生成模型与LLM推理优化

Benchmarks for vision-language models offer their answer choices in some convention: a letter, a color name, a pixel coordinate. That convention is treated as neutral. We find it is not, and that the limits a benchmark reports can belong to the readout rather than to the model. On 200 COCO photographs, Qwen3-VL-4B picks the correct one of nine locations for a named object 68.5% of the time when the locations are given in English and 20.0% when the same locations are given as pixel coordinates. Chance is 11.1%. The cost arises when the answer options are coordinates; giving the model a coordinate in the question instead costs 3.5 points and is not significant. The gap holds on a 4x4 grid, under 8-bit rather than 4-bit quantization, and in every slice by object size, boundary distance and category. It also decides which model wins. Two models that tie under English names differ by 39 points in one coordinate system and by 54 in the other, in opposite directions. On the color task, three of the four open models capable of the task show the penalty; on photographs, two of three open models do, and so does Gemini, at 11.1 points on parseable answers (p = 1e-4). GPT-4o does not. To ask whether a model reads a coordinate at all, we attach the wrong name to each one and record which the model follows. Color options written as hue angles are followed below chance; a normalized pixel convention is followed at four times chance. This tells apart conventions a model can use from ones it cannot, though it did not predict accuracy on two untried conventions. Five models also name the same color wheel five different ways, so a fixed answer vocabulary is not neutral across models either. Five times during this work we measured a capable model as incapable because our scorer and the model disagreed about what an answer looks like. We report each case. They are the phenomenon in miniature.

KV-cache eviction is typically evaluated through average quality-memory trade-offs, yet a small average loss can hide requests whose utility degrades materially. We reformulate eviction as a deployment risk-control problem: a material degradation occurs when eviction lowers task utility by more than a deployment-specified tolerance relative to full-KV inference on the same request, and deployment risk is the population frequency of such events. Given a reliability contract specifying a target risk level and confidence requirement, we use a compressor-agnostic post-hoc certification procedure to select a retention policy from calibration data with a finite-sample guarantee, falling back to full KV when no compressed policy is certified. Across multiple eviction methods, Llama and Mistral models, and LongBench and RULER-32K, the same contract supports substantially different levels of eviction: on Llama, it certifies SnapKV at 75% retention on LongBench but no tested compressed policy on RULER-32K, triggering full-KV fallback. Policies with empirical degradation rates below the 5% target can still fail finite-sample certification; on Llama LongBench, empirical thresholding selects uncertified policies that retain 5-10 percentage points less cache across fixed-budget methods. The proposed framework converts a deployment-level reliability requirement into a KV-memory operating point.

An estimated 1 billion people worldwide live with vision impairment, yet current vision-language models (VLMs) produce descriptions too vague for safe navigation by blind and low-vision (BLV) users. Large VLMs can generate high-quality audio-description-compliant narrations but cannot run on mobile devices; small VLMs offer competitive latency but lack spatial detail, directional cues, and hazard awareness for navigational assistance. We present Smol-VL-BLV, a compact VLM for blind and low-vision users that closes this gap using a 500M decoder transformer model and two post-training mechanisms: (1) teacher-student distillation and (2) Group Relative Policy Optimization (GRPO) with a composite BLV reward targeting directional language, metric distances, and hazard detection. Because multi-stage post-training can induce catastrophic forgetting, we add a lightweight finetuning stage after the last stage GRPO finetuning to recover general descriptive quality while preserving BLV-specific spatial grounding. Our best model substantially outperforms the baseline across various benchmarks, including tasks: VQA, BLV captioning, OCR, and latency. Compared with the baseline for relative improvement, it improves the Spatial score gain of 19.3%, and the Social score gain of 14.8%. It also increases OCR-Bench by 101.5%, and raises TextVQA accuracy by 44.2%. These results show that BLV-focused post-training improves both accessibility-specific spatial grounding and general visual-text reasoning. Deployed on a mid-range Android smartphone via Mixed-Precision Quantization, the model remains approx. 450 MB and runs entirely on-device, offline and without network dependency, generating descriptions with latency dependent on host hardware capabilities. Our model, dataset, and code is publicly released at https://smol-vl-blv.github.io/Smol-VL-BLV-website/

Long horizon language model agents continuously accumulate interaction history, increasing computational cost while making relevant information harder to preserve and reuse. Existing context management methods mainly focus on how to compress or retrieve history, but largely leave open whether the model itself already represents the need for these memory operations before they occur. We study the hidden state immediately before each agent action and find that compression and recall needs are already encoded in the model's internal representations. These signals cannot be explained by simple context length or interaction progress, and they exhibit distinct formation patterns across model depth. We further show that most memory decision information is preserved in a compact recent context, while selectively restored historical evidence complements the long range dependencies that recent context misses. Based on these findings, we propose Preaction Memory with Evidence Retrieval (PaMER), which combines state guided compression with external evidence retrieval. PaMER+ further introduces step level evidence selection to recover only the historical information required by the current task. Experiments on WorkBuddyBench, across multiple context management baselines and model backbones, show that our framework substantially reduces context consumption while maintaining competitive task performance.

