生成模型与LLM推理优化

Large language models represent context with a key-value (KV) cache. Caches are model-specific: for the same text, models with different architectures or weights produce incompatible representations. This makes it costly to switch models over a shared context: although the context has already been processed by one model, the incoming model must process it again to build its own cache. We introduce KV-Lingo, a method for translating the KV cache of a source model into one that can be read by a target model. KV-Lingo consists of a collection of linear maps, typically one per layer of the target model, that are applied independently on all tokens' key and value representations. We train these maps using distillation, minimising the divergence between the target model's predictions from its native cache and those from the translated cache. We consider several model pairs spanning multiple sizes and architectures, training one translator per pair on a generic text corpus. The resulting translators preserve strong downstream performance in both small-to-large and large-to-small transfers. Since a switch then costs a linear map and a single decoding step instead of a prefill, replacing re-prefill with cache translation reduces the time to first token after a model switch by 9.6x already on a 64-token prompt for Qwen models on an Apple M3 Ultra, and by up to 29x at 32k context length on an H100. These gains make KV-Lingo particularly useful for dynamic model routing: a context can be processed by one model and handed off to another only when needed, without re-prefilling the shared prefix. We finally show that KV-Lingo can be used for seamless model switching, staying close to re-prefill across repeated switches in our multi-turn evaluations.

Smaller activation outliers do not necessarily imply better low-bit quantization: their alignment with the quantizer matters. We introduce PrismQuant, a quantizer-aware rotation framework that aligns the leading activation eigenspace with the constant group subspace of asymmetric grouped INT4. The affine offsets represent the energy in this subspace without widening the range within the group. We formulate rotation design as a Ky Fan trace maximization and derive a closed-form solution that is provably optimal for this alignment objective. Compact Householder transformations and their compact-WY representation enable gradient-free construction and efficient application at both foldable and online sites. A predictive range law further connects unaligned activation energy and group size to quantization-relevant variation. Experiments on Llama, Qwen, and Mistral span dense models up to 70B parameters and a 30B mixture-of-experts model. Under W4A4KV4, PrismQuant sets the state of the art on Llama-3.2-3B among the compared methods in both perplexity and accuracy. On Llama-3.1-70B, it attains 3.85 perplexity and 72.46% average zero-shot accuracy, only 0.22 percentage points below full precision. In the deployment study on Llama-3.1-8B, our optimized implementation achieves 1.51x prefill and 1.22x CUDA Graph decode speedups over matched FP16 baselines, with 56.34% lower decode peak memory and only 2.35% additional Graph decode latency over Hadamard. Code is available at https://github.com/ForeverBlue816/PrismQuant.

The KV cache of a transformer entangles every token's representation with its entire prefix: a passage encoded once cannot be reused under a different prefix or removed without recomputing everything after it, so exact cache reuse is limited to shared prefixes. We present SMem, an architecture whose context representation is a cache by construction. A block-local encoder maps each block to memory rows independently of other blocks, and a reader conditions generation on their union through cross-attention. For every parameter setting, memory composes exactly at fixed block indices, deleting a block is an exact $O(b)$ update for $b$-token blocks, and the memory state is independent of the edit path. At $4\times$ the training context, under the shared recipe, SMem retrieves planted needles beyond any trained-length window (exact match 0.14-0.28 at distances of 31 and 63 blocks), where learned-position, RoPE, and Block-Attention-style transformers all score at most 0.02. A fully cached context is served by computing one block alone at a near-constant 3.1-6.2 ms, whereas cold prefill grows with context; batched decode stores 34-38% fewer KV rows and runs 1.4-1.7$\times$ faster when bandwidth-bound; and deletion beats suffix recomputation by 8.5$\times$ at 512 blocks and 452$\times$ at 4096 blocks (32-256$\times$ the trained length, probing the cost model rather than a served regime). The cost is a perplexity gap of -4.7% to +2.8% (negative favors SMem) against a parameter-matched transformer with the same positional scheme, at 160M-1.5B on FineWeb-Edu across two recipes and a learning-rate search. SMem also composes with RoPE: at 160M and 410M the composite matches or leads the matched transformer and closes 29-59% of SMem's gap to a RoPE transformer. Dropping prefix entanglement thus keeps perplexity comparable while making the cache exactly composable and editable.

