生成模型与LLM推理优化

This report describes Tail-Aware HiFloat4, our submission to the low-bit text-to-video generation quantization challenge. Our method adapts the public ViDiT-Q post-training quantization pipeline to Wan2.2 under the HiFloat4 numerical format. We quantize the main linear layers in both Wan2.2 transformer modules with W4A4 HiFloat4 fake quantization, keep numerically sensitive boundary modules in high precision, and introduce an activation-tail-aware percentile calibration module for channel-mask construction. Together with compact PTQ-state restoration, this design reduces the influence of rare calibration outliers while keeping the runtime HiFloat4 arithmetic and sampling pipeline unchanged.

Video spatial reasoning requires accumulating viewpoint-dependent evidence over time while retaining information useful to the question being asked. Existing spatial video-language models improve geometric perception and long-range context modeling, but often treat memory as a generic temporal cache, which can introduce redundant or irrelevant geometry and weaken long-horizon reasoning. We propose \ours, a question-guided geometric memory framework for video spatial reasoning. \ours injects camera-conditioned geometry into visual tokens and maintains two complementary memories: a Fine-Grained Context Bank for recent dense features and camera states, and a Semantic-Geometric Evidence Bank for compact long-range evidence. Each candidate frame is scored by the product of Q-Former-based question relevance and novelty with respect to the retained bank; this score is stored and reused during reading, while a capacity-based replacement rule keeps the bank compact. During reasoning, both memories are read before update and adaptively fused with the current frame representation. Experiments on VSI-Bench and VSTI-Bench show that \ours achieves state-of-the-art performance among evaluated spatial reasoning models, validating the effectiveness of question-guided geometric memory. Ablations further verify the contribution of the proposed evidence scoring mechanism.

Implicit neural representations have emerged as a promising paradigm for video compression, with recent methods achieving competitive performance on natural video. However, screen content video — common in remote desktop, online education, and cloud gaming — exhibits distinct statistics: sharp edges, limited color palettes, and strong temporal redundancy. Existing neural representation methods, designed for natural scenes, lack mechanisms to exploit these properties, leaving substantial room for improvement. In this paper, we propose NeR-SC, a neural representation framework tailored for screen content video. Building on the SNeRV backbone, NeR-SC introduces three screen-content-specific modules: (i) a learnable color palette that models the discrete color structure of screen content by restricting the low-frequency sub-band to a learned color set; (ii) a multi-gate dense fusion module that replaces sequential feature fusion with dense, attention-gated cross-stage interaction; and (iii) an embedding-level frame skip strategy that bypasses redundant decoder invocations for static frames, with zero training overhead. Experiments on DSCVC and VCD show that NeR-SC achieves 40.32~dB and 41.73~dB average PSNR, outperforming representative neural video representation methods and, at low bitrates, surpassing H.264 and H.265. The skip strategy enables real-time decoding with no loss in quality.

Omnimodal large language models enable unified audio video understanding, but long joint token sequences make inference costly, and existing benchmarks do not fully isolate audio visual association in noisy user generated videos. We introduce UGC-AVQA, a public UGC benchmark with 1,000 videos and 4,816 QA pairs, where an audio removal test ensures that benchmark questions require both acoustic and visual evidence. To reduce inference cost, we propose OMAC, a training free plug in compression method that preserves salient visual memory and temporally grounded audio anchors. To further make compact models robust to compressed inputs, we introduce O-MARC, a compression distillation framework for learning with memory compressed multimodal contexts. On Qwen2.5-Omni-3B, O-MARC improves the average score across four benchmarks to 45.8, outperforming full token inference at 44.1 and OmniZip at 41.0. OMAC also keeps inference efficient, reducing latency by 34.6% (1.53$\times$ speedup) and memory by 34.7% compared with full token inference.

We study Latent Recurrent Transformer (LRT), a lightweight augmentation of autoregressive transformers that reuses a high-level source-layer hidden state from the previous token as recurrent memory for the next token. Because this source state is already computed during ordinary decoding, LRT adds a cross-layer recurrent latent pathway across positions without inserting pause tokens or extra depth loops, and the standard attention mechanism and KV-cache interface are preserved. To pretrain this recurrence at scale without sequentially unrolling the transformer, we introduce interleaved parallel training: a single full-sequence initialization forward pass builds a shared buffer; then disjoint position subsets are refined in parallel and written back, so that all tokens receive recurrent-memory-aware supervision at roughly 2 times baseline compute. Across nanochat style backbones and a wide range of tokens-per-parameter budgets, LRT improves both language-modeling loss and in-context learning under matched effective compute while adding as little as 0.3% parameters.

