生成模型与LLM推理优化

Conversational AI agents require memory systems that are both scalable and semantically coherent across long interaction horizons. Existing approaches rely predominantly on large language model (LLM)-based summarisation at write time, which introduces non-determinism, escalating token costs, and opacity in pruning decisions. We present the Deterministic Memory Framework (DMF), a CPU-first approach that replaces generative memory compression with a fully deterministic pipeline grounded in classical NLP analysis, vector geometry, and mathematical scoring. DMF assigns each conversational interaction a Survival Score $Ω$ computed from deterministic content signals, conversational cues, and structured provenance, combined through a logistic projection. An interaction-count decay law, denoted as $Ω_{\mathrm{eff}}(Δn)$, governs how relevance evolves as new turns arrive, where $Δn$ is the number of newer interactions rather than wall-clock time, preserving full determinism. We present the mathematical formulation of DMF, its structured recall pipeline, the pruning decision procedure, and the evaluation protocol. Experiments are conducted on a purpose-built benchmark using the LoCoMo and LongMemEval datasets. We compare DMF against Mem0, a popular memory layer for AI agents. DMF achieves comparable accuracy while using zero tokens to prepare the memory context and 5x to 242x fewer tokens over the entire conversation. These results show that it is possible to eliminate LLM calls from the memory-management loop, reducing token costs to nearly zero and enabling deterministic memory systems for conversational AI agents.

Spiking Vision Transformer (SViT) models are promising low-power ViT models for solving vision-based tasks with state-of-the-art performance. However, their large sizes limit their deployments for resource-constrained embedded platforms, underscoring the needs of model compression. One of prominent compression techniques is pruning, and the state-of-the-art works employ unstructured pruning techniques to compress SViT models. Such techniques require specialized hardware architectures tailored for the sparsity patterns to maximize their efficiency benefits, making this approach not scalable. To address this, we propose PSViT, a novel methodology to perform structured pruning on SViT models, hence making it possible to efficiently accelerate their inference using the existing and widely-used computing architectures. To do this, PSViT employs several key steps: uniform channel-wise filter pruning to structurally eliminate the non-significant weights, sensitivity analysis to evaluate the impact of channel-wise pruning of individual layer on accuracy and network size, as well as fine-grained channel-wise pruning based on the sensitivity analysis and the given network architecture. Experimental results show that PSViT effectively obtains 22.4% memory saving through single-shot pruning, while maintaining high accuracy within 3% (70.3% without fine-tuning and 72.8% with fine-tuning) from the original non-pruned SViT model (73.3%) on the ImageNet-1K. These results also show that the PSViT methodology advances the effort in enabling efficient SViT deployments on resource-constrained applications.

The large sizes of Spiking Vision Transformers (SViTs) still hinder their embedded implementation, highlighting the need for model compression. State-of-the-art works compress SViT models through unstructured pruning, which needs specialized hardware accelerators for their specific sparsity patterns to maximize efficiency gains. Moreover, their manual approach requires a huge design time to find an appropriate pruning setting for each network, thus making this approach not scalable. To address this limitation, we propose PrimeSVT, a novel framework that performs automated memory-aware structured pruning on pre-trained SViT models, thereby maximizing their efficiency gains during inference amenable to widely-used computing architectures. To achieve this, PrimeSVT first sorts the SViT layers based on their sizes (i.e., number of parameters), identifies the targeted pruning layers based on their robustness under different pruning rates, then leverages this order for compressing the model layer-by-layer sequentially from the largest one to the smallest one (i.e., so-called prioritized compression policy), while considering the user-defined constraints (i.e., acceptable accuracy and memory saving). In each layer, PrimeSVT employs channel-wise filter pruning based on their L2-norm values to structurally remove the non-significant weights. Experimental results show that PrimeSVT saves 26.68% memory through automated single-shot pruning, while preserving accuracy within 3% (70.3% without fine-tuning and 72.9% with fine-tuning) from the original unpruned SViT model (73.3%), thus meeting the accuracy and memory constraints. These show that our PrimeSVT framework enables design automation for SViTs and their embedded implementation.

One-shot block drafters for speculative decoding generate the full draft in a single forward pass, achieving strong throughput by eliminating sequential token generation. However, they predict each draft token conditioned only on the prefix context, with no dependence on previously drafted tokens. This non-autoregressive conditioning causes the drafter's distribution to diverge from the verifier's true autoregressive distribution as draft depth grows. This problem becomes more severe in tree-based drafting, where distinct branches are forced to share the same marginal distribution for subsequent tokens. We propose TreeFlash, which addresses this by incorporating an MLP layer conditioned on the drafter's hidden state and the previous token to approximate an autoregressive distribution. TreeFlash retains the $\mathcal{O}(1)$ decoding time complexity of one-shot drafters by employing a two-stage approximation mechanism. TreeFlash achieves state-of-the-art performance across a variety of tasks and models, improving over marginal tree drafting by $12%$ higher block efficiency and $9%$ higher speedup.