The memory usage and decoding latency of LLM inference grow rapidly with context length. To reduce these costs, key-value (KV) cache compression methods selectively retain cached states based on token importance or differences in attention patterns across heads. However, we discover that retrieval capability varies substantially with relative distance, even within the same attention head. To exploit this structure, we introduce Distance-KV, which learns a static KV retention pattern over the joint space of layers, attention heads, and relative distances. The pattern is learned offline with the language model frozen and reused across inputs to prune and compact the KV cache without online importance scoring. Across three backbone models and four long-context benchmarks, Distance-KV consistently achieves the best overall performance among competing KV cache compression methods, exceeding the strongest compression baseline by up to 9.3 points on RULER at 128K. On Llama-3.1-8B-Instruct at 128K, Distance-KV reduces KV cache memory by 65.4% and achieves a $1.66\times$ decoding speedup relative to Dense. Together, these results identify relative distance as an important structural dimension for understanding how LLMs retrieve information over long contexts and for designing more efficient inference methods.

LLM-based agents execute long-horizon tasks through repeated model calls interleaved with tool execution and user interaction. As each call extends the history accumulated in previous turns, prefix caching avoids repeated prefill of the agent's entire context. However, the aggregate cache footprint grows with context length and concurrency, forcing serving systems to reclaim cached KV tensors. We identify recency synchronization: accesses to an agent's cached history refresh its cache nodes together, causing least recently used (LRU) eviction to concentrate on a few agents' private histories. When a fully evicted agent returns, it must recompute nearly its entire accumulated context, producing a large time-to-first-token (TTFT) outlier even if most other requests retain substantial cache reuse. We present RR-EVICT, a fine-grained prefix-cache eviction strategy that distributes reclamation across idle agent trajectories. RR-EVICT visits agents in round-robin order and evicts a tail chunk from each, preserving reusable prefixes for more agents when capacity remains for idle state. Returning agents can reuse these partial histories and recompute smaller missing suffixes. The policy requires no prediction of future arrivals or tool latency. We implement RR-EVICT in SGLang and evaluate conversational and coding-agent workloads under colocated and prefill-decode-disaggregated serving. Compared with LRU, RR-EVICT reduces P99 TTFT by up to 75.4% and P99 uncached prompt tokens by up to 65.7%.

RAG and retrieval-based agent memory both inject retrieved content into LLM prompts, as document chunks and recalled memory records, respectively. The same content can recur across requests at different prompt positions or after different preceding contexts, preventing reuse through conventional prefix caching. Our characterization finds that records recurring outside the matching prefix account for over 70% of injected memory tokens in agent-memory workloads. Composable KV-reuse methods enable reuse in such cases, but online serving introduces a management problem: a recurring unit's KV states may not yet exist, may have been evicted, or may reside on another node. We present Tessera, a disaggregated serving system that makes retrieval the control plane for KV reuse. By exposing the context units needed before model execution, retrieval allows Tessera to combine current demand with retrieval history, KV residency, and generation load to coordinate cache management and request routing. Generation nodes concurrently prepare locally cached, remotely cached, and missing states, while retaining newly computed states off the request's critical path. Across RAG and agent-memory workloads, Tessera lowers mean TTFT by up to 3.6x over SGLang and LMCache with EPIC at matched request rates, and sustains low TTFT at rates where the baselines saturate, while matching the answer quality of the underlying composition policy.

The widely used Transolver family of neural operators is based on physics-attention, which softly assigns the points of an unstructured mesh to a small number of slices, applies self-attention among the resulting tokens, and broadcasts the result back to the points. We provide a comprehensive empirical and theoretical analysis to elucidate the mechanisms which are responsible for model performance. To this end, we perform careful ablations on a challenging suite of nine 3D fluid dynamics benchmarks to find that replacing token attention with a constant linear map does not affect the accuracy. Thus, Transolver does not need a Transformer at all. However, removing the global mixing (slicing/deslicing) or doing it only once leads to performance collapse. We leverage the theory of averaging neural operators to explain and corroborate our findings by showing that just slicing/deslicing, in conjunction with pointwise MLPs, already suffices for universal approximation of continuous operators and attention is redundant in this context. Finally, we provide a novel FlashAttention-style efficient implementation of the key slicing/deslicing module of Transolver. This flashslice kernel streams slice and deslice over the points without materializing the heavy slice-weight tensor, while reproducing the best available implementation to floating point error. At the same time, it leads to very significant memory and compute savings, particularly at large slice counts.