At test time, large language models (LLMs) can encode historical information in activation memory (i.e., KV caches) and parametric memory (i.e., updated parameters). While activation memory is generally considered effective for factual recall and parametric memory for learning new tasks, their interplay remains unclear. In this work, we systematically investigate the role of memory in few-shot learning through controlled experiments. We find that activation memory is superior for recalling facts, whereas parametric memory does not consistently outperform activation memory in task learning. Moreover, our experiments show that the composite task, Conditional Arithmetic, requires the synergy of both memory types. Through neuron-level analysis, we find that the model activates distinct sets of neurons when accessing the same historical information through activation versus parametric memory. When both memory types are combined, the model recruits neurons from both sets, which is crucial for solving Conditional Arithmetic. These findings suggest that neither memory mechanism alone is sufficient for this composite task, highlighting the importance of their collaboration.

Generalized Langevin equations describe non-Markovian dynamics in which the evolution of resolved variables depends on their past. We propose a memory-conditioned diffusion method for learning stochastic flow maps of these dynamics from observed trajectories, without identifying a memory kernel or reconstructing unresolved variables. A compact, recursively updated bank of exponential filters enables the flow map to retain predictive history over multiple time scales without conditioning on long observation windows. The next-step distribution is conditioned on the current observation and this memory state, whose storage and update costs are independent of the history length for a fixed bank size. Predictive criteria guide the memory budget, with reference-assisted selection in the vector benchmark, and an optional linear projection further reduces the conditioning dimension. A kernel-based score estimator generates conditional samples without training a score network, and these samples are used to train a neural flow map for autoregressive simulation. Three numerical examples assess long-memory retention at small conditioning dimension, predictive compression in coupled vector dynamics, and non-Gaussian conditional distributions and intermittent events. The non-Gaussian example reproduces conditional asymmetry and burst statistics in a stochastic model of the plasma scrape-off layer.

Language models typically construct attention values from contextual hidden states, even when some of their content may be reusable across contexts. We investigate whether token-indexed memory can replace the dedicated value projection when complemented by contextual information. We propose Memory Attention (MA), which forms values by combining layer-specific token memory with contextual keys. The memory supplies token-specific representations, while the keys preserve context dependence. At inference, normalization can be folded into the memory tables, reducing value construction to lookup and addition. Token-indexed retrieval also enables CPU offloading with prefetching, reducing GPU parameter storage. Under matched training token budgets and with additional memory parameters, experiments across attention configurations show improved language modeling and average downstream performance.

We study the nonequilibrium dynamics of a minimal recurrent transformer with $N$ normalized tokens, $Q=K=I$, and a negative value map $V=-I$. Similarity-based attention selects nearby representations, while the negative value map drives tokens away from the selected field. This feedback can continually reorganize both the representation geometry and the attention network. For $d=2$, the tokens lie on a circle, where the regular polygon is an exact fixed point. As the attention feedback strength $γ$ is increased, the polygon loses stability through a flip bifurcation, giving rise to period-two motion, chaos, and cluster-exchange or cluster-flip states. Despite this temporal complexity, attention remains diffuse as $N\to\infty$ at finite fixed softmax sharpness $β$. Attention condensation instead emerges in the scaling regime $β\sim N^2$. In the hard-routing limit, repulsive updates amplify local perturbations and routing-partner switches transmit them ballistically, producing an emergent butterfly cone in representation space. High-dimensional geometry provides a distinct route to localization. For $d=N\to\infty$, simulations from Gaussian initial conditions provide evidence for a condensation transition at $β=O(1)$, driven by dynamically generated finite overlap gaps. Depending on $γ$, the resulting phases include diffuse simplex-like states, consensus flips, condensed active routing with signatures of chaos, and fragmented cluster flips. These results establish temporal activity, attention condensation, and geometric clustering as distinct collective phenomena, and show that sparse attention can sustain persistent dynamics rather than freeze it.