Quantization is a standard path to deploying large language models, and a quantized model is typically judged acceptable when its perplexity or downstream accuracy stays close to the full-precision original. Whether the model still computes in the same way, or whether the interpretable features identified in the full-precision model survive weight rounding, is rarely tested, even as safety audits and steering interventions increasingly rely on those features. We ask whether sparse autoencoder (SAE) features extracted from a dense full-precision model remain faithful once that model is quantized. Using a frozen SAE as a fixed measurement basis, we encode full-precision and round-to-nearest (RTN) quantized activations on identical tokens and quantify per-feature survival by Pearson correlation, sweeping bit-widths from INT8 to INT4 on Pythia-70M and Gemma-2-2B. We find that feature survival is graded: features degrade systematically rather than failing all at once, with 62.4 percent of active features surviving at INT6 on Pythia-70M and 51.3 percent surviving at INT6 on Gemma-2-2B, and with most non-survivors blurred rather than destroyed. Survival is predictable from full-precision statistics alone, with cross-validated AUCs of 0.92 to 0.97 and peak activation as the strongest marginal predictor. Critically, task metrics can miss this damage: on Gemma-2-2B, INT7 improves perplexity while degrading 18.7 percent of features. Finally, quantization and matched-perplexity magnitude pruning damage strongly overlapping feature sets, with Jaccard overlap of 0.79 to 0.86 and damage-score Spearman correlation of 0.98, suggesting a shared mode of compression-induced vulnerability. These results show that behavioral parity is insufficient evidence that interpretability findings transfer to quantized deployments, motivating feature-level audits of compression.

Large Language Models (LLMs) have become integral to modern applications, yet their deployment remains challenging. Beyond executing the models themselves, practical deployment must address cost efficiency, low latency, and optimal resource utilization. Conventional approaches typically assume that an entire model can be hosted on a single device, which does not hold in many real-world scenarios, particularly in Edge and Fog environments where device resources are constrained. In this paper, we introduce E2LLM, a framework designed to enable efficient LLM deployment in such resource limited settings. Rather than simply partitioning a single model across all available devices, E2LLM replicates the full model across multiple groups of devices (replicas) and applies model parallelism within each replica. Each replica is assigned a specialized role PREFILL or DECODER based on its efficiency in handling input and output tokens. This separation leverages the inherent differences between these two phases of LLM inference. To effectively organize devices, we utilize a Genetic Algorithm to form clusters that maximize system performance. Within each cluster, we apply Dynamic Programming to determine an optimal partitioning strategy that minimizes bottlenecks in model-parallel execution. Experimental results demonstrate that our approach adapts robustly to varying workloads, including scenarios with significant variation in input and output token lengths. Compared to the Splitwise baseline, E2LLM reduces average waiting time by over 50% under high-demand conditions

Disaggregated LLM inference forces the KV cache to traverse the datacenter network before decoding begins, so transfer time enters directly into the Time to First Token (TTFT) budget. Current schedulers route on compute load and prefix-cache locality alone, ignoring the topological distance and dynamic congestion between prefill and decode instances. We close this gap with a thin operator-to-scheduler interface, the network cost oracle, and we prove that ignoring the network term renders cache-aware-only scheduling arbitrarily suboptimal as context length grows. NetKV, the O(|D|) per-request greedy that consumes this oracle, has tier rankings that are provably robust to stale telemetry. On a 64-GPU four-tier fat-tree simulator driven by Mooncake traces, NetKV reduces mean TTFT by up to 21.2% over round-robin and 17.6% over a tuned cache+load-aware scheduler, lifts SLO attainment by up to 20.1 percentage points, and keeps the Time Between Tokens overhead below 0.5 ms in every condition tested, with no changes to the transport, inference engine, or hardware.

Reasoning models improve accuracy through extended chains of thought, but their long outputs create a memory and compute bottleneck. KV cache eviction methods reduce this cost by evicting unimportant key-value pairs from the cache, yet they often yield worse accuracy than selection-based sparse attention alternatives, which keep the full KV cache. We identify key factors crucial to KV cache eviction accuracy. First, a small fraction of value states have abnormally large magnitudes, and evicting them causes catastrophic failure where models enter repetitive reasoning loops. Second, introducing stochasticity during eviction improves accuracy by increasing cache diversity. Based on these findings, we propose Value-aware Stochastic KV Cache Eviction (VaSE), a training-free recipe that protects large-magnitude value states and promotes diverse eviction decisions. Across six reasoning tasks, Qwen3 models using VaSE with 4x KV cache compression yield higher average accuracies than SOTA selection method at the same sparsity, while outperforming the strongest eviction method by more than 4%. Overall, VaSE bridges the gap between efficiency and accuracy, supporting FlashAttention2 and enabling a static memory footprint for reasoning models.

