we introduce SAPER (Soft Attention PrunER), an end-to-end differentiable pruning framework based on the LapSum Soft Top-K approach. Extensive experiments on ImageNet-1K demonstrate that SAPER achieves a highly favorable accuracy-efficiency trade-off, outperforming the competitive RAPTOR baseline in FLOPs reduction while preserving strong classification performance.

Surgical augmented reality (AR) can provide contextual guidance by overlaying virtual annotations, tool cues, and procedural information onto the surgical workspace. However, the virtual content may obstruct task-relevant real-world information, such as surgical instruments, and interfere with users' perception during time-sensitive surgical tasks. In this paper, we investigate visual obstruction detection for surgical AR and present a latency-aware pipeline that combines vision-language model (VLM)-based surgical-object recognition with segmentation-based obstruction reasoning. To reduce inference overhead, the system adopts a cascaded small-to-large VLM architecture with segmentation-guided early exiting and attention-based visual token pruning. The small VLM handles easy frames when its key-object prediction is supported by segmentation consistency, while difficult frames are forwarded to a large VLM with pruned visual tokens. We construct a pseudo-AR surgical obstruction detection benchmark by overlaying virtual content onto surgical-tool images and labeling whether the virtual content obstructs task-relevant instruments. Evaluation results show that the proposed system achieves 87.43% obstruction detection accuracy with an average end-to-end latency of 479 ms, reducing latency by 62.90% compared with a cloud large-model baseline. These results demonstrate the feasibility of latency-aware obstruction detection for surgical AR and motivate future work on dynamic surgical videos, multi-object scenes, and clinically grounded AR guidance content.

Memory for self-evolving large language model (LLM) agents is often provisioned as if its byte budget only grows. Cloud platforms, however, adjust quotas with load and cost, and we show that capability does not follow the budget back up: after a squeeze-and-recover cycle, the agent settles below its pre-squeeze level, a gap we call memory hysteresis. The cause is structural. Deletion and one-way compression discard the material needed for later rebuilding, and we prove that any policy that only keeps or drops entries carries a residual-deficit floor. We propose CrystalMem (Crystallized Memory), an elastic memory sidecar that demotes entries across four fidelity states under a crystallization-energy schedule, orders demotions by advantage-weighted influence with dependency coupling, and recovers capability through verified recrystallization under explicit compute and byte caps. Across seven environments, seventeen methods, and six backbones, with multi-tenant serving and a physical edge-cloud deployment, CrystalMem achieves the highest restored capability in every setting and closes the loop left open by every baseline. From a 50% byte budget, CrystalMem matches the strongest budgeted baseline at full provision on every environment; at equal budgets, it leads by +4.6 pp on average.

The limited-memory BFGS (L-BFGS) Hessian update scheme is the critical kernel in many quasi-Newton optimization algorithms. The most common approach to implementing L-BFGS uses $2m$ sequential rank-1 updates as part of solving a linear system when there are $m$ history steps. The performance of this approach suffers when the latency of synchronization is significant, and its poor temporal locality increases the memory traffic when vectors do not fit in cache. The compact dense representation of L-BFGS results in an approach that has minimal synchronization latency and better temporal locality, but it requires an additional pass over the basis vectors and an additional basis that must be recomputed when the $B_0$ matrix changes as in variable-metric methods. In the Portable Extensible Toolkit for Scientific Computation and the Toolkit for Advanced Optimization (PETSc/TAO), we have implemented an intermediate dense formulation of BFGS that retains most of the good characteristics of both the recursive and compact dense approaches. We report single-node performance tests of these implementations on the U.S. Department of Energy's Polaris and Frontier machines, testing both GPU-based and CPU-based computations.

Deploying robot foundation models at scale is the next step towards realizing the potential of general-purpose robots. However, Vision-Language-Action (VLA) and other foundation models are computationally demanding, and on-device compute is constrained by power and space. In this paper, we introduce the problem of serving a robot policy to multiple robots from a remote GPU and formulate it as a scheduling problem. We build Armory, a serving system validated on fleets of both simulated and real robots. Our experiments show that naive scheduling heuristics perform well when all robots are the same, but fall short when robots consume action chunks at different rates, uncovering a mismatch between conventional batching methods and the closed-loop requirements of robot policy execution. To address this, we propose a scheduling algorithm that accounts for this heterogeneity and improves overall system throughput by up to $18%$ in real-world experiments. Additional details are available at https://gatech-rl2.github.io/actionchunkscheduling.