KV cache eviction methods such as SnapKV and PyramidKV rank tokens solely by mean attention over a small observation window. We study a unified score, $μ_i+λ_1σ_i+λ_2\mathrm{corr}(i,S)$, adding attention dispersion across window queries and redundancy relative to selected tokens. For $λ_2<0$, the score penalizes similarity to selected tokens as in maximal marginal relevance (MMR), without extra forward passes. To test whether this relevance-diversity balance should vary with depth, we compare fixed global coefficients with three-segment and quadratic profiles. Only these depth profiles are searched on a development split under a $\sinh$ reparameterization. On all 16 English LongBench datasets with Mistral-7B at a budget of 64 entries per layer, a single global diversification constant improves 13 of 16 datasets (macro +1.1); the gain holds at budget 32 and narrows at 128. Per-dataset search finds no detectable layer structure on most datasets; on passage retrieval it finds a large one: a mid-layer sign flip that rewards similarity and is worth +9.6 over the baseline at budget 64 and, without re-tuning, +13.2 over the global constant at budget 128. Ablations attribute the gain to the redundancy term; replaying every accepted search state on the held-out test set separates genuine structure from tuning noise.

Position-independent KV cache reuse aims to reduce latency in retrieval-augmented generation by reusing chunk-level KV caches across prompts. We show that current evaluations of KV cache reuse techniques rely on measurements that fail to faithfully capture the loss of accuracy attributable to reuse, often artificially inflating the reported effectiveness. We also show that existing datasets do not exhibit the reuse dynamics needed to thoroughly evaluate such techniques. To address these issues, we propose an evaluation methodology that measures this accuracy loss without ambiguity and we introduce Boxoffice, a tool that programmatically generates evaluation datasets that exercise challenging KV cache reuse patterns.

Low-Earth-orbit (LEO) satellites are increasingly expected to perform onboard learning for applications such as disaster response and environmental monitoring. However, conventional federated learning (FL) is ill-suited to onboard satellite learning, as it assumes computational, memory, and communication resources beyond the capabilities of resource-constrained LEO platforms, often necessitating the transmission of raw imagery to ground stations. We present COSMIC-FL, a resource-aware FL framework for efficient onboard learning in LEO satellite constellations. COSMIC-FL introduces two complementary Mixture-of-Experts (MoE) architectures: a Sliced design that shares backbone representations while activating task-specific channel subsets, and a Modular design that employs lightweight gating to route inputs to physically separated expert networks. A semantic class-to-expert mapping enables each satellite to train, update, and communicate only the expert paths relevant to its local data. To further improve efficiency, COSMIC-FL integrates staged optimization with three structured pruning strategies: server-side pruning, client-side fixed-ratio pruning with mean-vote aggregation, and adaptive client-side per-layer pruning based on aggregated importance and a MAD-based gap criterion. Combined with semantic expert routing, these techniques jointly adapt computation and model sparsity to both data semantics and layer importance, yielding a favourable accuracy--efficiency trade-off for heterogeneous space platforms. Experiments on six image classification benchmarks under highly non-i.i.d. settings show that COSMIC-FL maintains competitive accuracy while reducing communication, computation, and energy consumption by up to 80% over SOTA FL methods. We further validate COSMIC-FL on an NVIDIA Jetson AGX Orin, confirming its efficiency gains under realistic embedded deployment constraints.

Generative world models can provide visual rollouts for embodied planning, yet their feasibility on edge devices depends not only on the learned model but also on how the execution runtime represents its operations. We introduce a cache-aware lowering that expresses supported causal Conv3D calls as batched spatial Conv2D operations while preserving pretrained weights, temporal-cache semantics, convolution parameters, bias placement, and output layout. Across the complete Cosmos3-Edge image-to-video pipeline on a 64-GB NVIDIA Jetson AGX Orin, the proposed route accelerates VAE decoding by approximately $7\times$ and reduces complete-generation latency by more than $2\times$, while repeated decoder evaluations maintain complete fast-path coverage without fallbacks. The unchanged lowering also improves Cosmos3-Nano and transfers to LingBot-World's architecturally distinct Wan2.1 VAE. A clean-device comparison against fully specialized TensorRT shows that TensorRT provides a further $1.36\times$ steady-state improvement, but requires substantially greater per-module and per-runtime-state AOT specialization. Same-latent BF16 and FP32 evaluations characterize the finite-precision differences introduced by the alternative execution order. Together, these results position cache-aware lowering as a lightweight runtime optimization that recovers most of the available decoder acceleration without modifying the learned models themselves.