Test-time scaling is a powerful approach to obtain better reasoning in large language models, but it becomes memory-bottlenecked during long-horizon decoding, as the KV-cache grows. KV-cache quantization can help improve this, but current methods are evaluated under prefill-like settings and errors behave differently under autoregressive decoding. We show that in the latter regime, quantization errors accumulate across timesteps, driven primarily by incorrect token scales. We introduce KVarN, a calibration-free KV-cache quantizer that applies a Hadamard rotation followed by a dual-scaling variance normalization across both axes of the K and V matrices. We find that this combination fixes outlying token-scale errors and substantially reduces error accumulation over existing baselines. KVarN establishes a new state-of-theart for KV-cache quantization on generative benchmarks, including MATH500, AIME24 and HumanEval, at 2-bit precision. A vLLM implementation of the KVarN method is available at https://github.com/huawei-csl/KVarN

Causal video generators must predict from the past, but they need not learn only from it. In streaming autoregressive video diffusion, each emitted segment becomes a commitment that future segments must preserve. Standard training, however, only asks each causal state to explain the present. This creates what we call a representation-level planning gap: states that fit the current segment may discard identity, layout, and motion information needed for a consistent future. We introduce Video-Mirai, a training-only method that closes this gap without changing causal inference: the generator rolls out causally, a frozen foresight encoder reads the completed rollout non-causally, and a lightweight predictor distills the resulting stopped-gradient targets into causal states. Future frames supervise representations, never generator inputs. At inference, the encoder and predictor are discarded, leaving the original architecture, per-step FLOPs, and KV-cache behavior unchanged. Video-Mirai improves a strong Causal-Forcing baseline on 5-second VBench from 83.8 to 84.6 in terms of Total Score. On 30-second rollouts beyond the training horizon, subject consistency improves from 84.9 to 88.5 and background consistency from 90.2 to 91.9. Ablations identify future-conditioned targets as the key ingredient, and probes show that future frames become more decodable from current features. Causality should constrain inference, not representation supervision. Our study highlights that visual autoregressive models need foresight. Project page: https://y0uroy.github.io/Video-Mirai.

Deploying language-driven agents in robotics requires evaluations that reflect real-world task demands: natural-language instructions with reproducible outcomes. Such agents must connect language models to callable perception and control tools, and be assessed using deployment-critical metrics including latency, accuracy, and error modes. We present SCOPE (Simulation and Camera Operations for Perception and Evaluation), a modular agent for natural-language, open-vocabulary pan-tilt-zoom (PTZ) camera control and visual scene understanding, designed explicitly for edge deployment. SCOPE operates both in a Blender-based simulation environment and on a physical PTZ camera, executing all perception, planning, and control locally at the deployment site using edge-accessible compute. We release a 536-task benchmark spanning QA, single- and multi-step commands, counting, spatial reasoning, descriptions, and optical character recognition in a Blender-based simulation environment that exposes realistic PTZ control affordances. Execution traces are combined with an LM-as-Judge to evaluate latency, accuracy, and error modes. We evaluate 19 planner-perception model combinations pairing Qwen3 small language models (SLMs) with Moondream and Qwen vision-language models (VLMs). Stronger SLMs substantially reduce hallucinations and improve tool routing, leading to more reliable closed-loop behavior. Once a sufficiently capable SLM is used, perception becomes the dominant performance bottleneck. Mixture-of-Experts models on both the planning and perception side consistently match or exceed dense alternatives at latencies and memory footprints comparable to much smaller networks. Quantization provides additional efficiency gains with minimal accuracy degradation, identifying a practical, sim-to-real validated design point for real-time, edge-feasible language-driven PTZ control.