The growing deployment of Internet of Things (IoT) devices has increased the need for privacy-preserving intrusion detection systems that operate directly on resource-constrained hardware. Federated Learning enables collaborative model training without sharing raw data, but conventional federated models are often too large and unstable for deployment on microcontroller-class devices. TinyML techniques enable compact neural networks but are typically designed for inference-only workloads. This work investigates combining Federated Learning with TinyML-based model compression for intrusion detection in IoT environments. We evaluate compression strategies including knowledge distillation, structured pruning, and quantization within a federated training pipeline. Preliminary results show that training stability plays a critical role in federated TinyML systems. In particular, server-coordinated cosine learning-rate scheduling improves Attack Recall from 46.7% to 93.85% while enabling substantial model compression and efficient edge deployment. These findings provide insights for designing lightweight and privacy preserving intrusion detection systems for IoT devices.

DSpark-style parallel drafters have made speculative decoding highly effective, yet their draft phase remains serialized on the critical path of every round. Parallel speculative decoding (PSD) overlaps drafting with verification, yet existing methods must guess the accepted prefix and bonus token in advance: a wrong guess reverts the whole batch to serial drafting. We present DPara, a PSD framework that reuses effective parallel drafters yet guarantees backbone--verification overlap in every round, thereby eliminating this probabilistic fallback altogether. While the target verifies, DPara's diffusion backbone precomputes draft representations for every acceptance boundary with the bonus left unspecified; a lightweight autoregressive head then combines the revealed verification outcome with the matching precomputed representation to emit the next round's draft tokens almost instantly---fully parallelizing the dominant backbone forward with verification and leaving only the negligible head cost serial. Experiments on Qwen3-8B and Qwen3-14B across seven math, coding, and chat benchmarks show that DPara achieves average speedups of $3.21\times$ and $3.52\times$ over autoregressive decoding, surpassing the strongest serial and parallel speculative decoding baselines alike.

The key-value (KV) cache of autoregressive transformers can be viewed as a fourth-order tensor spanning attention heads, tokens, features, and grouped layers. We measure the singular-value spectra of all four mode unfoldings on Mistral-7B-v0.3 and LLaMA-2-13B and compare four standard tensor decompositions: Tucker, CP, tensor train, and t-SVD, at matched storage. The spectra partition the four axes into two classes. The token and feature modes carry low-rank structure, particularly for keys. The head and layer modes are nearly full-rank and resist compression at any practical error level. Among the four decompositions, Tucker achieves the lowest reconstruction error at every compression ratio from $2\times$ to $5\times$, because it can leave the full-rank modes untouched. Comparisons with two-dimensional unfolding baselines show that the preferred representation differs between keys and values: 2D methods achieve lower key error, while four-way Tucker achieves lower value error at matched storage. A mode-pinning theorem certifies the full-rank preservation from the measured spectra alone. Two further spectral properties affect the compressible modes without touching the full-rank ones: values reach a higher error floor than keys at every ratio, and post-RoPE keys lose $41%$ - $64%$ of their pre-RoPE compressibility on both models.

Memory-polynomial digital predistortion (DPD) evaluates a small fixed coefficient set over a sliding input history, so its reduction step is a complex-MAC workload with local reuse. We map this DPD reduction kernel onto In-Memory Accelerator eXtension (IMAX), a programmable CPU-Grounded Linear Array (CGLA) composed of a one-dimensional processing-element/local-memory pipeline. For a (P,M)=(5,5) odd-order memory-polynomial instance, the mapping keeps the 120 B coefficient set in local memory, advances the five-tap history over 1024-sample tiles, and realizes the 15 order-delay terms as a 33-stage streaming complex-MAC reduction. The evaluation measures kernel latency and modeled energy. All measured paths use the same single-precision complex workload of 32 sequences, each with 2048 complex samples, across an IMAX FPGA prototype, a CUDA implementation on an RTX 4090 system, and an ARM-NEON implementation on Jetson AGX Orin. With this 1024-sample tile configuration, the IMAX FPGA prototype reports 20.201 ms end-to-end latency and 1.948 ms kernel-only latency. Using the previously reported 28 nm IMAX frequency and power model, the projected IMAX configuration gives 3.14 ms end-to-end latency and 0.34 ms kernel-only latency. The RTX 4090 baseline has the lowest end-to-end latency at 0.484 ms. Under model-based platform power accounting and the stated power assumptions, the projected IMAX configuration gives 169.1 times smaller modeled end-to-end energy per batch than the RTX 4090 baseline. This value uses platform power assumptions rather than workload-dependent runtime power or a direct silicon power measurement. A controlled synthetic PA-model validation checks that the same 15-term form improves test-set NMSE by 26.1 dB and ACLR by 26.0 dB. These results characterize the mapped memory-polynomial DPD reduction on IMAX for the evaluated tile configuration and power model.