A 3D CT scan entering a vision-language model produces a long sequence of visual tokens, often thousands to tens of thousands per volume, and this sequence must be compressed before a language model can consume it. Token compression is well studied in general vision, but little of it targets 3D CT specifically. A common baseline is grid average, which pools regular grid cells and can blend distinct anatomy, lesion, and air into one token. We present ORCA (ORgan-Centroid Aggregation), a token compressor for 3D CT. It merges adjacent tokens with organ guidance and adds a sinusoidal encoding of each region's centroid to preserve spatial layout. This preserves the anatomical information a downstream model needs. ORCA is training-free and plug-and-play, producing an adjustable token set without any model change or text query. We evaluate it across two datasets (CT-RATE and Merlin) and five encoders. The evaluation spans two task types: attribute prediction over five families (size, density, location, texture, and disease) and text generation (visual question answering and report generation). At matched token budgets, ORCA improves consistently over existing compression methods. It shrinks the visual context $64\times$ and its KV-cache $50\times$, and is $31\times$ faster to process each volume. Code released at https://github.com/renjie-liang/ORCA-3DCT.

Deploying state-of-the-art deep neural networks (DNNs) at the wireless edge is severely bottlenecked by the strict energy and resource constraints of mobile devices. While federated split learning (FSL) mitigates on-device computation by offloading workloads to an edge server, this may introduce systemic overheads, while the continuous exchange of cut-layer data, and submodels still incurs significant energy consumption (EC). To address this, we propose a green quantized FSL (GQ-FSL) framework that incorporates stochastic quantization for both local collaborative training and wireless transmissions. Notably, GQ-FSL supports asymmetric precision levels for the client- and server-side submodels, effectively decoupling device energy constraints from global convergence degradation. To quantify these tradeoffs, we develop parameterized energy models for the split architecture and derive a theoretical convergence bound under statistically heterogeneous data. Building on that, we formulate a joint optimization problem to configure the DNN split point and precision levels, minimizing the total system EC while satisfying a strict target accuracy constraint. Ultimately, we demonstrate that GQ-FSL enables large-scale DNN deployment on resource-constrained devices, achieving superior energy efficiency compared to quantized federated learning and full-precision FSL.

Zero-shot image captioning (ZIC) describes images without paired image-caption supervision during captioner training, relying on text-only corpora and frozen pretrained image-text scorers. Existing retrieval-augmented methods score image-text alignment once, at retrieval, then commit the captioner's autoregressive beam under language-model probability alone, leaving the decoder without further visual grounding feedback. Progress has stalled, with no method improving on the strict-regime best since 2024. We propose Adjudicated Captioning, an inference-time multi-agent framework that restores grounding feedback at multiple checkpoints over an unchanged IFCap captioner. First, we install a stronger frozen Retrieval Encoder at the input. Second, between retrieval and decoding we insert a frozen Cross-Attention Verifier that re-ranks the top-9 retrievals to top-5. Third, at the output beam we attach a learned Reranker pairing TriFuse, a multilayer perceptron, with MemAttend, a memory-attended transformer, the pipeline's only learned components; both are trained self-supervised by Borda-consensus distillation across the three frozen scorers, using no paired image-caption labels and no reference captions. Under the inductive headline protocol, with rerankers fit on the disjoint COCO Karpathy validation beam and applied frozen to test, the framework reaches CIDEr 117.6 and SPICE 21.9 on COCO Karpathy, up from 108.0 and 20.3 for IFCap, a +9.6 CIDEr gain, and +7.7 above NES, the strongest synthetic-image-augmented method at 109.9, without retraining the captioner. A training-free fixed-fusion baseline reaches 115.8 CIDEr, so +7.8 of the +9.6 gain comes from the non-learned architectural intervention and the remaining +1.8 from the learned rerankers. The same recipe transfers off-COCO without captioner retraining: +8.1 CIDEr on Flickr30k Karpathy and +5.7 on NoCaps overall.