Large Language Model (LLM) inference relies on key-value (KV) caches to avoid redundant attention computation. While approximate KV cache retention techniques reduce memory usage by sacrificing model accuracy, lossless approaches instead evict KV cache blocks from GPU memory and reconstruct them on demand to preserve exact outputs. Existing lossless KV cache management systems primarily base eviction decisions on access frequency or positional heuristics, without considering how different KV cache blocks affect the execution efficiency of GPU attention kernels. In this paper, we propose AsymCache, a computation-latency-aware KV cache management system for LLM inference that explicitly aligns cache residency decisions with GPU attention kernel performance, including three key components: Multi-Segment Attention (MSA) for efficient non-contiguous KV context processing, a cache eviction policy that jointly optimizes hit rate and position-aware recomputation cost, and an adaptive chunking scheduler for high hardware utilization. Experiments show that AsymCache reduces TTFT by up to 1.90-2.03x and time-per-output-token (TPOT) by 1.62-1.71x over latest baselines, confirming the effectiveness of the method in common workloads and validating its design goal of balancing computational efficiency with cache hit rate. Moreover, the low-level design of AsymCache allows seamless integration into agent serving systems such as Continuum, where it further reduces average job latency by up to 18.1%.

Diffusion large language models promise parallel token generation, yet inference remains bottlenecked by deciding which masked tokens can be safely committed together. Fast-dLLM addressed this with KV caching and confidence-guided parallel decoding, but its decoding theory uses a homogeneous high-confidence assumption that effectively reduces each candidate set to its weakest selected token. We argue that this leaves speed on the table because real decoding steps exhibit heterogeneous confidence profiles. We propose Fast-dLLM++, a training-free extension that introduces Fréchet profile decoding: selecting parallel commit sets from the full sorted confidence profile rather than a single worst-case confidence. The resulting rule is a heterogeneous-confidence generalization of Fast-dLLM's factor selector and it recovers the previous rule exactly in the equal-confidence case and adds a provable heterogeneity bonus when the selected tokens have uneven confidences. Fast-dLLM++ leaves the model, diffusion process, and cache implementation entirely unchanged, making it a drop-in replacement for existing Fast-dLLM decoding. Experiments on GSM8K, MATH, HumanEval, and MBPP with the LLaDA-8B model show that the theoretical improvement translates directly into empirical gains: profile-aware selection improves the accuracy--throughput frontier by exploiting safe parallelism that weakest-token rules miss, achieving up to 37% higher throughput at comparable accuracy. Our anonymous code release is at https://github.com/Ringo-Star/FastdLLM_plusplus.

Financial AI agents often fail for a simple reason: they make users carry the complexity. A user must repeatedly restate goals, risk preferences, portfolio context, past judgments, and shifting market assumptions, while the agent answers, retrieves, acts, and forgets. In finance, this is not just inconvenient. In tasks such as market analysis, copy-trading review, and trade preparation, forgotten context and stale memory can create latency, repeated errors, weak auditability, and unsafe decisions. We propose the interaction-native knowledge harness (InKH), an architecture for financial LLM agents that absorbs complexity into the system. InKH converts user, market, portfolio, and tool events into structured operational knowledge. It uses passive knowledge injection to assemble a bounded working context buffer before the main model step, temporal graph memory for low-latency retrieval, a wiki audit surface for human-readable governance, and background extraction with maturity, decay, and write-time invalidation. We evaluate InKH on a reproducible controlled synthetic benchmark with 24 random seeds, 4 rounds, 80 episodes per round, and 6 baselines, producing 46,080 baseline-conditioned evaluations. InKH achieves mean task quality of 0.815 at 900 ms latency. Compared with agent-driven wiki-walk memory, it reduces latency by 82.95 percent, token cost by 82.29 percent, and stale-knowledge usage by 96.58 percent, while improving quality by 0.108 and traceability by 0.461. Compared with a temporal-graph system without invalidation, it improves quality by 0.050 and reduces stale-memory usage by 96.58 percent with comparable serving cost. The results support a design thesis for financial AI: adoption happens when complexity is absorbed by the system rather than transferred to the user. The benchmark validates architecture-level behavior, not live trading performance.

Speculative decoding accelerates inference by having a lightweight drafter propose tokens verified in parallel by the target language model. Block diffusion drafters such as DFlash generate an entire draft block in one pass, yielding per-position marginals; DDTree uses these to build a candidate tree that maximizes expected acceptance length under a fixed node budget. We observe, however, that acceptance length is non-decreasing in budget: it always favors larger trees regardless of verification cost, offering no principled basis for budget selection. We introduce CaDDTree (Cost-aware Diffusion Draft Tree), a method that directly optimizes token throughput (expected tokens generated per unit time) by jointly selecting the tree structure and node budget. We model draft and verification latencies explicitly, show that the throughput objective decomposes into a per-round one-dimensional search over the budget, and prove that under a convex verification cost the throughput function is unimodal, enabling an efficient greedy stopping rule. CaDDTree requires no offline budget search, adapting the budget each round from the current per-position distributions and verification cost. Experiments on Qwen3-4B and Qwen3-8B across eight benchmarks spanning reasoning, coding, and instruction-following tasks show that \caDDTree{} matches or surpasses DDTree with oracle budget selection on nearly all tasks.