Model internals encode rich information about how a large language model (LLM) processes its training data; however, post-training data engineering largely relies on external signals and ignores rich intrinsic signals lying in model internals. We propose SAERL, a data engineering framework for LLM reinforcement learning (RL). It models three intrinsic data properties: diversity, difficulty, and quality, using model internals extracted with Sparse Autoencoder (SAE), an advanced mechanistic interpretability tool. Each property grounds a concrete data engineering operation: SAE-space clustering with moderate batch mixing for batch diversity control, a difficulty proxy for easy-to-hard curriculum ordering, and a quality probe for data filtering. SAERL improves average accuracy by 3.00% over vanilla GRPO and reaches target accuracy with 20% fewer training steps on Qwen2.5-Math-1.5B, with consistent gains across model scales and RL algorithms. Experiments show that SAE transfers effectively across model families and scales, serving as a lightweight and reusable data engineering tool. These results demonstrate that model internals are a powerful and practical source of signals for post-training data engineering.

For the independent reference model with popularity vector $p\inΔ_N^\circ$, let $H_C(p)$ denote the exact stationary hit rate of an LRU cache of capacity $C$. We prove that, for every $1\le C<N$, the uniform popularity vector is the unique global minimizer of $H_C$ on the interior simplex. More sharply, along every nonconstant segment from the uniform vector to an interior point, the LRU hit rate is strictly increasing. The proof uses the standard exponential-age representation of the stationary LRU cache and gives an explicit positive pair-square formula for the radial derivative. Equivalently, for the move-to-front rule, the stationary search-cost distribution improves strictly in the usual stochastic order along every nonconstant ray away from uniform. This proves the radial restriction of the Fill--Holst Schur-concavity conjecture for move-to-front search-cost tails. In particular, all LRU miss probabilities and all nonconstant nondecreasing stack-depth costs decrease strictly along such rays. The result is radial rather than Schur-convex: full majorization monotonicity for LRU is known to fail, and the proof identifies the special positivity that survives on rays from the uniform vector.

Diffusion-based generation is increasingly powering production content pipelines; however, deploying these models at scale remains a significant challenge. Model weights frequently exceed the memory capacity of commodity GPUs, while the encoder, diffusion transformer (DiT), and decoder stages exhibit highly imbalanced computational and memory footprints. A natural remedy is disaggregated serving-running stages as separate services on heterogeneous GPUs-yet this introduces new bottlenecks, including stage handoff overheads and fast-changing workloads that make cross-stage provisioning and scheduling brittle. This paper presents DisagFusion, enabling asynchronous pipeline parallelism and elastic scheduling for disaggregated diffusion serving. First, DisagFusion introduces asynchronous pipeline parallelism that overlaps computation and stage-to-stage communication to reduce pipeline bubbles and mitigate network jitter. Second, DisagFusion employs a hybrid instance scheduling strategy that combines lightweight performance prediction with runtime feedback to continuously rebalance instance ratio across stages under workload shifts. We implement DisagFusion and evaluate it with modern diffusion models. Compared to a monolithic baseline, DisagFusion improves throughput by 3.4x-20.5x and reduces end-to-end latency by 18.5x, while enabling flexible, cost-efficient deployment across heterogeneous GPUs.

Large language model (LLM) inference is limited by high computational cost and memory bandwidth demands, making deployment on heterogeneous many-core processors challenging. Taking the MT-3000 processor used in the Tianhe supercomputer as an example, its limited main-memory bandwidth and distributed memory hierarchy exemplify these bottlenecks, making it difficult to directly migrate existing GPU-based inference frameworks. To address this problem, we propose THInfer, a hardware-aware inference framework that maximizes data locality under bandwidth-constrained conditions through hardware-software co-design and parallel strategy optimization. THInfer incorporates three key techniques: (1) a high-performance operator library for the VLIW SIMD architecture, providing hand-optimized FP16 kernels that achieve up to 70 percent of the peak performance per cluster; (2) a density-driven computation graph fusion and unified kernel scheduling mechanism, combined with a staged pipelined attention fusion method; and (3) a Prefill-Buffer-Decode (P-B-D) pipeline and bounded buffer management strategy, which supports hybrid parallelism and enables efficient multi-cluster collaboration through two-level communication based on MPI and hthreads. Experiments on the Llama model series show that THInfer improves throughput on the 7B model by 62 percent to 73 percent over DeepSpeed on two V100S GPUs and by 67 percent to 84 percent over the A800 GPU. The 13B and 30B models also demonstrate comparable or better performance. Moreover, THInfer maintains stable performance on the 70B model, whereas typical GPU-based frameworks fail to run under the same setting. Overall, THInfer significantly enhances throughput, reduces latency, and improves scalability, providing a feasible system solution for efficient and scalable LLM inference on heterogeneous many-core architectures.