Autonomous agents challenge conventional LLM serving by coupling repeated inference with persistent context and sandboxed tool execution. We present Aries, a full-stack experimentation framework that separates task semantics from execution configurations, reconstructs cross-component agent trajectories with correlated system telemetry, and exposes stateful tool execution through a consistent interface across heterogeneous sandbox substrates. We use Aries to conduct reproducible experiments on open agent harnesses and benchmarks. We complement these experiments with production traces from a commercial platform, grounding low-level systems research in observed production behavior. Our results show that (1) token-centric metrics miss non-inference bottlenecks, (2) retaining additional context yields diminishing accuracy benefits while reducing serving capacity, and (3) tool sandboxes alternate between long idle periods and short resource bursts, while current snapshot-based state management makes aggressive suspension costly. A complementary security analysis further highlights the need to reduce the sandbox attack surface. We then discuss the vision for agent-native serving systems designed around trajectory-level metrics, adaptive context management, elastic sandbox resource management, and sandboxes with minimized attack surface.

Although world-action models (WAMs) enhance long-horizon robot control by predicting visual evolution before acting, long-horizon reliability demands repeated re-grounding in real observations--not recursive rollout. Existing WAMs address this by refreshing history or KV cache with ground-truth data between chunks. However, such chunk-wise feedback operates at a coarse temporal granularity and thus fails to correct prediction errors at the individual time-step level. To address this, we propose Feedback Flow Matching (FBFM), a training-free inference mechanism that pushes re-grounding inside the actively generated chunk. During flow matching, FBFM applies a masked pseudoinverse correction to the conditional velocity field: it leverages the preceding action chunk to guide generation of the next action chunk, and uses the image observed after executing that preceding chunk to guide the next frame prediction. This cross-chunk pairing--where feedback from one chunk arrives in time to shape the next--creates an asynchronous loop that corrects errors without waiting for chunk boundaries. Being training-free, the mechanism improves responsiveness to unexpected events and suppresses drift in long-horizon tasks. We evaluate FBFM on both a joint-generation WAM (DreamZero) and a stage-wise WAM (LingBot-VA). On selected LIBERO and RoboTwin2.0 tasks, it improves success rates by over 5% in favorable settings, and real-world robot observation-prediction diagnostics show notably better tracking. We argue that FBFM offers a new paradigm for fine-grained online correction, bridging open-loop flow generation with closed-loop real-world dynamics.

We address the challenge of securely and efficiently outsourcing AI computations from a trusted but computationally weak client to an untrusted but powerful server, in the setting where the client holds both the input and the model, and the server must learn neither. We present MOSAIC, whose core is a novel matrix-multiplication masking protocol that scales to far larger matrices than prior work, enabling the safe outsourcing of modern workloads such as large transformer inference. By introducing small amounts of noise to the multiplication result and thereby relaxing correctness, MOSAIC achieves optimal asymptotic client overhead and concrete runtimes orders of magnitude faster than prior work. Its security reduces to the decisional LWE and LPN assumptions. Because this noise accumulates across the many layers of a transformer, a key technical challenge is bounding error growth; MOSAIC addresses this with an error-scaling mechanism based on random Hadamard rotations. On large 70B transformer models, MOSAIC's perplexity is comparable to popular quantization approaches and even matches full-precision BF16 inference on HumanEval. Finally, we present an end-to-end implementation showing how ideas like MOSAIC can promise a path towards large-scale confidential AI in modern data centers. Non-confidential inference is already distributed across phase (prefill/decode), layer, and time to maximize utilization of heterogeneous hardware, using RDMA-like networking to move activations, cached KV values, and weights across nodes. MOSAIC enables scaling of confidential compute by keeping the trusted computing base (TCB) small and outsourcing the bulk of the AI computation to untrusted accelerators.

Diffusion models have revolutionized generative tasks but incur high latency due to iterative denoising. While cache-based strategies accelerate inference by reusing intermediate features, they largely rely on static, sample-agnostic schedules. We argue that this rigidity overlooks two facts empirically validated in this paper: (i) generation difficulty varies across prompts, requiring adaptive resource allocation--complex inputs demand more computation while simpler ones require less; (ii) error sensitivity fluctuates across timesteps, where static policies may cache high-error steps or waste computation on low-error ones. We therefore propose OnlineCache, a dynamic caching framework that jointly learns when to cache and how to correct approximation errors. We leverage policy gradient to train a lightweight network for adaptive speed-quality trade-offs, and incorporate a learnable corrector to mitigate caching-induced errors. Both modules are jointly optimized under a bilevel optimization framework, with the policy targeting global generation quality and the corrector minimizing local errors. Our method automatically allocates computational resources across both samples and timesteps, improving overall generation quality. Extensive experiments demonstrate clear superiority. On FLUX.1-dev model, OnlineCache achieves nearly 3 speedup while preserving generation fidelity. On DiT and CogVideoX, it similarly delivers competitive acceleration without compromising quality; across all scenarios, it consistently outperforms existing cache-based acceleration baselines.

