The paper introduces a method for quantizing the fixed‑size recurrent states of hybrid language models to as few as four bits per token. By deriving distortion weights from the observability Gramian and combining them with normalized state ranges, the authors achieve mixed‑precision bit allocation without requiring calibration data, rotation, or additional training. The approach also logarithmically quantizes decay rates, yielding significant reductions in excess negative log‑likelihood—up to 27.9× better than seven baselines—while maintaining near‑FP32 performance at six bits.
By Hongren Chen, Jiayang He
The paper introduces effective depth (Deff), a scalar diagnostic that treats a transformer’s layer‑wise residual stream as a discrete‑time process and measures how representation similarity decays with layer distance. Across sixteen decoder‑only language models, Deff reveals that most models exhibit a lower similarity decay than the closed‑form reference, indicating correlated residual updates rather than unused depth. The study also shows that this effect is robust to various controls and persists early in training, suggesting Deff is a global accumulated‑state diagnostic rather than a capability score.
By Barak Gahtan, Ido Galil, Alex M. Bronstein
The paper introduces a reproducible benchmark for evaluating attention mechanisms in tabular foundation models, focusing on the distinct row and column attention patterns that differ from language model attention. It compares several backends—Torch SDPA, FlashAttention variants, vLLM, and SageAttention—across realistic tabular shapes on A100, H100, and B200 GPUs, revealing that optimal backend choice varies by attention type, hardware, and model specifics. The study finds FlashAttention generally performs best, but CuDNN can outperform it for column attention on longer sequences, while SageAttention excels for large row sequences beyond 16k rows.
By Maximilian Schambach, Clemens Biehl, Sam Thelin
The paper investigates how different adaptation strategies, model architectures, parameter scales, and quantization settings influence the performance, efficiency, and robustness of large language models (LLMs) for log anomaly detection. Across three public log datasets, the study finds that adaptation strategies lead to significant performance variations, model scaling offers dataset‑dependent gains, and models with similar accuracy can differ markedly in computational cost. Low‑bit quantization largely preserves detection performance, and the authors also assess robustness to structural, semantic, and label noise at varying perturbation levels.
By Bin Li, Dongdong Wang, Siyang Lu
The paper discusses position‑independent KV cache reuse, a technique designed to cut latency in retrieval‑augmented generation by reusing chunk‑level KV caches across prompts. It argues that current evaluation methods overstate the accuracy of such reuse because they do not accurately capture the loss of accuracy, and that existing datasets lack the necessary reuse dynamics for thorough testing. To remedy this, the authors propose a new evaluation methodology that unambiguously measures accuracy loss and introduce Boxoffice, a tool that programmatically creates datasets with challenging KV cache reuse patterns.
By Samuel Cestola, Tianxiang Xia, Pengfei Zheng, Weiyan Zheng, Bo Wang, Yi Zhao, Diego Didona
The paper investigates the generalization behavior of the OPTQ quantization algorithm and its stochastic variant. It derives bounds on the expected squared error when a test point is drawn from a fixed distribution, linking this error to the calibration dataset and to the regularization parameter λ. The authors use these theoretical insights to propose a new recommendation for choosing λ, which shows improved performance in experiments compared to previous suggestions.
By Erin George, Rayan Saab
The paper argues that reporting scale‑invariant statistics such as cosine similarity without their noise floor is misleading when evaluating interpretability transfer from full‑precision to quantized neural networks. It derives a closed‑form expression for the expected cosine similarity based on a dimensionless parameter κ = n
ho^2/d, measures the class separation ρ on real activations, and shows that a reported cosine of 0.996 between full‑precision and INT4 models cannot be interpreted as preservation without knowing the sample size n. The authors demonstrate that at INT4 the direction of the interpretability artifact rotates beyond the estimator’s own noise, while at INT8 no significant movement is detected, and they highlight that scale‑invariant metrics cannot distinguish between translation and attenuation of a transferred decision variable.