Microscaling quantization techniques are increasingly used to represent neural network parameters with 8 bits or fewer while preserving near-full precision accuracy. However, applying these methods efficiently in convolutional layers is not straightforward. A naive approach transfers full-precision weights and activations to processing units and quantizes each tensor twice, resulting in much more memory movement than expected. Additional overhead comes from the activation tensors, whose sizes grow substantially because of the im2col transformation applied before quantization. We propose MicroQonv, a way to combine microscaling with convolutional layers' forward and backward operations by quantizing each tensor only once and quantizing the activation tensor before applying a modified version of im2col: channel-batch-first im2col. MicroQonv reduces the quantization cost by a factor of $\times2$ for weights and gradients, and by up to $\times9$ for activations, at a negligible accuracy cost. It reduces memory movement and storage by up to $\times7.53$ compared to their full-precision counterparts. This way, MicroQonv reduces microscaling-quantized activation memory movement by $\times3.5$ for state-of-the-art object detection models YOLOV8nano and $\times2.2$ for YOLOV26nano. It also enables 4-bit microscaling in a quantized latent replay strategy for continual learning at the edge, improving accuracy by +5.7% to +11%.

Unlearning ensures LLM compliance by removing the influence of private or copyrighted training data. However, since LLM models typically undergo post-training compression, like quantization, in practical deployment, it has been observed that the unlearning effect can be substantially weakened, with the forgetting behavior degrading more severely than that of model utility. This paper proposes a quantization-robust unlearning framework that makes forgetting robust to quantization while maintaining overall model utility. We analyze this gap through the lens of loss landscape. Specifically, our analysis reveals a curvature-based criteria that pinpoints sensitive weights in the unlearned model that leads to both non-robust forgetting and reduced utility. We therefore propose sensitivity-guided noisy regularization, which is applied on the sensitive parameters to steer the model convergence towards a smoother minima of uniformly low forget and retain losses. Balancing unlearning and utility, we further propose forget-critical optimization, which updates only forget-critical layers, preserving most of the network to retain useful knowledge. Extensive experiments on the MUSE and TOFU benchmarks across multiple LLM unlearning algorithms show that our approach achieves substantially more quantization-resilient forgetting while maintaining utility.

Driving world-action models improve planning by coupling multimodal reasoning with future prediction, but their growing inference cost increasingly conflicts with the real-time latency requirements of vehicle control. Existing acceleration methods reduce tokens, layers, or sampling steps with policies selected prior to deployment, yet leave residual runtime variation largely unexploited after offline profiling and static scheduling on shared onboard compute. We observe that the largest admissible compute budget varies systematically with the residual runtime state, while recent realized latency provides a direct signal of the available compute slack. Motivated by this observation, we propose SlackDrive, a pre-inference compute allocator that reuses realized latency to select the compute budget of each control step before model execution. SlackDrive profiles the latency and planning utility of a small discrete budget set once, estimates online compute state from completed forwards, and selects the highest-utility budget predicted to remain within the admissible latency envelope, complementing existing profiling and resource scheduling while preserving the driving backbone and its compute actuator. On NAVSIM v2 with DriveDreamer-Policy, SlackDrive improves latency-constrained EPDMS by $21.7%$ over the strongest baseline under a stringent latency regime, while the full-budget model and preconfigured token-pruning baselines exceed the admissible latency envelope under runtime contention.

Modern GPU kernels fuse increasingly more work into a single kernel, and intra-kernel tracing has become the mainstream method to profile them. Tracing inserts probes into the kernel to record its runtime states, and the fidelity of the trace determines the efficiency of performance optimization. Unfortunately, existing tools insert probes before compilation. These tools interfere with the compiler's optimizations, so they trace a different binary from the one the GPU executes. They also add significant runtime overhead. Xtrace is the first GPU kernel tracing system with near-zero compile-time interference and minimized runtime overhead. Xtrace inserts probes directly into the compiled kernel binary. It reuses only the registers that hold dead values at the insertion address and resolves all hazards with the compiler's hazard tables. It further schedules the instruction order, register allocation, and control bits to minimize the runtime overhead the probe introduces. Xtrace supports 19 NVIDIA and AMD GPU architectures, and is publicly available for use at https://g-watch.github.io. We evaluate Xtrace on major production large language model (LLM) kernels against the state-of-the-art tracers Neutrino and IKET from NVIDIA. On H100, B300, and MI300X GPUs, Xtrace preserves 94-98% of the instructions of the kernel, while existing tools preserve only 8-48%. Xtrace adds only 0.9-2.8% overhead, while existing tools add 3.8-75.6%. Xtrace guides a coding agent to reach the same FlashAttention-3 performance with 3.9x fewer iterations than existing traces do. Thanks to our binary-level instrumentation, Xtrace also traces the faster closed-source cuDNN kernel, which guides the agent to lift the open-source FlashAttention-4 by 5.2-13.3% in throughput.