Training-free acceleration makes diffusion-based multimodal large language models (dMLLMs) more deployable, but it may silently change generated content. We study this serving-time consistency problem on 300 real images, comparing Fast-dLLM outputs with the same model's unaccelerated outputs. Across the mild parallelism induced in our long-form setting (1.05--1.25 committed tokens per step), confidence-threshold tuning changes decoding behavior but not baseline agreement. State-refresh ablations and an image-swap intervention instead identify stale visual and generated-text states as contributors to drift. For the tested Fast-dLLM implementation, shortening the KV-cache refresh interval yields a monotonic speed--agreement frontier and near-exact agreement at a measured 1.3x speedup. The initial diagnosis also appears with dLLM-Cache and LaViDa, although dLLM-Cache recovers agreement only after both caches are tightened, which removes its speed advantage. Independent prompts and images reproduce the threshold-insensitivity and refresh recovery. A targeted audit finds genuine content substitution in half of 50 low-agreement pairs. In a separate blinded two-annotator evaluation, the pooled accelerated-minus-baseline factual-error difference is 0.00 (95% CI [-0.17,+0.17]); this sample detects no difference but does not establish factual equivalence. Finally, none of the tested adaptive or smoothed-refresh variants beats the fixed interval at matched compute. Our contribution is a paired diagnostic and an implementation-scoped consistency control, not an accuracy or safety guarantee.

LLM serving stacks cache prompt KV state, yet the front end still re-tokenizes the full request text on every call. Coding agents pay the most: each call resubmits a long transcript after a small append, and reuse is hard because a short append can move token boundaries near the end of the prior sequence. Across 153,951 agent calls, the median append is 1.4K characters; only 1.0-3.6% of calls start or rebuild a session, but those carry multi-million-character contexts. At the fleet's 94.1% prompt-cache hit rate approaching 0.99, tokenization grows from 10% to 64% of time to first token. TokTier is a stateful CPU+GPU tokenization service for this two-mode workload with one contract: emitted token IDs are always identical to full reference tokenization of the request text. For session continuations it re-tokenizes a small window around the append and splices only when a per-request check finds a stable pre-tokenization boundary, else it widens or falls back. For calls without a reusable prefix it decomposes GPT-family regex pre-tokenization into run-local rules and runs exact pre-tokenization and BPE on a GPU. A sampled shadow verifier re-checks live traffic. Differential campaigns over 17 production tokenizer families ($1.5\times10^{10}$ split checks, a 12.4TB real-text corpus, 93,000+ replayed agent steps) show zero divergence. Incremental repair takes 0.5-1.1ms from 100K to 3M characters, up to $437\times$ faster than HF tokenization and $2.1\times$ faster at 1M characters than the strongest cache-based baseline (Gigatoken) fully prewarmed. GPU full tokenization encodes 1M characters in 0.87ms, $491\times$ below HF and $23.4\times$ below the fastest published CPU method. With vLLM, median time to first token drops 16-34% and P99 23%; under a 50ms P99 objective, four repair cores plus one GPU sustain 1,821 requests/s where a 16-core stateless front end saturates at 40.

Large language model (LLM) serving commonly increases batch size to improve throughput, but performance eventually reaches a deployment-dependent plateau beyond which larger batches provide marginal gains while increasing latency and GPU memory consumption. Previous studies have attributed this behavior to HBM/DRAM bandwidth limitations, but the underlying causes have primarily been supported by conceptual arguments or high-level performance observations. As our first contribution, we present a detailed GPU characterization using hardware profiling techniques, demonstrating that throughput saturation originates in the attention kernels during the decode phase. Specifically, we show that their nearly constant arithmetic intensity as active-context lengths increases -not merely larger batch sizes- drives DRAM-bandwidth saturation, while the achieved compute throughput remains far below the hardware limit. Building on this analysis, we present the Batching Configuration Advisor (BCA), which selects the highest-throughput batching configuration satisfying a target latency constraint and identifies up to 55 GB of GPU memory allocation that can be avoided for the evaluated OPT models with minimal throughput loss. To enable these recommendations, we introduce SLIM (Saturation-Aware Lightweight Performance Model), a semi-analytical model that predicts LLM inference throughput and latency from analytical formulations of Transformer computation and memory traffic. Across the evaluated scenarios, SLIM outperforms representative performance-modeling baselines while successfully generalizing to previously unseen operating conditions.