By Pranav Varshney
LUMO (Lightweight Unified Multilingual Orchestrator) is a privacy‑preserving offline voice assistant that runs entirely on edge hardware, specifically a Raspberry Pi 5 with 8 GB RAM. It integrates local ASR, a 4‑bit GGUF‑quantized LLM, and TTS to deliver end‑to‑end response latencies of 2.0–4.0 s, a 6.8 % WER on short English utterances, and lower peak power consumption (~9 W) compared to existing edge assistants. The system also supports Bangla speech, enabling multilingual use in low‑resource settings.
By Md. Mehedi Hasan Naeem, Mst. Kamrunnahar Ruma, Nafiza Anjum, Shakila Sultana, Md. Sujan Ali
The paper introduces Distributed Learning as a Service (DLaaS), a platform that lets developers launch distributed/federated learning jobs through a single admin dashboard. It offers declarative options such as Differential Privacy, Split Learning, Hierarchical Aggregation, and Knowledge Distillation without requiring changes to client code. The authors demonstrate the full service lifecycle on an industrial smart‑home Wake‑up Word task using the Ok Aura dataset, showing live operation across Android clients and Dockerized aggregators, and releasing the source code and video walkthroughs.
By Tianyue Chu, Filippo Vannella, Dimitra Tsigkari, Paula Delgado-Santos, Fernando L\'opez, Pablo Gomez Guerrero, Sotirios Spantideas, David Solans Noguero
The paper introduces the Neural Action Codec (NAC), a convolutional encoder‑decoder architecture that treats short robot action trajectories as multi‑channel 1D signals and compresses them using a multi‑scale residual vector quantization (RVQGAN) model. NAC replaces traditional discrete action tokenizers with a compact, ordered token space via offset codebooks, allowing standard autoregressive policies to operate over short, structured sequences while a Vocos‑style decoder reconstructs the actions. Experiments on LIBERO‑10, RoboMimic, and real‑world manipulation tasks show that NAC achieves higher reconstruction fidelity and better average success rates than existing binning, FAST, and VQ‑based tokenizers at comparable or improved compression rates.
By Ahad Jawaid, Yu Xiang
The paper argues that for large‑context autoregressive language‑model inference, memory bandwidth—specifically the Key‑Value (KV) cache—becomes the limiting resource rather than arithmetic throughput. It analytically derives how arithmetic intensity decays with context length for NVIDIA H100, NVIDIA B200, and AMD MI300X, identifies crossover points where KV traffic overtakes weight traffic, and evaluates representative techniques across five compression domains. The study finds a three‑regime behavior: below the crossover, weight traffic dominates and KV compression offers little benefit; beyond it, KV traffic dominates and compression methods trade quality for bandwidth, with paging and prefix sharing being lossless but capacity‑limited, while quantization and eviction directly reduce bandwidth at the cost of accuracy.
whyItMatters":"The work provides a unified analytical framework and a standardized protocol that enable consistent comparison of KV‑compression techniques across hardware and workloads, guiding practitioners in selecting appropriate methods for long‑context inference."
By Tejinder Singh
EAServe introduces an encode-aware disaggregated serving framework for multimodal large language models (MLLMs), restructuring the traditional Prefill-Decode pipeline into a three-stage Encode-Prefill-Decode (EPD) system. By treating Encode as the control point, EAServe coordinates load‑adaptive micro‑batching, rate‑controlled offloading to prefill workers, and dynamic SM partitioning to balance GPU utilization across stages. Its Hybrid Auto Selection (HAS) layer optimizes GPU allocation, encode batch size, and offload ratio using capacity profiling and Bayesian optimization, achieving up to 4.3× higher goodput compared to NVIDIA Dynamo and 1.7× higher than vLLM on various MLLM architectures.
By Kunxiong Zhu, Zhihao Shu, Hangyu Zheng, Minghai Qin, Miao Yin, Gagan Agrawal, Wei Niu
TeDiServe is a cluster‑level serving system designed for diffusion language models (DLMs). It addresses DLM‑specific challenges such as the speed‑quality tradeoff from confidence‑based denoising, variable parallelization under fluctuating load, and non‑uniform per‑step costs from approximate KV caching. By employing deadline‑aware scheduling, adaptive load control, and a quality‑aware optimization for cluster reconfiguration, TeDiServe achieves up to 56.6 percentage points higher SLO attainment and reduces end‑to‑end latency by up to 46% with less than 1% accuracy loss.