Large-scale transformer training and deployment are increasingly constrained by the transfer of activations, gradients, and optimizer states across accelerators. Low-bit quantization offers a natural remedy, but transformer activations are often heavy-tailed and outlier-dominated, making simple quantization highly lossy. We show that this difficulty is not only a property of the quantizer, but also of the architecture. Specifically, residual connections can drive transformer activations away from Gaussianity during training. Using controlled comparisons between residual and residual-free transformers, we demonstrate that this effect leads to substantially higher quantization error and accuracy degradation at low precision in residual models. We explain the phenomenon through an excess kurtosis analysis, showing that residual mixing can amplify non-Gaussianity, whereas dense mixing in residual-free contracts non-Gaussianity. We then show that residual-free transformers can be made trainable using orthogonal initialization, spectral or second-order optimization, and depth-aware scaling of attention temperature. In language tasks, while there is a small drop in full precision performance, these models retain near-Gaussian activations and exhibit significantly improved robustness to low-bit quantization. Our results identify an accuracy--compressibility trade-off in transformer design and motivate architecture-level approaches to quantization-friendly foundation models.

We present the first end-to-end demonstration of fine-tuning and serving Google's Gemma 4 31B model on TPU hardware, providing an empirical comparison of TPU and GPU platforms for large language model adaptation. Using LoRA on a Google TPU v5p-8 for training and TPU v6e-8 (Trillium) for inference, we document the full set of code-level adaptations required to port a GPU-native training recipe, built on PyTorch, HuggingFace TRL, and FSDP, to the JAX + Tunix/Qwix stack. These adaptations span mesh configuration, LoRA module naming conventions, sharding annotation corrections, gradient checkpointing, data pipeline restructuring, and a custom Orbax-to-safetensors checkpoint merging procedure. For inference, we detail the vLLM-TPU Docker setup necessary to serve Gemma 4 on v6e-8 and characterize the resulting latency and throughput profile. Compared with a 2xH100 GPU baseline under identical hyperparameters, TPU training completes 1.61x faster at 2.12x lower cost. Inference throughput is within 3% across platforms, while TPU achieves 2x lower time-to-first-token (235 ms vs. 475 ms). Together, the TPU configuration is 1.82x cheaper for a representative train-plus-service workload. Our work removes a critical gap in the open tooling ecosystem and provides practitioners with a reproducible, production-ready recipe for Gemma 4 deployment on TPU infrastructure.

Recent advances in few-step diffusion distillation have enabled efficient image generation, yet aligning these models with human preferences remains challenging. We propose Reward-Tilted Distribution Matching Distillation (RTDMD), a two-stage framework that unifies distribution matching distillation with reward-guided reinforcement learning for few-step flow generators. We show that minimizing the KL divergence to a reward-tilted teacher distribution naturally decomposes into a distribution matching term and a reward maximization term. In the first stage, we introduce Ambient-Consistent Distribution Matching Distillation (AC-DMD), which performs subinterval-wise distribution matching and augments the fake score objective with a consistency regularizer to help the fake score model track the shifting generator distribution under limited updates. In the second stage, we jointly optimize both terms: for the reward maximization term, we derive a hybrid policy gradient that combines a GRPO-style estimator for the stochastic intermediate transitions with direct reward backpropagation through the deterministic final step, and further introduce step-subset GRPO (SubGRPO) to reduce variance. Experiments on SD3, SD3.5, and FLUX.2 demonstrate that RTDMD establishes new state-of-the-art results across preference, aesthetic, and compositional metrics with only 4 inference steps, outperforming previous few-step text-to-image generation methods. Code and models are available at https://github.com/Harahan/RTDMD.