Deploying deep learning models on edge CPUs is bottlenecked by computational and memory constraints. Mixed-precision quantization promises to reduce inference latency while preserving accuracy. However, quantization affects different layer types in inconsistent ways, so identifying where accuracy loss is minimized and latency reduction is maximized is critical, as the effect accumulates over a full deployment into substantial savings or unacceptable task degradation. Such identification relies on sensitivity metrics, proxies that estimate layer-wise degradation without evaluating the task accuracy of every candidate policy. Nevertheless, widely used metrics fail systematically on modern architectures. We present a systematic empirical study of 13 sensitivity metrics for layer-wise INT8 quantization across four distinctly different neural networks, and validate the resulting policies on two ARM64 platforms. Gradient-based sensitivity methods fail on 4 out of 8 model-hardware configurations and weight-based statistics on 2. In contrast, the Jensen-Shannon Divergence achieves zero catastrophic failures, reliably isolating the layers that cannot be safely quantized. A sensitivity metric alone does not define a policy, and the fixed thresholds typically used for that step are fragile over the highly skewed distributions of modern architectures. We address this with K-Means clustering, achieving near-lossless accuracy and a mean speed-up of $1.81\times$ over the full-precision model. Finally, we reveal that excluding from quantization the layers whose speed-up is negligible, regardless of their sensitivity, can be counterproductive, as it induces computational graph fragmentation and disables operator fusion. Our results yield concrete allocation policies for practitioners and researchers deploying quantized vision models on heterogeneous edge CPUs, without GPU access or gradient computation.

Transformer language models process sequences token by token in an autoregressive manner, making growing contexts increasingly expensive. Yet many adjacent token spans are highly predictable or frequently occur as stable units, suggesting that their representations may be compressible. We introduce a method for distilling sequential computation by replacing spans of input tokens with collapsed representations, computed on the fly by a lightweight merge module. This module generates a single surrogate embedding from a sequence of static token embeddings that captures the functional role of the multiple tokens, allowing pretrained models to operate on compressed inputs without architectural changes or re-training. We apply this approach during inference to compress both prompts and intermediate decoding steps, using a rollback mechanism to substitute stored multi-token KV cache entries with their single-step surrogates. Experiments across diverse models show that the merge module can be used to reduce effective sequence length by up to 40% with minimal accuracy degradation across language modeling evaluations and downstream tasks, including question answering, summarization, commonsense reasoning, and long-form mathematical reasoning. Additional lightweight adaptation of the merge module further improves the accuracy-compression trade-off in selected settings. These results demonstrate that sequential token computation in Transformers can be effectively approximated through condensed surrogate representations that approximate the original behavior without model updating.

Large language model (LLM) inference transfers model weights and activations for every generated token, making memory traffic and its energy cost part of the decode path. BitNet b1.58 represents its low-bit weights by ternary values and uses integer activations. However, this arithmetic does not match conventional int8 or floating-point general matrix multiplication, and existing BitNet accelerators implement it in specialized datapaths. We instead map this operation to a CPU-Grounded Linear Array (CGLA), a programmable ASIC with explicit direct memory access, local memories, and reusable compiler-visible integer lanes. The mapping adds OP_SMA4 as a reusable signed-int4 multiply-accumulate instruction rather than a BitNet-only datapath. Each ternary weight occupies one signed 4-bit lane. Each int8 activation is split into two signed-int4 fragments and reconstructed by shift-and-add. Frequency scaling of the 145 MHz FPGA measurement to an 840 MHz 28 nm CGLA achieved 0.390 ns per signed-int4 product. We showed that CGLA-offloaded BitNet C++ execution measures 2.52 tokens/s.