Large Reasoning Models (LRMs) rely on long reasoning traces, making inference expensive. While low-bit quantization reduces per-token decoding cost, we show that aggressive 2-bit inference can fail to deliver end-to-end speedup because instability in the generation process inflates total token count. Instead of merely lowering answer accuracy, 2-bit quantization often produces much longer traces with repetitive loops, budget exhaustion, delayed commitment, and unclosed reasoning segments. We analyze full reasoning traces of Qwen3 reasoning models across mathematical and commonsense benchmarks and show that accuracy degradation is tightly linked to these process-level failures. To address them, we introduce two lightweight controls: FP16 planning, which gives the 2-bit model a short high-precision outline, and loop rescue, which detects repetitive traces and either commits to an earlier answer or falls back to FP16. On MATH-500, loop rescue improves Qwen3-8B accuracy from 17.2% to 74.2%, while planning plus loop rescue improves Qwen3-32B from 65.0% to 87.2%. Overall, our results show that extreme low-bit reasoning becomes practical when its failures are treated as controllable generation pathologies: with lightweight detection and selective FP16 support, 2-bit inference can recover accuracy while preserving real end-to-end speed. Our code is available at: https://github.com/brain-lab-research/quantized-reasoning.

Vision-language-model-based graphical user interface (GUI) agents have shown broad automation capabilities, yet deployment is bottlenecked by a key-value (KV) cache that grows linearly with interaction steps. For instance, UI-TARS-1.5-7B consumes 76 GB of GPU memory on merely five screenshots, approaching the capacity of mainstream 80 GB accelerators. Existing KV compression methods share two structural assumptions: aggregating visual-token importance into a single shared saliency map, and applying a fixed top-B cutoff to the fused score distribution. Pilot measurements refute both: spatial specialization lives at the attention-subspace level and migrates across layers, while the score distribution drifts in shape along a trajectory. We propose STaR-KV (Spatio-Temporal Adaptive Re-weighting), a training-free KV cache compression framework that calibrates token importance along three axes: (i) subspace-aware scoring driven by online spatial mutual information; (ii) a temporal stability discount that suppresses redundant cache entries from persistently attended subspaces; and (iii) an entropy-derived temperature that adaptively reshapes the score distribution. Across four GUI benchmarks, STaR-KV achieves the strongest average accuracy among state-of-the-art KV compression methods (e.g., GUIKV, SnapKV) at matched budgets, with no compression-stage FLOPs overhead (-0.07%) and cutting peak GPU memory by nearly 40% at a 20% KV-cache budget. Code is available at https://github.com/kawhiiiileo/STaR-KV.

4-bit quantization reduces the memory footprint and latency of large language model inference, but its aggressive precision reduction can severely degrade accuracy. Prior methods address this by decomposing each weight matrix into two components (e.g., via singular value decomposition) and quantizing them separately, assigning the bulk of values to a low-precision residual component while handling outliers with a high-precision low-rank component. However, such decompositions are designed to minimize the real-valued energy of the residual, rather than the post-quantization error of the residual and low-rank components. We propose TwinQuant, a 4-bit quantization framework that learns quantization-friendly decomposed subspaces and jointly reshapes both the low-rank and residual components. TwinQuant learns component-specific transformations via a joint optimization over the Stiefel and general linear manifolds, flattening their distributions and reducing dynamic-range imbalance. To enable efficient end-to-end execution, we further design a fused dual-component kernel that pipelines the two-stage low-rank computation on-chip and merges both components with a single epilogue, avoiding intermediate global-memory traffic. Across LLaMA3 and Qwen3 models, TwinQuant preserves near-FP16 accuracy and delivers up to $1.8\times$ end-to-end speedup over an FP16 baseline.

Massive activation spikes in Large Language Models (LLMs) severely degrade quantization by stretching dynamic ranges. While prior hypotheses characterize these as high-level scalar biases, we argue that they are merely the scalar intermediates of rigid, structural vector biases in the spike-carrying tokens. We show that these tokens converge to constant vectors after normalization that drive the attention sink and value-state drain mechanisms. We geometrically substantiate this by analyzing the coordination of projection weights: $W_K$ contrastively amplifies the vector, $W_Q$ aligns semantic tokens toward it, and $W_V$ projects it into the spectral null-space. Furthermore, we reveal that the model actively preserves these structural biases against Rotary Positional Embedding (RoPE) perturbations by localizing them in "zones of rotational stability" utilizing low-frequency bands and coherent channel pairs. Leveraging this, we propose INSERTQUANT, a post-training quantization (PTQ) framework that clamps spikes and restores their function via pre-computed template vectors. This renders activations strictly spike-free, enabling robust low-bit quantization with high fidelity. INSERTQUANT achieves parity with state-of-the-art per-tensor quantization methods on LLMs and uniquely generalizes beyond text to other modalities such as ViTs.