Linux is the foundation of the digital age, accounting for the majority of the cloud and mobile OS markets. Any device that runs Linux uses the Linux page cache, a central pillar in OS and application performance, serving to reduce extraneous disk access. Many page cache eviction policies have been developed but remain bound by the rigidity of heuristics. The rise of AI-driven tools in recent years, melded with the ever-increasing variety of workloads for Linux devices, sets the stage for machine-learning-driven cache eviction policies. Promising research has been done in this field, but only in the field of user-space applications such as CDNs. We develop LearnedCache, an eBPF-integrated single-layer perceptron-based cache eviction policy for the Linux page cache, trained on real kernel data from diverse workloads. We demonstrate median AUCs of nearly 80% over multiple linear models modeling page reuse time, then take a step further by embedding these models inside the Linux kernel for real-time performance evaluation. Through statistical testing over 50 paired trials against a baseline of FIFO for each workload, LearnedCache reveals that machine-learning-derived cache eviction policies are practical in the Linux kernel under representative empirical workloads and are able to surpass conventional FIFO by statistically significant margins of up to 10% in insertion rate, a frequency-adjusted derivation of cache hit rate, in specific workloads while incurring minimal overhead.

The deployment of Large Language Models (LLMs) and Vision Transformers (ViTs) on edge devices is significantly constrained by memory limitations and the critical timing bottlenecks introduced by dense Multiply-Accumulate (MAC) arrays. In the ultra-low bit regime, logarithmic Power-of-Two (PoT) quantization provides a hardware-efficient alternative by replacing MAC operations with bit-shifts. However, the non-uniform exponential lattice is inherently limited by a Low Angular Resolution Regime, a structural flaw that becomes particularly pronounced at sub-4-bit thresholds, leading to a notable degradation of high-dimensional feature manifolds. To address this geometric limitation, we propose Orthogonal Residual Projection (ORP), an algorithm-hardware co-design framework. By formulating quantization as a dual-basis geometric projection, ORP adaptively synthesizes a higher-resolution residual lattice using strictly shift-and-add operations. Furthermore, ORP's analytical solver offers a practical alternative to computationally intensive gradient-based optimization, reducing the full-model calibration time for LLaMA-2-7B to approximately 15 minutes. Extensive evaluations demonstrate ORP's applicability across modalities and its hardware efficiency. Under the 3-bit (W3/A16) constraint, ORP achieves a perplexity of 6.10 on LLaMA-2-7B, comparing favorably to conventional MAC-intensive baselines like AWQ without relying on asymmetric scaling, while maintaining competitive accuracy in 4-bit scenarios. At the silicon level, standard-cell RTL synthesis at a 28nm node indicates that ORP effectively mitigates the timing bottlenecks associated with dense multiplier trees.

The convergence of large language models (LLMs) with 6G networks is fostering a paradigm of autonomous multi-agent cooperation, which in turn is expected to substantially increase east-west traffic. Although latent-space interaction mechanisms can enable more efficient collaboration than symbolic natural-language (NL) exchanges, prior work often abstracts away the associated communication overhead under practical wireless constraints. In embodied multi-agent settings, heterogeneous interaction media incur disparate inference and transmission costs, thereby inducing an inherent end-to-end (E2E) latency trade-off. To address this, we propose a joint design that integrates communication-media selection with wireless resource allocation. Through analytical characterization and simulation-based evaluation, we show that neither token-based transmission nor key-value (KV) cache-based transmission is uniformly optimal across operating regimes, as performance depends critically on system parameters such as available computational resources and channel conditions. Accordingly, we formulate a joint optimization problem aimed at minimizing the E2E latency of multi-agent collaboration and develop a low-complexity joint media selection and resource allocation (JMSRA) algorithm. Numerical results further confirm that, by adaptively coordinating the interaction media and bandwidth allocation over heterogeneous links, the proposed scheme achieves markedly reduced E2E latency relative to conventional NL-only and KV-cache-only baselines, enabling efficient and robust multi-agent collaboration in future wireless networks.

Large Language Models (LLMs) are increasingly expected to operate over long contexts, yet standard softmax attention incurs a KV cache that grows linearly with sequence length, quickly becoming the bottleneck for long context inference. A practical remedy is to evict less important KV entries; however, existing eviction policies are largely heuristic and struggle to capture the rich, input-dependent distribution of token importance. In this work, we introduce a learnable indexer that predicts KV importance, enabling more accurate retention of critical tokens. Meanwhile, naively evicting tokens permanently discards their information, leading to irreversible forgetting and degraded retrieval over long ranges. To address this, we propose a lightweight latent memory module that compresses evicted tokens into a compact, online-updated state and provides residual readouts to compensate for the attention contributions lost through KV eviction. Collectively, our method enables accurate long-context inference under a bounded KV budget, delivering consistent improvements on RULER (4K/16K) across Qwen, Mistral, and Llama models (up to 25 points under aggressive eviction), markedly more stable Needle-in-a-Haystack retrieval, and superior LongBench scores and compression curves compared to existing eviction policies.