Activation probes are increasingly used to monitor LLMs in deployment. A probe is typically trained under one inference configuration, then used under whatever batch size and numerical precision the serving stack uses. Because common GPU kernels are not batch-invariant and floating-point formats round differently, the activations seen at deployment are not the ones the probe was trained on. We measure what that costs for Llama-3.1-8B, Qwen3-8B and Gemma-3-4B across batch sizes 4, 8 and 16 and float32, bfloat16 and float16, training 768 probes on one configuration, evaluating each on every other, and comparing verdicts example by example. Probes are stable, but aggregate accuracy is the wrong instrument for showing it: it understates how many verdicts change by a factor of two to nine. At the prompt, accuracy never moves by more than 0.47 percentage points across 1,392 transfers and only 0.076% of verdicts change; under float32 with only the batch size varied, none of 201,960 verdicts change. During decoding the flip rate rises to 2.8%, but rows whose realised tokens matched flip in only 0.12-0.15% of cases, while rows whose tokens diverged flip in 12.9%: the cause is the text, not the arithmetic. A bfloat16 batch-size change flips the first generated token for 2.1% of rows and leaves 25% on different tokens by token 20. Flips are symmetric, Cohen's kappa stays above 0.94, and AUROC moves by at most 0.05 points. Underneath, activations move about as much as the format's rounding: a bfloat16 batch-size change perturbs them by a median relative L2 of 1e-2, roughly 8x the float16 figure. Probes absorb this; the model's own next-token argmax does not. Robustness evaluations of activation monitors should report per-example agreement rather than aggregate accuracy, separate representational noise from input change, and state the serving configuration.

How many bits does a network need before its accuracy collapses, and how does this grow with depth? We study the precision floor, the perturbation level or bit-width at which accuracy falls halfway to chance, in MLPs, CNNs, Vision Transformers and nine pretrained language models, under post-training quantization (PTQ) and quantization- or noise-aware training (QAT). (i) A first-order theory sets the floor through one full-precision quantity, the predictive amplification $G$: $η_c=Λ/G$, and $G^2$ grows linearly in depth at a rate proportional to the squared residual branch scale. (ii) The predicted equality $α_{PTQ}=ρ$ of depth exponents holds within 95% intervals in twelve of thirteen trained architectures and in GPT-2 from 12 to 48 layers, with $Λ=1.45\pm14%$ across trained architectures. (iii) Residual branches scaled by $1/\sqrt{D}$ and pre-normalisation remove the depth penalty, and each quantizer turns noise into bits at a rate fixed by its step rule, giving $b_c=(α/γ)\log_2 D+C$. (iv) A QAT paradox: noise-aware training roughly doubles the tolerable noise of shallow networks, but the gain decays with depth, so the depth law steepens ($α_{QAT}/α_{PTQ}=1.45$-$1.47$ on two datasets, ten seeds each). Decision margins, cross-layer error cancellation and heavy tails do not set the floor.

Multi-agent systems distribute large contexts across agents that communicate to solve a task. Text messages are compact but require decoding and may omit evidence the receiving agent needs. Recent latent communication instead transfers KV caches. This avoids text generation and can improve accuracy and latency. However, a full KV cache grows linearly with both the context an individual agent processes, and the number of agents that coordinate together. This raises memory and context costs, often far exceeding available GPU resources and context window sizes. Our key observation is that agents need only send what the receiving agent requires for its local task — which we call receiver-conditioned communication. The receiver agent passes the sender a small description of its information needs, which serves to filter and compress the sender agent's KV cache. CacheBack is a simple, robust, training-free instance of receiver conditioning based on the sender's attention weights. On FanOutQA, CacheBack with Qwen 3 removes 75% of the state the agent would otherwise receive, improving accuracy by 14.7 percentage points and reducing median task-completion latency by 3.2x relative to text communication. We show comparable improvements across model families that span dense Transformers, Mamba-attention hybrids, and sliding-window attention.

In this paper, we propose I-Parakeet, an integer-only implementation of NVIDIA's Parakeet-CTC (0.6B parameters) that runs on a smartphone NPU without any floating-point operator or CPU fallback. Modern Conformer ASR models are hard to deploy on edge devices because of their size, and quantized models still fall back to floating point for numerically sensitive operations. This prevents them from fully exploiting integer accelerators such as mobile NPUs. To achieve this, our contributions are threefold. First, we derive an integer formulation of the relative-positional self-attention at the core of the Conformer. We fuse its two score branches with different quantization scales and the relative shift into integer-only operations. Second, we introduce a minimax-optimized Swish approximation that minimizes the maximum error of the Swish output. Third, a layer-wise range analysis of activations yields two targeted remedies: an INT16 grid for the BatchNorm output and percentile calibration for the heavy-tailed pre-encoder activations. I-Parakeet achieves 4.97% WER on LibriSpeech test-other, running on a Qualcomm NPU at a real-time factor of 0.048, 7.5x faster than a CPU baseline.