Autoregressive decoding in Transformer-based language models relies on the KV cache, whose memory footprint grows linearly with sequence length and becomes the primary bottleneck for long-context inference. KV cache eviction addresses this by retaining a fixed-size subset of key-value pairs and discarding the rest. We identify that a primary source of output degradation is not the residual attention mass on evicted tokens, which existing methods already minimize, but a directional mismatch between the retained and evicted token sets. Specifically, the evicted tokens in practice are often near-orthogonal to the retained ones. Thus, even a small evicted mass could have an oversized impact on the resulting direction distribution and amplify into substantial output error. This reveals a fundamental limit in existing strategies. To address this, we propose MomentKV, which maintains compact, small-size moment statistics over the evicted token set, including a count, key mean, value mean, and value-key covariance. During eviction, the moment statistics is leveraged to identify tokens already well aligned with and captured by the accumulated summary, keeping the evicted set geometrically regular. During inference, they yield a closed-form first-order approximation of the evicted attention output, forming a mutually reinforcing loop between selective eviction and accurate correction. On LongBench and RULER with LLaMA-3.1-8B-Instruct and Qwen3-4B-Instruct, MomentKV outperforms all baselines at every cache budget, with the largest gains under aggressive compression.

Deployers of online LLM services usually seek to maximize cluster-wide performance given a fixed number of GPUs. Tensor parallelism (TP) is necessary to fit modern models but scales sub-linearly as the TP degree t grows, due to cross-GPU communication and non-scalable runtime work, as predicted by Amdahl's Law. Conversely, increasing t improves memory efficiency and alleviates KV-cache contention and swapping. We identify and validate an empirical optimal TP degree t_e that balances these effects. We present Albireo, a parallel inference system that raises the attainable t_e by shrinking the non-scalable portion via overlap of scheduling and I/O with compute and sequence-parallel sampling, without changing model architectures. Across models and benchmarks, Albireo achieves up to 1.9x higher throughput, 48% lower latency, 28% higher GPU utilization, and 54% lower energy than vLLM; in production it yields up to 2x higher throughput.

LLM-based agents resolve a user task through many turns of dependent inference and tool calls, producing a workload whose total cost is unknown when the task arrives. Existing multi-turn systems keep the turn as the scheduling unit and decide, turn by turn, whether to disaggregate prefill from decode. That decision rests on the turn's decode length, tool behavior, and KV growth, quantities that are not observable when the scheduler must act, forcing the system to predict them. We show this dependence on prediction is imposed by the scheduling unit, not the workload. Raising the scheduling unit from the turn to the conversation converts turn-level irregularity into a stable, two-phase structure: 1) a compute-bound turn-1 prefill followed by 2) a long, memory-bound tail. Thus, with the conversation as the scheduling unit, placement reduces to reading the first-turn input length and per-decoder KV occupancy, both directly observable. We instantiate this principle in ConServe, which routes the first-turn prefill to a high-throughput prefiller, transfers the KV cache exactly once, and pins the conversation to a single decoder for its entire tail, with no learned model of decode-side cost. Against a per-turn prediction baseline, ConServe reduces p95 time-to-first-effective-token (the latency of a conversation's first user-visible output) by 51.08% and improves energy efficiency by 7.51% while preserving last-turn TBT and SLOs; mapping the two phases onto heterogeneous GPU tiers adds a further 22.75% in energy efficiency.

Diffusion large language models (dLLMs) have recently emerged as a promising alternative to autoregressive (AR) LLMs, offering faster inference through parallel or blockwise decoding. However, their masked language modeling formulation remains incompatible with standard token-level speculative decoding, one of the most effective acceleration techniques for AR models. In AR decoding, the causal mask preserves temporally valid token-level contexts, enabling a target model to verify multiple drafted tokens in a single forward pass. In contrast, dLLMs rely on mask tokens and bidirectional attention, causing the effective context to change across denoising steps and preventing direct token-level speculative verification. To bridge this gap, we propose a simple but effective speculative decoding algorithm for diffusion language models, named SimSD, which mainly adopts a plug-and-play masking strategy that equips dLLMs with temporally valid token-level contexts for speculative decoding. Our method explicitly introduces reference tokens from draft-model predictions and designs an attention mask that regulates their interaction with current-step tokens, allowing dLLMs to compute valid logits for drafted tokens in a single forward pass. This restores the key verification ability provided by causal masking in AR models while preserving the parallel decoding advantages of dLLMs. The proposed method is training-free and can be flexibly integrated with other acceleration techniques such as KV cache and blockwise decoding. Experiments on SDAR-family dLLMs across four benchmarks show that our method achieves up to 7.46x higher decoding throughput while maintaining and even improving average generation quality.