Multi-agent tool calling is becoming the dominant interaction pattern for LLM-based systems, yet existing inference frameworks treat each tool call as an independent request, re-processing the entire conversation from scratch even though 85-95% of the prompt is unchanged from the previous turn. We present a stateful inference architecture that converts the $O(n_t)$ per-turn cost of conventional serving into an $O(Δ_t)$ delta-only cost: a persistent KV cache lives across turns and advances by ingesting only the new tokens, while a radix prefix cache extends this across interleaved multi-agent traffic and a prompt-lookup speculative decoder accelerates structured output. Against vLLM and SGLang on novel, fully-generated workloads, the reference implementation is $2.1\times$ faster per turn on a 6-turn agentic workflow and $4.2\times$ on the median turn of a 35-turn one, halving end-to-end wall time. The advantage comes from stateful reuse and speculation, not caching.

Chunk-wise autoregressive video diffusion models rely on a KV cache of previously generated chunks to avoid redundant computation, but this cache quickly becomes a memory bottleneck as videos grow longer. Methods that quantize the KV cache to low bitwidths reduce memory pressure but degrade video quality. We show that a key driver of this degradation is a systematic bias in attention weights: due to the convexity of the exponential in softmax attention, quantization noise inflates the contribution of cached keys, a phenomenon we call the Jensen bias. This effect causes quantized keys to steal attention mass from the unquantized current chunk. We derive a per-attention-score correction that removes this bias in expectation, computed on the fly from the quantization step sizes of the cached keys and the query norm. Using a second-order Taylor approximation, the additional computational overhead is negligible, and no additional memory is needed alongside the cache. Evaluated on MAGI-1, SkyReels-V2, and HY-WorldPlay at INT2 quantization, our correction recovers most of the quality lost to aggressive quantization, reaching near-BF16 video quality, and can outperform INT4 quantization while using 50% less memory.

Diffusion Transformers require repeated denoiser evaluations during iterative sampling, making inference computationally expensive. Cache-based acceleration reduces this cost by reusing intermediate representations across denoising steps, but can introduce representation deviations and degrade generation quality. In this paper, we analyze these deviations and show that effective calibration should consider both the direct mismatch caused by reuse and the subsequent trajectory shift induced by earlier corrections. To address this challenge, we propose Trajectory-Consistent Calibration (TCC), a training-free method that calibrates cached representations toward their full-computation counterparts. Specifically, rather than estimating all calibration priors from a single uncorrected cache trajectory, TCC uses an offline iterative procedure so that each prior accounts for the trajectory shift induced by preceding calibrations. Experiments on PixArt-alpha and DiT-XL/2 show that TCC consistently improves FID across representative cache-based acceleration methods while preserving their underlying reuse policies. Notably, in a representative PixArt-alpha cache-acceleration setting based on FORA, TCC reduces FID from 29.83 to 27.35, slightly surpassing the full-computation baseline.

Long-horizon LLM inference turns the key--value (KV) cache into the dominant GPU memory consumer and makes per-token attention increasingly expensive. Many common eviction policies use static recency windows or historical attention, leaving unused a signal computed on every decoding step: the model's current uncertainty. We introduce CONF-KV, a KV-cache manager that converts the next-token distribution into a scalar confidence score and uses it to choose the per-step cache budget, retaining more context when the model is uncertain and pruning aggressively when it is confident. Within each budget, tokens are ranked by a composite of accumulated attention mass and recency, while a protected recent window preserves local coherence. We combine the policy with blockwise online-softmax attention, mixed FP16/INT8 storage, and a pyramidal per-layer budget variant. Across four model families and generated lengths up to 4K, CONF-KV stays near the footprint of a fixed 512-token sliding window while remaining within 1.5--2.1 perplexity points of full KV. On Needle-in-a-Haystack up to 32K tokens, CONF-KV reaches 91.4% retrieval accuracy versus 53.8% for sliding windows and 80.6% for H2O; on 75 VisualWebArena tasks it retains 95.3% of full-KV success at 2.8 times lower peak memory.