By Tzu-Tao Chang, Benjamin Yuanyang Hong, Kiet Pham, Shivaram Venkataraman
The paper investigates why per-layer pruning of IoT intrusion detectors can cause severe class-level failures. On the CICIoT2023 dataset, a two-layer convolutional detector pruned at 80% sparsity loses 16 accuracy points and half its macro‑F1, with 17 of 34 classes heavily damaged. The failure is traced to the first layer’s weight starvation, and the authors show that protecting those 192 weights or pruning globally, as well as recomputing normalization statistics, can prevent or repair the collapse.
By Md Anas Biswas
The study investigates what knowledge a student model inherits from its teachers beyond accuracy when using knowledge distillation for encrypted‑traffic classification. By distilling a 101k‑parameter student from two teachers of equal accuracy but different construction, the authors test ten hypotheses over a year of real TLS traffic, finding that unknown‑traffic detection and shortcut reliance can transfer depending on temperature settings and model size, while other abilities do not. The results show that distillation can propagate teacher habits, but some inherited capabilities can also be achieved without a teacher.
By Mahmoud Abbasi
Tetra introduces a new Leech‑lattice based codebook that reduces the memory footprint of quantized LLMs to about 2.15 bits per weight, enabling efficient 2‑bit quantization without a massive lookup table. The method employs a 64‑state Golay trellis and a shared 16 KiB table, decoding each 24‑weight block with only six table loads and two small lookups. When applied to Qwen3 models (4B, 8B, 14B), Tetra achieves 2.70–2.73 bits per parameter, scoring 63–75 on MMLU and generating 57–114 tokens per second, while maintaining close performance to 4‑bit AWQ and outperforming llama.cpp’s IQ2_XXS on 4B.
The paper investigates how hybrid attention mechanisms—combining full softmax attention with recurrent alternatives—affect multilingual language models, especially for long sequences and poorly tokenized languages. Interpretability analysis reveals that cross‑lingual representations form patterns linked to the ordering of recurrent and full‑attention layers, with a notable spike in alignment after the first full‑attention layer. Distillation experiments show that alternative layer orderings consistently outperform the standard arrangement, achieving up to 2.5× faster learning, suggesting that starting with a full‑attention layer may benefit multilingual models.
The paper investigates post‑training quantization of transformer attention blocks by optimizing a joint loss over the Q, K, V projections rather than individual weight matrices. Using this joint attention‑based objective (JAB), the authors achieve significant compression on Mistral‑7B, recovering 77‑90% of the performance gap at 3 bits, but the method fails when MLP layers are included. A role‑aware offset rule that ignores sensitivity estimates outperforms JAB on GPT‑2 and full Mistral‑7B, demonstrating that the matrix a weight belongs to is more critical than sensitivity metrics.
The paper introduces a training‑free visual token pruning strategy for vision‑language models that separates early vision‑guided pruning from later text‑guided reselection. By first pruning tokens with vision‑encoder attention, retaining candidates until the decoder midpoint, and then applying text‑to‑visual attention, the method preserves task‑relevant visual information. Across eight benchmarks and three models, it achieves an average performance recovery of 11.10 and 16.84 percentage points at 80% and 90% pruning, respectively, while maintaining comparable or lower LLM‑prefill latency.
PMOPD introduces a projection-based approach to multi-teacher on-policy distillation, addressing the capability seesaw problem by constructing subspace memories from task-specific parameter displacements and projecting gradients and optimizer updates to avoid cross-task interference. It also includes a lightweight conflict probe for task interaction analysis, a task ordering strategy, and a cycling mechanism to balance subspace estimation and task revisitation. Experiments on Code, Reason, and Math tasks demonstrate that PMOPD improves all evaluated capabilities, raising average scores by 2.54 points on Qwen2.5-7B and 2.09 points on Llama-3.1-8B.