Recent video-language models increasingly adopt hybrid architectures that interleave linear and full attention layers for efficient long-context processing. While the recurrent state of linear attention remains fixed in size, the KV cache of full attention continues to grow with the video stream, making eviction necessary under a bounded memory budget. The key challenge in streaming is that eviction must occur before the question arrives, so what to retain has to be decided without the question. Existing eviction methods derive token scores from the KV cache itself, using position, attention, or key-value representations, and attention-based scores further require proxy queries or extra computation. Hybrid backbones offer another source of signal. In gated-delta linear attention, the recurrent state is updated by the residual between each input and what can already be retrieved from the state, so its change over a chunk of frames reflects how much new information the chunk brings. We propose DeltaS, a query-agnostic, training-free method that retains video chunks inducing larger normalized state change, or state drift. In a controlled comparison with the budget and retention policy held fixed, state drift outperforms position-, attention-, and key-value-based signals. With a signal costing only 1.9% of the forward pass, DeltaS surpasses the strongest query-agnostic bounded-memory baseline by 2.1 points on average across six long-video benchmarks and by 5.6 points on the longest benchmark. These results suggest that the two memories of hybrid architectures can work cooperatively. Code is available at https://github.com/MaumAI-Company/DeltaS.

Prefix caching is critical for efficient large language model (LLM) serving, particularly for agentic workloads that repeatedly invoke the model with a growing conversation and tool-use history. By reusing the key-value (KV) states of previously processed prefixes, prefix caching avoids redundant prefill computation. Its effectiveness, however, depends on retaining a sufficiently large set of KV cache states. Provisioning enough cache to preserve all historical KV states is prohibitively expensive and often unnecessary, whereas insufficient capacity can substantially degrade the cache hit rate. Determining the KV cache working set, defined as the minimum cache capacity required to achieve a target hit rate, is therefore essential for efficient cache provisioning and system design. We present KVSET, an online analyzer that estimates the KV cache working set of LLM serving workloads. KVSET uses the Mattson stack algorithm to efficiently estimate cache hit rates across a wide range of cache capacities. For each KV cache page, KVSET computes its LRU stack distance and compares it with the page number of each candidate capacity. This comparison determines whether the page would be a hit at each capacity without independently simulating every capacity configuration. KVSET therefore substantially reduces the computational and memory overhead of conventional capacity-by-capacity simulation and makes online working-set analysis practical. KVSET further determines the minimum cache capacity based on the maximum LRU depth among the prefix pages required to achieve the target hit rate. We validate KVSET using traces collected from production LLM workloads and show that its estimates closely match measurements from real cache deployments. The open-source implementation supports both online request processing and offline trace replay.

Large Reasoning Models (LRMs) have achieved state-of-the-art performance in reasoning tasks by utilizing Chain-of-Thought (CoT) reasoning. To achieve fast execution speed, speculative reasoning techniques adopt a lightweight draft model for candidate token generation followed by process reward models (PRMs) for verification and a strong target model for refinements. This paper identifies that the existing speculative reasoning paradigm follows a strictly forward-only reasoning trajectory, which lacks robustness and can lead to severe error propagation if early reasoning steps are suboptimal. Furthermore, executing these disparate inference schemes, including sequential drafting and parallel verification on homogeneous GPU platforms, can lead to severe resource underutilization. To address this, we propose HeteroReason, an algorithm-hardware co-designed heterogeneous FPGA-GPU inference paradigm specifically tailored for LRM speculative reasoning. At the algorithmic level, we introduce a backtracking-enhanced workflow that enables the system to recover from low-quality states and explore alternative reasoning trajectories, significantly improving reasoning robustness. At the system level, the draft model is offloaded to the FPGA while deploying the PRM and target models on GPUs. A specialized workflow is optimized to achieve prefill-decode disaggregation, which exploits shadow synchronization to overlap GPU-side refinements with FPGA-side token updates to effectively hide synchronization latency. To mitigate inherent sequential constraints, we propose a step-ahead speculation and refinement scheduling scheme, transitioning the system from a sequential execution scheme to a parallel pipeline. Experimental evaluations show an average 4.2% accuracy improvement, with 1.01x-1.42x latency speedups and 1.25x-1.57x improvements in energy efficiency compared to homogeneous GPU baselines.