Large language model (LLM) serving is fundamentally limited by inefficient hardware utilization. Autoregressive (AR) decoding underutilizes GPUs due to its strictly sequential execution, while diffusion LLMs (DLLMs) improve throughput by decoding multiple tokens per iteration. However, fixed block-size diffusion decoding exhibits strong load sensitivity: large blocks exploit idle GPU resources under low load, but saturate early and incur substantial redundant computation under high load. As a result, throughput gains vanish beyond saturation, and no single decoding granularity performs well across dynamic serving workloads. We present Optimus, a serving system that enables elastic decoding for diffusion LLMs by dynamically adapting decoding granularity to runtime load. The key idea is to treat decoding granularity as a runtime control variable, balancing GPU utilization and token efficiency. Optimus combines chunked decoding, which enables fine-grained execution without retraining, with saturation-aware scheduling, a closed-loop mechanism that selects chunk sizes based on runtime conditions. Together with system-level optimizations and customized attention kernels, Optimus achieves significant performance improvements while preserving model accuracy. Experiments show that Optimus delivers up to 6.1x throughput improvement over AR decoding and 4.3x improvement over fixed-block diffusion LLM, while maintaining stable performance across diverse load regimes and improving end-to-end serving capacity under latency constraints. The source code is available at https://github.com/dubcyfor3/Optimus.

Deploying deep neural networks on resource-constrained 6G edge devices demands aggressive compression with minimal accuracy loss. Quantization-Aware Training (QAT) has emerged as a leading compression approach; however, existing mixed-precision methods typically operate at coarse layer- or channel-level granularity. These methods often rely on heuristic or search-based bit-allocation strategies, which may overlook fine-grained variability at the neuron level. We propose Neuron-Level Mixed-Precision QAT (NMP-QAT), where each neuron independently learns its own discrete precision during training. Starting from low-bit precision, NMP-QAT expands bit-width only when training signals demand it, via differentiable surrogates and straight-through estimators, while preserving a fully discrete inference graph. This adaptability extends to both weights and activations, reducing memory movement. Evaluated on telecom and non-telecom datasets across MLP and tabular foundation model architectures, NMP-QAT achieves superior compression-accuracy trade-offs over mixed-precision QAT baselines, making it well-suited for Green AI deployments at the network edge.

Large Language Models (LLMs) are widely used by our increasingly digitalized society, but raise sustainability, performance, and financial concerns, especially as inference workloads grow. To improve the design and operation of LLM ecosystems, we envision simulators and simulation-based digital twins becoming primary decision-making tools. LLM ecosystems leverage many heterogeneous components, making simulation a non-trivial, yet critical operation. The simulation challenge is exacerbated by the absence of a comprehensive reference architecture of LLM ecosystems; the lack of such a conceptual model can be costly and could misguide the designers and engineers. Without a reference architecture, even the most experienced stakeholders could tinker in researching, engineering, or maintaining LLM ecosystems. In this work, we bring a three-fold contribution to the scientific community. Firstly, we synthesize, propose, and validate a reference architecture (RA) of LLM ecosystems under inference. Then, adhering to the reference architecture, we design Kavier, the first simulation instrument able to predict the performance, sustainability, and efficiency of LLM ecosystems under inference, through discrete-event and cache-aware simulation, focusing on Key-Value-(KV-)Caching and prompt prefix caching policies. Through experiments with a Kavier prototype and real-world traces, (i) we measure the accuracy of Kavier and its performance in massive-scale simulations, (ii) we compare the performance of different KV-Caching policies, and (iii) we analyze the performance, sustainability, and efficiency of LLM ecosystems under various prefix caching policies. Overall, we show that Kavier enables operators, researchers, and engineers to predict LLM ecosystems in a time, performance, and cost-efficient way.