Quantum-inspired evolutionary optimization (QIEO) is a new class of population-based metaheuristic optimization algorithms which represents design variables as a set of qubits and searches a continuous, multi-dimensional landscape through rotation of the qubit's amplitude pair. Every generation rotates those amplitudes toward a single elite, which corresponds to that generation's best. The per-generation cost scales as $O(N_p N_g)$ for $N_p$ chromosomes and $N_g$ genes (decision variables). Production use of such solvers is rarely confined to a single machine class. Prototypes are run on laboratory servers- before moving to rented cloud workstations for more involved campaigns. The largest problems are reserved for leadership-class accelerators. This paper asks whether a single OpenMP~5 source of QIEO, offloaded with \#pragma omp target, is a viable production path in each of those settings. We report three independent, campaigns of the 0/1 knapsack problem against a same-source multi-core Intel CPU baseline. The study comprises approximately 3,000 runs spanning varying chromosome and gene counts, evaluated using both chromosome-level and gene-level offload strategies on the NVIDIA Tesla V100 SXM2, NVIDIA A100 80GB, and AMD Instinct MI300X GPUs. Deployment-specific nuances such as Volta's constant-memory cliffs, Ampere's L2 persistence and cp.async, CDNA~3's Infinity Cache and XCD occupancy are addressed to ensure high performance of these platforms. Results reveal gene-parallel offload achieved geometric-mean speedups of 90$\times$, 136$\times$, and 155$\times$ over a single CPU core on the V100, A100, and MI300X, respectively, and 12$\times$, 17$\times$, and 16.6$\times$ over 72 host threads. Furthermore DetermineElite, the $O(N_p)$ selection of the generation-best chromosome, is found to be better suited to the host than to the device.

Text-based memory and context compression support reuse of past interactions. Resizing continuous memory changes the input to a frozen LLM, coupling capacity allocation with readout. We propose Hard-Origin Adaptively Softened Memory (HasMem). Frozen hard-prompt embeddings provide a verifiable initial state. A controller adjusts memory widths, a Writer re-encodes resized entries, and Reader and Global provide readout adaptation and cross-turn state. On all $535$ questions in a reconstruction probe derived from the Multi-Session Chat (MSC) development split, the main configuration achieves lexical F1 of $95.3$ ($+4.4$ percentage points) at $93.6%$ of the hard reference's framed memory positions. With approximately matched per-question target body budgets, six configurations at mean per-entry retention around $0.83$--$0.91$ exceed rule-based re-encoding by $8.0$--$23.6$ exact-match (EM) percentage points. With fixed model parameters and rule target width ratio $0.75$, Global's EM gain passes a user-level exact paired test with Bonferroni correction over eight comparisons. On all $500$ LongMemEval-S questions, local lexical F1 rises from the hard reference's $3.4$ to $8.9$, and answer negative log-likelihood (NLL) falls from $12.257$ to $5.274$. F1 gains accompany lower EM on both evaluations.

Hierarchical vision-language-action (VLA) systems consist of a high-level vision-language planner and a low-level action expert that generates continuous actions. This hierarchical design has practical value only if the planner can generate plans fast enough to meet real-time control requirements, and the resulting plans actually contribute to the generation of action. We study one such system, a waypoint hierarchy pipeline adapted from $π_{0.5}$, and find that neither requirement is satisfied. This baseline relies on token-level autoregressive decoding (Token-AR) to generate a waypoint plan, requiring 57 very expensive vision-language model (VLM) forward passes. However, we find that erasing the waypoint endpoints has little effect on task success. Two findings reveal the misalignment of planner-executor: the planner generates outputs at an excessively fine granularity, and the executor underuses plans as a control condition. We address the latency issue with waypoint-aligned block-autoregressive decoding (Block-AR), and plan underuse issue with normalized goal modulation (NGM), a layer-wise goal path constrained by phase gating and anti-shortcut training so that the waypoint influences action generation maintaining other signals. Our method reduces the maximum number of VLM forward passes from 57 to 8 on LIBERO, including one prefix prefill, and achieves an $8.7\times$ reduction in planning latency on a Rokae dual-arm robot. With normalized goal modulation and anti-shortcut training, Block-AR's success rate on LIBERO-Long increases from 91.0% to 96.2%, while its average success rate across the four suites increases from 95.85% to 98.45%. On three bimanual tasks with this robot, success rates remain comparable across methods.