Video multimodal large language models (MLLMs) have made rapid progress on general and long-form video understanding, yet their ability to preserve brief answer-critical visual evidence remains underexplored. Many practical questions are determined by momentary visual events: localized actions or state transitions that may last only a few frames. Such evidence can be skipped by sparse frame sampling, suppressed by visual-token compression, or diluted by coarse temporal aggregation, causing failures that language-side reasoning cannot reliably recover. We introduce Moment-Video, a benchmark for diagnosing the temporal fidelity of video MLLMs through momentary visual event understanding. Each question is grounded in a localized, visually observable, and sampling-sensitive event, requiring models to notice, count, describe, or reason about transient evidence rather than rely on persistent objects, global scene context, or language priors. Moment-Video contains 1,000 human-verified video-QA pairs across 7 domains and 25 fine-grained subcategories, covering four task types: Temporal Occurrence, Temporal Counting, Action Description, and Temporal Reasoning. We evaluate 33 proprietary and open-source MLLMs on Moment-Video. The best-performing model, Seed-2.0-Pro, achieves only 39.6% overall accuracy, while most open-source models remain below 25%, revealing a substantial gap in momentary visual event understanding. Diagnostic analyses show that denser frame sampling improves some models but does not eliminate the bottleneck, and longer videos introduce stronger temporal-localization challenges. These findings suggest that current video MLLMs still lack temporally faithful representations for capturing, preserving, and using brief but decisive visual evidence.

Video Large Language Models (Video-LLMs) achieve strong performance in video understanding, but their excessive visual tokens bring substantial computational overhead. Existing training-free compression methods improve inference efficiency by reducing visual tokens, yet they often rely on local adjacent-frame similarity for temporal redundancy estimation or allocate token budgets mainly according to segment length. Such designs are sensitive to frame-level noise and fail to capture the non-uniform information distribution of real-world videos. To address these challenges, we propose InfoMerge, a training-free visual token compression method that improves token utilization through robust redundancy estimation and content-aware budget allocation. Specifically, we propose the Temporal Fingerprint Difference: a segment-level second-order temporal redundancy estimation strategy, which models the temporal similarity structure of tokens at the same spatial positions within each segment. We further introduce Content-Aware Budget Allocation (CABA), which dynamically allocates segment-level token budgets based on segment uniqueness and spectral-entropy-based representational richness. By reducing repeated preservation of redundant static regions and allocating more tokens to informative segments, InfoMerge makes better use of the limited token budget while maintaining strong performance. Extensive experiments show that InfoMerge achieves strong efficiency--accuracy trade-offs across multiple benchmarks and backbones, with more pronounced advantages under aggressive compression. On LLaVA-OneVision-7B, InfoMerge retains 98.8% of the original average performance while reducing 85% of visual tokens and achieving a 4.24-fold speedup in the prefill stage.

Discrete visual tokens should provide a compact representation for both token-based world modeling and planning in autonomous driving. However, most tokenizers are inherited from image generation and are optimized mainly for pixel reconstruction, which may leave a gap between what is easy to generate and what is useful to decode for driving decisions. We present a representation-guided and geometry-enhanced tokenizer that learns discrete tokens under joint supervision. The tokenizer aligns its discrete bottleneck with a frozen DINO feature space through feature decoding, while preserving appearance via RGB reconstruction with perceptual and adversarial losses. To inject geometric state-related cues, we add adjacent-frame depth and relative-pose supervision during training and stabilize joint objectives with multi-codebook quantization. We evaluate the same learned tokens with a lightweight planning readout and a GPT-style next-token world model. Experiments on NAVSIM show improved reconstruction fidelity and representation consistency, competitive planning performance under a fixed decoder, and better generative quality under matched settings.