We apply the influence-adaptive Walsh geometry of a companion theory paper (arXiv:2605.01637) to extreme low-bit weight-only LLM quantization. The recipe is one math-invariant transformation: WHT-rotate each linear layer's weight matrix and rescale its columns by per-coordinate Walsh-basis activation energy before handing off to a reconstruction-error quantizer (Intel auto-round). This biases per-group integer rounding toward high-spectral-energy channels. On four pretrained decoder-only models from 135M to 1.5B parameters, BBT-spectral reduces wikitext-2 perplexity by 15-58% relative to vanilla auto-round at W2A16; we also report a TinyLlama-1.1B auxiliary data point. Three extensions transfer the recipe to families it failed on: a per-head PCA matrix-Gamma replacement of q_norm/k_norm for Qwen3 attention (PPL 136.76 -> 88.99 on Qwen3-0.6B); an SO(2) per-pair rotation that commutes with RoPE (PPL 36.93 -> 21.84 on Qwen2.5-1.5B); and an MoE-aware input-side absorption fix identified by architectural fuzzing of Laguna-style fused-expert layouts. A W2-vs-W4 ablation gives a deliberate negative control: the redistribution payoff falls within the +/-0.5 PPL noise floor at W4, consistent with the Schur-convexity intuition that the cost of unconcentrated influence vanishes as the noise budget shrinks. All quantized weights export to OpenVINO IR and run on Intel NPU + Arc dGPU + CPU with PPL invariant to device within +/-0.1. We do not claim a formal Boolean-to-real-valued transfer of the theory paper's majorization argument: the WHT activation energy used here is not the Boolean influence of the theory paper, the link is intuitive, and the contribution is engineering value rather than a transferred theorem. Head-to-head benchmarks against SpinQuant, QuaRot, QuIP-sharp, AQLM, OmniQuant, and ButterflyQuant at matched calibration are the main future-work item.

We study the rate-distortion limits of online KV cache compression in autoregressive language models, formulating it as sequential Wyner-Ziv source coding on the filtration induced by the model, with the next-step query as decoder side information. Empirically, across four models spanning two families and $0.5$-$3$B parameters, we find that the next-token distribution's sensitivity to context truncation decays polynomially rather than geometrically: a power law improves on an exponential fit by an order of magnitude in extrapolation, the fitted exponent is recovered independently from a sink-plus-recent KL measurement, and the decay is verified to be free of positional-encoding artifacts by a position-preserving ablation. Under a corresponding polynomial truncation-sensitivity assumption, our main result characterizes the per-token memory requirement of suffix-only cache policies: a sliding-window scheme attains distortion $\varepsilon$ with window $w = O(\varepsilon^{-1/α})$, and — under an additional two-sided Bayes-risk condition — a converse shows $w = Ω(\varepsilon^{-1/α})$ is necessary within this policy class, so the scaling is $Θ(\varepsilon^{-1/α})$ for suffix-only policies. Whether recurrent or propagating cache summaries can beat this scaling is left open. An explicit block-Markov scheme achieves the upper bound; its rate-of-convergence exponent matches the converse under additional forward-decay and regularity hypotheses (not implied by truncation sensitivity alone), and differs by a factor of two otherwise. Empirically, the polynomial law predicts the degradation curves of concrete cache policies: recency-based eviction (sliding, sink-plus-recent) suppresses distortion by roughly two orders of magnitude over random retention at equal budget, with a power-law decay in the budget.

Large Language Model (LLM)-based agents demonstrate strong reasoning and execution capabilities on complex tasks when guided by structured instructions, commonly referred to as workflows. However, existing workflow-assisted agent serving systems typically rely on predefined templates and shallow matching mechanisms, which limit their ability to capture deep semantic relationships and generalize to previously unseen tasks. To address these limitations, we propose a new workflow management paradigm that represents workflows using a unified graph, termed wGraph, where each node corresponds to an atomic operation. wGraph serves as a shared substrate from which task-specific workflows are dynamically instantiated. Building on wGraph primitives, we introduce GraphFlow, a system that efficiently integrates workflows into agent serving through two key designs. First, adaptive workflow generation dynamically constructs workflows from wGraph based on task semantics and constraint requirements. Second, workflow state management exploits wGraph structure to efficiently manage Key-Value (KV) caches, reducing redundant computation during agent serving. Extensive experiments across five benchmark datasets show that GraphFlow consistently outperforms state-of-the-art methods, yielding an average performance improvement of approximately 4.95 percentage points, while achieving an approximately 4$\times$ reduction in memory footprint.