Analog compute-in-memory (CIM) arrays have emerged as a promising substrate for energy-efficient LLM inference, particularly for weight-stationary computations in linear layers. However, extending analog CIM to attention mechanisms introduces a fundamental challenge: KV cache operations demand repeated in-situ weight updates, and the resulting mismatch with the weight-stationary paradigm exposes dynamic computations to significant hardware noise, a critical problem that remains largely unexplored. In this paper, we present the first systematic study of dynamic attention computation on analog CIM arrays, revealing that initial and recent tokens exhibit disproportionate vulnerability to hardware noise. Motivated by this token-level insight, we propose a hierarchical token protection strategy that keeps sink tokens and a sliding recent-token window on a higher-precision digital path while processing the bulk KV cache on analog CIM. A co-designed scheduler combines analog programming, ownership transition, and bulk-MVM tile formation to bound digital overhead. Evaluations on nine LLMs show that our approach lowers average perplexity under analog noise from 33.91 to 11.95, close to the clean baseline of 11.06, while improving dynamic-KV programming-row utilization from 23.1% to 91.2%.

Deploying large language models in realistic server environments poses challenges, as the system needs to provide high-quality responses with low latency. Quantization is a common approach to reduce the memory footprint and improve inference efficiency, yet its impact on latency and throughput is rarely evaluated under controlled, orchestration-level workloads. In this work we study the quantization trade-offs of two translation model families, EuroLLM \citep{martins2025eurollm} and Hy-MT2 \citep{zheng2026hy} across five models ranging from 1.7B to 22B for efficient deployment on a single A100 or H100 GPU. We demonstrate that combining a document-chunking strategy with W4A8 or W8A8 quantization improves the latency-throughput Pareto-curve under a wide range of workloads. Furthermore, since standard machine translation (MT) benchmarks rely on isolated sentences and fail to capture long-context dynamics, we introduce a document-level evaluation from WMT24++ to assess how text chunking strategies affect translation quality under quantization. Our results reveal that standard segment-level evaluation can fail to predict the interaction between quantization and long-context document translation. While Hy-MT2 remains robust under quantization, EuroLLM shows strong sensitivity and translation quality collapses rapidly for all considered quantization formats. Overall, our experiments show that the trade-off between inference efficiency and translation quality depends not only on the quantization format, but also on the choice of text chunking strategy.

Heterogeneous Large Language Model (LLM) systems increasingly rely on shared contexts, retrieved evidence, and multi-agent dialogue histories, yet their internal key-value (KV) caches remain model-specific and cannot be reused across architectures. Consequently, each model must repeatedly prefill or store caches for the same context, limiting the scalability of multi-model reasoning and long-context generation. We propose Mixture-of-Translators(MoT), a cache translation framework that maps context KV caches from a source LLM into the cache space of a target LLM. Unlike prior approaches that depend on a single projection path or global shared latent space, MoT uses multiple translator modules to capture diverse source--target mappings. To further reduce residual translation error, we introduce a Context Correction Loss that aligns the replayed target trajectory with the native target trajectory. We reveal two competing failure modes in cache translation: propagated translation shift from early injection and last-state shift from late injection. MoT addresses them through translator mixtures and target-side correction. Across homogeneous and heterogeneous translations among Qwen2.5, GPT-2, and OPT models, MoT preserves downstream QA performance, including Qwen2.5-7B-scale translation with 51.0% average closed-set QA accuracy and 0.43 average extractive QA F1. In practical case studies, MoT enables quality-preserving memory reuse for multi-agent reasoning and retains 96.3% of direct-context quality in long-context cache-augmented generation, demonstrating scalable KV cache reuse across heterogeneous LLMs.