Recurrent-depth language models, such as looped Transformers, repeatedly apply shared network blocks to refine latent representations without generating explicit intermediate reasoning tokens. However, each step recomputes full attention over the entire context, repeating costly global routing. We study how attention routing evolves across recurrent depth and find a consistent separation in convergence timescales: attention support and distributions stabilize substantially earlier than hidden states and attention outputs. This suggests two stages of recurrent inference: early discovery of a sparse working set, followed by representation refinement over largely stable routing support. Motivated by this finding, we introduce WISE (Working-set Inference with Support Exploitation), a training-free method that uses unrestricted attention during early recurrent steps to discover a block-structured working set, then reuses its support in later steps while keeping attention weights and recurrent refinement dynamic. Controlled interventions show that multi-step discovery yields more effective working sets than first-step selection, and that support reuse better preserves model behavior than more restrictive forms of attention reuse. Across multi-hop QA benchmarks, WISE largely preserves full-attention performance. Matched context-scaling experiments reveal an increasingly favorable quality-efficiency tradeoff as routing support becomes sparser with longer contexts. A sparse-attention implementation achieves up to a 1.76x late-step attention speedup over native FlashAttention at 4K context. Code: https://github.com/tbn5pj/WISE_code.

Scaling a language model is not only a question of final quality: the architectural choice determines how much computation is spent during training, prompt processing, and autoregressive decoding to achieve certain model quality. An ideal model architecture should lower all above computation costs to facilitate scaling to a larger model, while ensure the larger model indeed outperforms smaller baselines. We introduce KV-Invariant Transformer Expansion (KITE), a scaling paradigm that achieves this goal. It trains the model from a smaller size to a larger size (i.e., saving training costs via upcycling), while places newly added parameters in regions that do not affect attention KV. Consequently, during inference, prefilling KV only relies on the smaller part of the model, so the inference costs are saved. As a concrete instantiation, we present Step Scale Transformer (SST), a two-tower decoder in which one tower produces KV and the other reads them. At comparable cumulative training compute, SST, a 67B MoE model with 2.15B active body parameters per decode token, achieves lower training loss than 47B and 63B MoE Transformers with 1.48B and 2.02B active body parameters, respectively, while reducing estimated inference cost by 6.7% and 31.6%.

Collaborative medical AI platforms allow researchers to train models on sensitive imaging data while restricting data export. However, trained models can serve as covert carriers of patient information: medical images may be encoded within model parameters and reconstructed outside the secure environment. Existing defenses rely on lightweight sanitization (e.g., fine-tuning, pruning, quantization) and limited statistical auditing, creating a realistic insider exfiltration risk. We introduce a high-capacity neural steganography attack that encodes medical images as continuous latent representations embedded into model initialization. A StyleGAN2-based adversarial autoencoder learns compact latent codes regularized to match standard weight initialization statistics, keeping embedded parameters statistically consistent with clean models. Noise injection during training improves robustness to export-time mitigation. The carrier model remains functional on its intended task and hidden images can be reconstructed directly from its weights after export. This continuous encoding enables robust and scalable exfiltration, allowing up to 99 brain MRI volumes to be embedded within a 30MB model, and remains recoverable under mitigations that disrupt prior bit-level schemes. While reconstructions are approximate rather than pixel-exact, embedded content remains anatomically recognizable and recoverable at scale, exposing a privacy risk distinct from prior bit-level approaches. Experiments on MIMIC-CXR, BraTS, and LiTS demonstrate effectiveness across modalities, tasks, and architectures, highlighting the need for structural defenses beyond parameter-level sanitization. Code is available at https://github.com/ElieThellier/high-capacity-robust-medical-image-exfiltration.

Multi-GPU servers have become the standard building block of modern data centers, providing aggregated capacity through high-bandwidth interconnects. At the same time, workloads such as LLM inference exhibit highly dynamic memory demands, which can cause one GPU to exhaust its local memory while others remain underutilized. This mismatch motivates a model of elastic resource sharing across GPUs. We present EMA, a memory sharing system that allows GPUs within a server to borrow and reclaim memory from each other, forming an elastic pool of capacity. EMA ensures performance transparency for both borrowers and lenders. For borrowers, prefetching hides remote access costs so that applications experience remote and local memory as indistinguishable in performance. For lenders, borrowed resources remain reclaimable on demand, guaranteeing that performance never falls below that of static partitioning. While our design focuses on memory, the same principle naturally extends to other GPU resources. Our evaluation shows that EMA improves individual user throughput by up to 52%, achieves 96% of the throughput of a system provisioned with 2X capacity, and maintains latency similar to the static local baseline.