Large language models (LLMs) have demonstrated promising performance in log anomaly detection, yet how their adaptation strategies, architectures, and deployment configurations affect detection effectiveness remains insufficiently understood. To investigate these factors, we conduct a systematic empirical analysis across three public log datasets, examining different adaptation strategies, model architectures, parameter scales, and quantization settings. Our results reveal substantial performance differences across adaptation strategies, while model scaling yields varying detection gains across datasets. We further observe that models with comparable detection accuracy can exhibit markedly different computational costs, and that low-bit quantization largely preserves detection performance in the evaluated configurations. Finally, we examine detection robustness under structural, semantic, and label noise at different perturbation levels. These findings provide empirical insights into the performance, efficiency, and robustness of LLM-based log anomaly detection, highlighting practical considerations beyond conventional accuracy-oriented evaluation.

Reliable voice interaction is essential in environments with limited internet connectivity and strong privacy. However, most existing voice assistants depend on cloud-based services, which leads to latency issues, dependency on internet access, and privacy vulnerabilities. This research presents LUMO (Lightweight Unified Multilingual Orchestrator), a privacy preserving offline voice assistant designed for edge computing environments. This system integrates local Automatic Speech Recognition (ASR), locally deployed quantized Large Language Model (LLM), and Text-to-Speech (TTS) synthesis into a fully offline pipeline running on a Raspberry Pi 5 with 8 GB RAM. To enable efficient operation on resource constrained hardware, the language model is compressed using 4-bit GGUF quantization, which reduces memory usage while preserving practical conversational capability. Existing edge based voice assistants Mycroft provides partial offline functionality without a generative LLM, with an approximate latency of ~5 s and power consumption of ~12 W, while Rhasspy supports full offline operation but lacks generative capabilities, with ~3 s latency and ~11 W power usage. In contrast, LUMO achieves a Word Error Rate (WER) of 6.8% for short English utterances in low noise conditions, an end-to-end response latency of 2.0-4.0 s, and a lower peak power consumption of approximately 9.0 W. The system also achieves effective offline recognition for Bangla speech, supporting multilingual accessibility in low resource settings. By operating entirely offline, LUMO provides strong data privacy, reduced need for cloud connectivity, and suitability for privacy sensitive edge execution such as rural healthcare, education, and disaster response scenarios.

Hybrid language models maintain fixed-size recurrent states, but existing quantizers typically use eight bits or more. Quantization errors persist according to channel decay rates. We derive distortion weights from the observability Gramian and combine them with normalized state ranges for mixed-precision bit allocation, without calibration data, rotation, or training. We also quantize decay rates logarithmically. With per-token state quantization, a four-bit mean payload reduces excess negative log-likelihood by factors of 3.3--27.9 relative to the best of seven baselines across three hybrid models; metadata costs vary. At six bits, negative log-likelihood differs from the FP32-state baseline by less than 0.005 nats. Ablations separate gains from variable bit widths, decay weighting, and range normalization. With less frequent write-backs, gains diminish and depend on the model and budget.

Tool observations dominate the context of software-engineering agents, making long interaction histories costly to maintain. Existing context compression methods can discard information needed by later actions, while adapting agents to soft-token representations can compromise their original behavior. To reduce context while preserving action-critical information and agent behavior, we combine Latent Observations, Hard Actions (LOHA), a context layout that separates compressed history from text needed for exact reference, with Anchored Context Distillation (ACD), a training method that enables latent reading while constraining behavioral drift. LOHA compresses older tool observations into soft tokens while retaining the agent's own turns and the last K observations in text, providing compact access to historical information and exact access to recent content. To enable the agent to use this representation, ACD distills the base model's full-text predictions into the latent view while anchoring its behavior on plain-text inputs to the same base model. On SWE-bench Verified, K=3 reduces context per call by 43% for Qwen3-4B and 57% for SWE-Master-4B-RL, with resolve rates of 12.1% and 21.8% versus 14.5% and 27.5% for their uncompressed bases. A single-run recency sweep reaches 14.4% and 23.0% at K=8, with larger windows generally favoring task performance over compression. Under a 32K-token limit, Qwen3 with K=3 resolves 21.1% of a 199-instance subset versus 11.1% for the same adapted agent using full text. In concurrent single-GPU serving, it achieves 1.9 times that full-text agent's instance throughput.

Whether an information extraction pipeline should process page images or parsed text depends on the document, and the answer flips across the layout spectrum. We study this trade-off under a constraint that rules out (closed) cloud services: privacy-sensitive documents processed on-premise by small ($\le 8\mathrm{B}$ parameter) text-only and vision--language models, evaluated on both accuracy and energy over a design space spanning input representation, model family, and inference configuration. Benchmarking on the near-plain-text Kleister-NDA contracts and the layout-rich VRDU forms, we find that batching is the dominant energy lever, cutting energy per page by 38-85% at no cost in accuracy, while FP8 quantization saves 27-32% when requests are served one at a time but less than 1mWh per page (9-19%) once batching is applied. Preprocessing dominates what remains: neural OCR costs $17\times$ more energy per page than classical OCR and never reaches the Pareto frontier. Which representation wins flips with the type of document: vision--language models on layout-rich documents and small text-only models with a cheap parser on near-plain text, where they are both more accurate and cheaper than any vision--language configuration. Our work yields concrete guidelines for energy-efficient, privacy-compliant local information extraction.