AI coding agents like GitHub Copilot, Claude Code, and Codex interleave multi-step LLM inference with tool execution, creating a workload different from chatbots. We present the first production-scale characterization of this workload using sampled GitHub Copilot traces from June 2026, comprising 3.2M users, 13M sessions, 761M LLM calls, and 95T tokens. Our analysis reveals distinctive workload properties with important systems implications. For example, agentic coding sessions consist of sparse user-initiated turns, each unfolding into an autonomous agent loop of LLM calls almost always coupled with tool execution. This structure yields KV cache hit rates averaging 90% within a turn, but falling to 55% across turn boundaries and drastically invalidated after events like model switches or context compaction. Diverse workflows and user behaviors are observed with variable and long-tailed token consumption, time span, and tool calls. We highlight the difference between quick agentic turnaround times and the minutes-long user idle periods at turn boundaries, and design a lightweight idle-time predictor that captures 86-90% of total idle time, enabling proactive decisions for efficient resource orchestration. These findings challenge assumptions underlying current LLM-serving systems and provide an empirical foundation for agent-native infrastructure.

Large language model (LLM) workloads motivate multi-partition GPUs as a path to scaling compute and memory capacity, but their non-uniform memory access characteristics and inter-partition communication can amplify contention and degrade locality, leading to suboptimal kernel latency. To address this, we analyze performance-critical LLM kernel implementations spanning weight projection, mixture-of-experts, and attention variants of state-of-the-art serving engines to present a characterization of data access patterns in multi-partition GPUs. First, we introduce memory trace analysis methodology to derive workgroup-level data access and sharing behavior, then evaluate the locality implications on latency using a cycle-level simulator. Using these tools, we categorize LLM kernel operands into three inter-workgroup sharing patterns (global, partial, or private) and show that the required optimization strategies differ across categories, from simple per-workgroup pinning to subgroup-aware co-scheduling. Our findings highlight the need for placement-aware kernel programming and smarter architectural support for work and data locality in multi-partition GPUs.

KV-cache quantization is validated today by offline benchmark averages; a deployed system cannot tell whether compression is damaging the request it is serving right now. We give it a provably sound runtime meter, a "DTrace for KV quantization": a per-(layer, head, step) upper bound on the total variation between exact and compressed attention. The meter has two tiers: a deterministic band-norm-witness bound, sound for any cache-preserving black-box quantizer and for any query (adaptive-safe, worst-case Cauchy-Schwarz plus RoPE band-unitarity), and a tighter probabilistic certificate for a controlled subtractively-dithered INT8 quantizer under an explicit request-level failure budget (stated for non-adaptive queries; core theorems machine-checked in Lean 4). Three results. Observability: the meter enters SGLang through an environment-guarded patch, and any scheme registered as one tensor function is measured in live serving. Repair: meter-driven gating, risk-ranked where the witness is saturated and certified where it is informative, empirically restores the quality floor at benchmark scale. For example, raw-cast FP8 improves from 22.8 back to 79.7 on hard RULER tasks, with the difference from uncompressed bounded at [+0.0, +0.8] by a paired test. Analysis: aggressive schemes survive on cross-layer error cancellation, not per-step fidelity. In a 28-layer sweep, no single layer's pollution alone loses anything (0/28), and the certified INT8 cache serves 1.88 times more KV tokens at the same memory in SGLang.

Grounding human motion in language, and language in motion, is a central step toward physical AI systems that can understand, generate, and communicate human behavior. Unified motion--language systems first coupled the two directions through a shared discrete motion codebook, but quantization limits generation quality. The strongest generators buy quality back at growing cost: stacked residual codebooks enlarge the representation; masked decoding stages, long autoregressive rollouts, and denoising chains of tens to hundreds of steps stretch inference; even the continuous-latent designs among them reach their latent only through an iterative diffusion head; and none of this decoding machinery serves understanding. We therefore propose MUGEN, a unified motion--language framework that pays neither cost: no codebook, one draw. A single adaptive-length autoencoder compresses any-length motion into a few continuous latent slots, the system's only motion representation: the language model generates them for text-to-motion and reads them back for motion understanding. Depth-routed hidden states let each slot read from the transformer depth it needs, and a calibrated head predicts a joint distribution over the full latent set, so a single draw carries the text-conditional, cross-slot variation a description permits. At a decoding cost of K language-model steps, one draw, and one decoder pass, MUGEN leads language-model baselines on FID on HumanML3D while raising retrieval precision above the real-motion reference under the standard evaluator, achieves the best CIDEr and BLEU@4 scores, and surpasses the discrete-token state of the art on every retrieval and alignment metric on SnapMoGen.