As serving capacity demand surpasses that of training, serving efficiency becomes increasingly important. Prefill-decode (P/D) disaggregation improves serving efficiency through specialization and isolation of the two phases. These benefits rest on a static partitioning. Phase demand, however, is not static. We observe that in a large LLM fleet the ratio of uncached input to output tokens has peak-to-mean ratios up to 4.7x at minute timescales, and that in a public agentic trace the hourly ratio spans a median 24.5x within a single day, while reassigning a replica takes tens of minutes. Agentic traffic sharpens the mismatch. Sizing each pool at its ninety-fifth percentile leaves up to 17% of cluster capacity unused; sizing below it converts the same imbalance into queueing and unrealized throughput. We present Crossflow, which makes this boundary elastic without changing node roles. Each decode node publishes a short-lived, revocable lease that bounds local-prefill compute, KV capacity, transfer work, and projected output. Across public and internal traces, Crossflow improves token throughput by 16.2-17.4% on geometric mean over static P/D, and by up to 43.4% at high load, while reducing mean TTFT at every evaluated point.

Large Vision-Language Models (LVLMs) incur high computational costs from redundant visual tokens. Although training-free attention-based multi-layer pruning in the vision encoder stage has been explored as an effective strategy, we find that pruning in shallow layers consistently degrades performance. In this paper, we aim to understand this problem and seek a solution. By analyzing attention patterns across network depth, we find that shallow layers primarily function as edge detectors with chaotic attention maps, while deeper layers transition through local subject recognition and unstable semantic aggregation. To address the misalignment between pruning strategies and network stages, we propose STD, a hierarchical token pruning framework that adapts token selection mechanisms to the functional role of each network stage. STD employs High-Frequency Spectral Analysis in shallow layers to deterministically preserve structural edges, uses Gaussian-Smoothed Attention in intermediate layers to maintain spatial coherence, and introduces a Stability-Adaptive Trigger in deep layers to execute pruning only during semantically stable phases. Extensive experiments show that STD outperforms state-of-the-art pruning methods by 1.1% on LLaVA-1.5-7B with 88.9% token reduction, while also being plug-and-play and highly effective when combined with other methods, and by 2.1% on LLaVA-NeXT-7B with 94.4% reduction, delivering a 3.9x speed-up in the prefilling stage. Our code will be released at https://github.com/Twilight03/STD.

Fine-grained mixture-of-experts (MoE) architectures have become a mainstream design for open-weight LLMs, with hundreds of experts and increasingly many selected per token. This shift makes dynamic expert pruning an attractive route to cheaper inference. Yet existing evidence comes largely from coarser architectures and likelihood-scored multiple-choice benchmarks, leaving three central questions open in the fine-grained regime: how redundant per-token expert selection is, how effectively existing pruning methods exploit that redundancy, and what governs a model's sensitivity to pruning. We fill this gap with a systematic empirical study of twelve fine-grained MoE checkpoints spanning nine architecture families, with a core suite of eleven benchmarks covering knowledge QA, mathematics, code generation, and general reasoning. We find that expert selection is far more redundant than the field's operating points assume: uniformly retaining about two thirds of the selected experts preserves 98.8% of unpruned performance on average, requiring only a one-integer change and delivering 1.2-1.7x measured speedup across two serving backends. This simple baseline leaves little room for dynamic allocation at conservative budgets: even the best published rules differ from it by under 1% at matched expert budgets. Their value emerges under aggressive pruning, where the best rules recover up to 3.0% over uniform truncation, with gains concentrated in the generative tasks that suffer the sharpest degradation. Sensitivity to aggressive pruning also depends on the model: larger and thinking models are more resilient, whereas multimodal models are more vulnerable. Together, these findings reveal how much expert computation fine-grained MoEs can dispense with, and establish when dynamic allocation earns its complexity, informing both practical deployment and future pruning methods.

Dynamic graphs are widely used to model time-evolving relational systems in real-world applications. Dynamic graph neural networks provide an effective framework for capturing both structural dependencies and temporal dynamics in such data. However, they typically intertwine temporal graph propagation with every optimization epoch and often maintain large trainable representations for each node-time pair. This design repeatedly recomputes largely unchanged historical structures, leading to substantial training and parameter overhead. To address this critical issue, we propose CacheDyG, a Cache-refine framework for efficient Dynamic Graph learning. Specifically, it decouples temporal propagation from routine parameter updates by constructing a time-ordered temporal dependency cache that stores graph-aware node-time representations in non-trainable buffers. During standard training epochs, CacheDyG reads from the cache and updates only a lightweight cache refiner, an adaptive residual gate, and the link predictor. Selective cache refresh further keeps cached representations aligned with the supervised objective while avoiding epoch-wise sparse propagation. Experiments on five dynamic graph benchmarks show that CacheDyG adopts substantially fewer trainable parameters and lower runtime to obtain more competitive predictive performance than baselines. These results demonstrate that cache-based decoupling provides an effective principle for scalable dynamic graph learning.