Cross-domain few-shot learning requires adapting a classifier to a new visual domain from very few labelled examples without target-time parameter updates. We isolate one question: under a fixed global representation, what does joint query-support adaptation contribute to prototype construction? The Within-Instance Prototypical Transformer (WIPT) implements single-query test-time prototype adaptation by jointly transforming one unlabelled query and the labelled support embeddings, then forming query-specific class means. Using a shared frozen ViT-S/16 encoder, miniImageNet source training, and CUB, EuroSAT and ISIC targets, we replicate the key comparisons across five independent training seeds. In 1-shot evaluation, WIPT improves frozen ProtoNet in every run on CUB (+0.21 percentage points) and EuroSAT (+2.07), but decreases ISIC (-0.22). In 5-shot evaluation, ProtoNet remains strongest overall, while WIPT consistently improves a capacity-matched support-only Transformer on ISIC (+0.99). Joint processing of up to five queries yields no reliable accuracy gain; in a head-only 5-shot benchmark, g = 5 reduces analytical attention-token pairs by 73% and peak allocated memory by 29% relative to g = 1, although latency is non-monotonic. Across all target/shot conditions, WIPT changes uncertain ProtoNet decisions far more than confident ones, and rescue/break decomposition accounts for the observed gains and losses. Source-shift and scorer controls further show that the benefit is not universal. Overall, WIPT provides a streaming-compatible form of test-time prototype adaptation that can improve difficult low-shot cross-domain decisions without target-time optimization.

Large vocabularies make output heads a substantial inference cost in small language models. We introduce softmax reparameterization, a post-training method that searches over functionally equivalent output heads before quantization. The method subtracts a scalar multiple of the vocabulary-row mean from every output row and selects the coefficient by validation KL. For linear-softmax heads, these shifts preserve full-precision predictions exactly and require no decoder retraining; a rank-one correction extends the construction to nonlinear logit paths. Across seven output heads and three quantizers, W4 gains are largest where baseline quantization substantially distorts predictions: test KL falls by 93% on XGLM under RTN and by 73--77% on Phi, BLOOM, and BLOOMZ under activation-weighted MSE. Heads with low baseline error change little; at W2, used as a compression stress test, benefits extend more broadly. On Phi, the gains persist under stronger GPTQ calibration; a separate untouched holdout reproduces the improvements on Phi and BLOOM. Frozen WikiText-selected coefficients also transfer without retuning to C4 and OpenWebMath. Residual analysis on Phi shows how fidelity can improve despite greater total logit error: the selected representative reduces error on likely outputs and lowers its Fisher-weighted cost. For shift-compatible heads, the shift adds no inference operation. With the decoder held in BF16, a packed W4 Phi output head reduces batch-one generation latency by 10.8%, and reparameterization preserves this speedup.

Full-duplex speech language models continuously accumulate acoustic key-value (KV) states, making long-running interactions memory-intensive. During listening, the model can finish processing an audio unit before the next arrives; we term the remaining interval listening-time slack. We propose acoustic-to-text KV compression, which introduces a transcription side channel to convert incoming speech into compact textual memory within this interval. When the cache exceeds a target budget during inference, older acoustic states are evicted while transcripts and recent acoustic context remain. We train the side channel with LoRA using cross-entropy on transcription segments. To preserve listening and speaking behavior, we apply knowledge distillation to the original model's token-level output distributions at native prediction positions. On ten-minute LongSpeech sessions, our MiniCPM-o 4.5 implementation reduces peak streaming KV-cache size by 64.6% compared with the same model without eviction. The proposed method also improves transcription, temporal question answering, and summarization over the baseline. Full-Duplex-Bench evaluations further show comparable pause-handling, turn-taking, and interruption performance.