Spiking Neural Networks (SNNs) communicate through sparse binary spike events rather than dense activations, enabling energy-efficient inference on neuromorphic hardware and motivating their use in always-on, battery-powered edge systems. We show that this same efficiency advantage creates a distinct security risk: sponge attacks can increase inference-time spike activity and synaptic workload, inflating energy consumption while remaining difficult to detect through correctness-based monitoring alone. Prior input-space efficiency attacks on SNNs have focused on per-sample optimization, primarily in rate-coded settings. We extend this threat to native event-based binary inputs and study two attack models. First, we develop a per-sample sponge attack that crafts a custom adversarial spike train for each input via gradient-based optimization. This attack increases per-inference SynOps by 1.5-2.6x on three SNN models for the NMNIST, SHD, and IBM DVS Gesture datasets, while preserving the predicted class on at least 98% of evaluated samples. Second, to the best of our knowledge, we introduce the first universal sponge attack for native event-based SNN inputs: a fixed binary perturbation computed offline and applied via XOR to all subsequent inputs. Although weaker, it still inflates SynOps by 1.09-1.24x across all three datasets and represents a more realistic deployment threat because it requires no per-input optimization. Mapping SynOp inflation to estimated Loihi-1 energy yields per-inference overheads from 14 $μ$J to 13.24 mJ. These results show that native event-based SNNs are vulnerable to practical input-space efficiency attacks, and that reusable universal perturbations can accumulate into meaningful battery drain in continuously deployed edge systems.

Mathematically equivalent expert-reduction orders can produce observably different sparse-MoE executions. We isolate this effect in native DeepSeek-V4-Flash by freezing local MoE state and varying only aggregation semantics. Four schemes separate operand representation from accumulator precision. At one layer-5 fork, 720 A-mode orders yield 10 continuation basins; 720 B-mode orders form 360 exact structural classes and 11 basins. Under one Chinese prompt, the B classes split into 202 layoffs, 113 hiring, and 45 other continuations. Maximum-L-infinity B-branch selection separates 12, 24, and 36 of 50 prompts by 8, 16, and 32 tokens. Across 192 persistent trajectories per scheme, P32, A, and B change every native-reference route trajectory, while C preserves routes, token sequences, and texts. A separate 192-trajectory C check matches native MoE, post-mHC, next-router, and LM states bitwise. For one controlled B branch, exact post-mHC endpoint reconstruction reproduces the measured downstream trajectory. At the next decode boundary, exact FP64 reconstruction of the branch's full persistent state yields agreement for 301 downstream post-mHC states, 301 persistent-state checkpoints, 301 routes, predictions, and text over seven steps, given the same naturally generated next input. These controls identify post-mHC as an intra-token boundary and full persistent state as a cross-token continuation boundary. Identical tokens need not imply identical autoregressive state: divergence can survive a token boundary and become visible later. These results make expert operand conversion, accumulator precision, and reduction order part of a numerical compatibility contract for sparse-MoE runtimes and hardware backends. They establish controlled causal possibility, not deployment incidence; C's order invariance is limited to evaluated six-term states and schedules.

Time-varying implicit neural representations (INRs) provide a compact representation of scientific volumes and, for modalities such as dynamic X-ray computed tomography (CT), are often the only practical way to represent the data. However, interactive volume rendering of INRs is challenging, as cheap memory lookups are replaced by expensive neural inferences, hindering the performance. Therefore, conventional volume rendering methods such as ray marching with dense sampling are often impractical. While resampling, caching, and retraining can mitigate this cost, they compromise convenience and accuracy and become impractical for time-varying data. We tackle these challenges using a query-efficient stochastic volume rendering framework based on delta tracking. Our system employs a four-stage pipeline that exploits heterogeneous parallelism, using ray tracing cores for traversal and tensor cores for batched neural evaluation. Furthermore, we present strategies to reduce INR queries via ray budgeting and query pruning, thereby increasing per-frame performance. Using our renderer, many time-varying INRs can be rendered directly from their original representation. The system achieves ~30-40 FPS at 1024x1024 resolution on an RTX 4090 GPU and converges to high-fidelity images. Moreover, the system enables interactive temporal exploration of the continuous domain, with timestep updates taking approximately 1-2 ms.