Scientific machine learning often requires combining known physics with unknown parameters or correction terms learned from data. Existing approaches either ignore known structure, encode it as a soft penalty, or require hand-written PyTorch code for each equation. We present The Neural Compiler, a system that translates programs written in a first-order Scheme-like expression language into frozen, differentiable PyTorch modules. These modules match the source program to floating-point precision and provide gradients through autograd. In hybrid models, the compiled module encodes known physics exactly while learned components model the unknown remainder. We evaluate the compiler across six experiment domains: Feynman physics equations, Lotka-Volterra dynamics, a damped pendulum, a one-dimensional heat equation, three-dimensional vector mechanics, and compositional generalization. Compiled modules match hand-coded PyTorch implementations numerically for single equations, showing no accuracy loss from compilation. With only 1 to 4 trainable parameters, compiled models recover physical constants to less than 1 percent error in most cases, while standard PINN baselines with more than 8500 parameters show 7 to 93 percent error. Compiled modules also compose with zero error, while neural approximations can accumulate large errors in deep composition chains. The main value of the compiler is not improved accuracy over hand-coded equations, but systematic composability: it generates correct, differentiable modules from symbolic specifications without rewriting each equation by hand. The system supports 51 primitive operations, including vector and matrix algebra, enabling PDE discretizations and hybrid scientific models. This string-in, module-out interface also provides a natural target for large language models that translate scientific descriptions into executable differentiable modules.

The KV cache used in large language models has linearly growing time complexity, so LLMs face memory blow-up and reduced decoding efficiency when they process long contexts.Current KV Cache eviction has become an important research direction; however, existing methods based on fixed Soft Tokens (e.g., Judge Q) rely on a static parameter set as the query to evaluate the importance of KV pairs, so they cannot adapt dynamically to different input prompts, and they cannot precisely capture complex and changing task relevance.Also, evicted KV pairs are discarded permanently, so this causes irreversible information loss and context breaks. To address this problem, we propose Meta-Soft, a dynamic compression framework based on probe-driven context integration. Specifically, we build a meta-library with a learnable orthogonal basis matrix $\mathcal{L}$, and we use a selector network with Gumbel-Softmax to produce differentiable sparse combination weights, so we dynamically synthesize the most targeted $k$ Soft Tokens from the input prompt features.We append these Soft Tokens to the end of the input sequence to probe key information. We also introduce an attention-flow based integration mechanism, which redistributes the semantic information of removed tokens into retained tokens, and this keeps the dropped context information effectively.Experiments on multiple datasets show that our method outperforms existing state-of-the-art eviction methods and provides a new solution for KV Cache compression.

Recent progress in LLM reasoning has increasingly shifted from single-pass generation to explicit search over intermediate reasoning states. Tree-of-Thoughts (ToT) organizes inference to tree-structured search with branching and backtracking, but it substantially amplifies the Key--Value (KV) cache: retaining KV states for a frontier of partial trajectories quickly becomes a memory bottleneck that limits throughput and constrains search depth and width under fixed hardware budgets. We address this challenge by observing that KV reuse in ToT-style inference is governed by search dynamics: near-term decoding depends primarily on the active branch and its ancestors, whereas inactive subtrees have low short-term reuse probability yet must remain recoverable for backtracking. Motivated by this, we propose ArborKV, a structure-aware eviction framework that couples a lightweight value estimator with a tree-aware allocation policy, and performs purely token-extractive eviction with lazy rehydration to support revisits. Experiments on ToT-style reasoning benchmarks show that ArborKV achieves up to ~4x peak KV-memory reduction while preserving near-full-retention accuracy, enabling larger search configurations under fixed device budgets that would otherwise run out of memory.

Long streaming video QA remains challenging due to growing visual tokens and limited reasoning length of large language models (LLMs). KV-caching stores the Key-Value (KV) of the historical tokens via LLM prefill and enables more efficient streaming QA. However, existing methods cache every one or two frames, causing redundant memory usage and losing fine-grained spatial details within frame or temporal contexts across frames. This paper proposes MuKV, a method that features a multi-grained KV cache compression module and a semi-hierarchical retrieval approach to improve both efficiency and accuracy for long streaming VideoQA. For the offline KV cache, MuKV extracts visual representations at patch-, frame-, and segment-levels. The multiple levels of granularity preserve both local cues and global temporal context, while maintaining efficiency with a dual signal token compression mechanism guided by self-attention and frequency. For online QA, MuKV designs a semi-hierarchical retrieval method to retrieve relevant KV caches for answer generation. Experiments on long-streaming VideoQA benchmarks show that MuKV significantly improves answer accuracy, without sacrificing memory and online QA efficiency. Moreover, our compression mechanism alone brings consistent benefits across answer accuracy, memory, and QA efficiency over baselines, showcasing highly effective contribution.