Tabular in-context learners such as TabPFN, Mitra, or ConTextTab rely on alternating row and column attention over 2D sequences of latent embeddings. These attention patterns differ markedly from the one-dimensional case in language models: row attention involves longer sequences while column attention operates on much shorter ones, and the strided memory layout of tabular data makes producing contiguous tensors costly. Moreover, the hidden dimensions used in current models are small compared to recent language models. Yet efficient attention has been studied mostly for one-dimensional sequences, leaving the two-dimensional tabular setting unexplored. To this end, we create a reproducible benchmarking setup and study the unique characteristics of tabular attention across several backends — Torch SDPA (efficient and cuDNN), FlashAttention-2/3/4, and the inference-only backends vLLM and SageAttention — measuring forward and backward throughput across realistic tabular shapes on three GPU generations (A100, H100, B200). We find that the optimal backend choice differs between column and row attention and varies across hardware as well as model specifics: While the FlashAttention implementations tailored for each GPU generation perform overall best, they are at times outperformed by CuDNN in the case of column attention at longer sequences with cross-over points depending on the head dimension. Among inference-only backends, SageAttention performs well for row attention and large sequences beyond 16\,k rows. Our reproducible benchmark lays the foundation for future improvements to table-native attention. The self-contained benchmarking and evaluation code is openly available at: https://github.com/SAP-samples/tabular-attention-benchmark

Agentic LLM workflows decide their execution paths at runtime. Downstream computation may be predictable, or may have run before, yet it cannot begin until the model or the user resolves the branch. We call this serialization the branch-resolution barrier. Caching alone does not hide it: the key that identifies a reusable result is not known until then. In this paper, we propose DynBranch, which makes an unresolved branch addressable before it resolves. Its stable coordinate lets candidate subgraphs run during resolution and completed subgraph results be reused across later requests. A two-level controller admits this work when its expected benefit exceeds the load price. DynBranch sits at the model-API boundary and requires no changes to agent harnesses or model execution engines. Across four agentic workloads with Qwen3-32B on 4x H200 GPUs, DynBranch reduces mean latency by up to 32% over each workload's strongest prior system and by 46-66% against a no-reuse floor, while preserving workflow results. The benefit persists across backbone families and on a commodity Qwen3-8B/RTX 4090 deployment.

Looped transformers reuse weights across recurrence steps, making low-bit quantization especially attractive. We identify two distinct failure modes of standard post-training quantization. On Huginn-3.5B, per-channel INT4 fails primarily at the non-residual loop-entry adapter, while quantizing the residual core is much less damaging. We call this feedback exposure: a quantized layer perturbs the recurrent state without an identity path, and the resulting error is fed back at later steps. Controlled experiments on linear filters and Mamba state-space models show that feedback exposure also occurs outside transformers. Grouped INT4 reveals a separate failure, calibration blindness: our one-step GPTQ baseline builds its Hessian from step-0 activations, leaving input directions used later in the recurrence nearly unweighted. Across nine checkpoints from seven looped architectures, one-step GPTQ is worse than round-to-nearest (RTN) on the primary task metric for five checkpoints. Accumulating the GPTQ Hessian across recurrence steps outperforms both one-step GPTQ and RTN on all nine checkpoints and recovers bf16-level accuracy on Huginn. These results separate two questions for PTQ on looped models: where quantization error enters the recurrence, and which states calibration sees.

Disaggregating the two stages, Prefill and Decode, onto separate GPU pools is now a standard optimization for (text-only) LLM serving. However, multimodal LLMs (MLLMs), which add a third phase, Encode, pose new challenges for resource allocation. Encode turns images, video, or audio into embeddings that the language model can consume, yielding a three-stage Encode-Prefill-Decode (EPD) pipeline. Existing frameworks offer only partial answers: text-only PD systems lack Encode, while EPD frameworks expose it as a separate service without regulating downstream request flow. The pipeline also carries a structural resource imbalance: every request enters through Encode before downstream work can begin, yet per-request execution leaves the encode GPU severely underutilized even at high loads, starving the downstream Prefill and Decode workers. Addressing this, we reposition Encode as the control point of the EPD pipeline, exposing three tightly coupled dimensions: when work enters downstream, where prefill executes, and how the GPU is shared. We instantiate this in EAServe across two co-designed layers. Its runtime manages load-adaptive micro-batching, rate-controlled partial offload to a co-resident prefill worker, and dynamic SM partitioning for predictable co-location. The configuration layer, Hybrid Auto Selection (HAS), navigates the joint space of GPU allocation, encode batch size, and offload ratio by pruning unbalanced allocations with per-stage capacity profiling and refining the remainder through TPE-based Bayesian optimization. Evaluated on three MLLM architectures spanning image, video, and audio, EAServe delivers up to 4.3x and 1.7x higher goodput than NVIDIA Dynamo and vLLM, respectively, under identical SLO constraints, sustains more balanced and higher GPU utilization across the EPD pipeline, and reaches near-optimal configurations faster than baseline search methods.