Large Language Model (LLM) generation remains expensive because autoregressive decoding calls the model once for each new token. Speculative decoding reduces this cost by drafting multiple tokens and verifying them with the target model in one step, but its speedup depends on how many drafted tokens are accepted. Parameter-free draft sources can propose long continuations at low cost in structured and agentic workloads, yet a cache match that looks promising at one generation step may have low payoff at the next. We propose Hybrid Verified Decoding, which predicts the accepted length of a cache draft before verification and uses this payoff estimate to choose between cache verification and a model-based drafter. Across three LLMs and sixteen datasets, Hybrid Verified Decoding is especially effective on agentic workflows, where it outperforms EAGLE3 in every setting with a 2.73x average speedup. Our analysis shows how prompt structure creates cache opportunities, how high-payoff cache drafts concentrate in a small part of the draft space, and how payoff-guided selection reduces sequential decoding work, pointing to runtime draft selection as a promising direction for speculative decoding.

Post-training quantization is widely used for compressing large neural networks, but aggressive low-bit quantization can significantly degrade model quality. A common remedy is to augment the quantized weights with a low-rank correction, leading to approximations of the form $W\approx Q+LR$. In this paper, we study this low-precision plus low-rank representation through the layer-wise reconstruction objective $\|XW-X(Q+LR)\|_F^2$, where $X$ is a calibration matrix. We establish, to our knowledge, the first information-theoretic lower bounds for this problem under finite-alphabet and bounded low-rank compensation constraints. We then propose GPTQ-intrinsic LoRA, a training-free algorithm that incorporates the low-rank correction directly into a GPTQ-style quantization pass by appropriately augmenting the calibration Hessian. For the choice $L=V_r$, where $V_r$ contains the top right singular vectors of $X$, we prove layer-wise reconstruction error bounds in which the usual GPTQ dependence on $\|X\|_F^2$ is replaced by the rank-$r$ residual $\|X-X_r\|_F^2$, up to regularization terms. Under natural structural assumptions, these bounds match the information-theoretic lower bounds in their dominant scaling, up to constants and mild factors. We also introduce Bid-Up, a fixed-grid quantization refinement step that can be alternated with optimal low-rank compensation with guaranteed non-increasing layer-wise reconstruction error. Experiments on Qwen3 language models and DeiT vision transformers show that GPTQ-intrinsic LoRA improves over GPTQ and GPTQ followed by low-rank compensation, with additional gains from refinement loops.

Long-form automatic speech recognition (ASR) requires both high accuracy and low latency, but existing systems force a trade-off between the two. Chunk-based pipelines process audio in parallel windows for low latency, but lose cross-chunk context and need brittle heuristics to align speakers and timestamps at boundaries. Long-context ASR models resolve everything in a single pass for better accuracy, but are an order of magnitude slower. We propose Murmur, an inference system that overcomes this trade-off by operating at two levels. At the inter-chunk level, we revisit the chunk-based pipeline for modern long-context ASR, treating chunk size as a tunable hyperparameter, and show that intermediate chunk sizes strike a good balance of accuracy and latency. At the intra-chunk level, we exploit attention sparsity through a sliding window KV cache eviction policy applied to both output and speech tokens. On AMI-IHM, Murmur matches single-pass accuracy while reducing latency by 4.2x, with further gains from token eviction at less than 1% relative tcpWER degradation. The code of Murmur is available at https://github.com/uw-syfi/Murmur.

Frontier LLMs increasingly decide what a query attends to with a sparse-attention indexer that picks a few KV-cache blocks per query: attention's unit is now a small, reusable chunk. Agentic workloads hammer it: many sub-agents query one large codebase, reusing the same blocks. When that corpus outgrows one GPU it is partitioned across instances, so a query and the blocks it selects often sit on different GPUs: answering it means attention across instances. The reflex of prior cross-instance KV systems is to move the cache: pull the selected blocks to the requester. Multi-head Latent Attention inverts the arithmetic, compressing each token's key and value into one narrow vector, so a routed query row is only ~1 KB, smaller than the chunk it attends; routing the query is then often cheaper than moving the cache. Which primitive wins, over which fabric and request shape, is uncharted, least of all on device-initiated RDMA that makes per-request cross-node transfers cheap. We characterize cross-instance MLA attention on a real multi-node H100 cluster, distilling two reusable artifacts: a topology-aware cost model (probe / transfer / compute / return / merge) and a closed-form route/fetch/local predicate, whose constants we measure on real IBGDA, where the model tracks batched round-trips to within ~7%. At decode it routes the query, trading the cost of moving the cache (a ~3 ms re-adaptation splice for a contiguous chunk, or a scattered gather under selection) for a tens-of-microsecond round trip, and picks the fabric by probe latency, not peak bandwidth. We instantiate the cost model and predicate for MLA, but neither is MLA-specific: they apply wherever compression or sparse selection shrinks attention to small chunks (DeepSeek-V3.2, V4, and GLM-5.1 today). Extending them to a new architecture requires measuring just two coefficients: the routed payload and fetch's move-the-cache cost.