Quantization
-
Structured Transforms for Low-Overhead Quantization of Language Models
2026-09-10 15:15:50Daria Cherniuk, Alexander Rudikov, Boris Kashin, Ivan OseledetsAbstract
We revisit Kashin-decomposition-based weight quantization for large language models and propose an improved algorithm with stronger convergence properties and structured, efficient orthogonal transforms. The method retains the core factorization of each weight into two components -- one with bounded infinity norm and the other with bounded infinity norm after an orthogonal transformation -- but replaces the dense random orthogonal matrix with a sign-randomized Discrete Cosine Transform (DCT), reducing the per-iteration cost from $\mathcal{O}(N^2)$ to $\mathcal{O}(N \log N)$. The proposed greedy algorithm with alternating updates guarantees the four-peak distribution required for stable 2-bit clustering of each factor and admits closed-form initialization of cluster centers, removing the multi-restart k-means bottleneck of prior work. Composed with OPTQ-style sequential error compensation and QuIP-style incoherence preprocessing, the resulting JAX pipeline is competitive with OPTQ, QuIP, QuIP-RG and a fine-tuning- and vector-quantization-free variant of QuIP# at 4-bit per channel on OPT, Llama-2 and Pythia, with favorable wall-clock scaling. The bounded-$\ell_\infty$ factorization is also notably robust: on stress configurations where QuIP variants diverge to four-digit perplexity (Pythia-6.9B) or abort with NaNs in LDL back-substitution (Mistral-7B), Kashin-DCT remains numerically stable and stays close to FP16 baseline. At inference time, each weight decomposes into two 2-bit factor codes per channel that are structurally suited to native-2-bit hardware.
Abstract (translated)
URL
https://arxiv.org/abs/2609.11687
PDF
-
OmniKVQuant: KV Cache Quantization for Omni-LLMs
2026-09-10 14:14:12Suho Yoo, Hyunjong Ok, Jongmin Choi, Jihoo Jung, Joon Son ChungAbstract
As Omni-modal large language models (Omni-LLMs) take in audio, video and text together, their KV cache memory cost grows. KV cache quantization is the de facto approach in text-only LLMs, but its application to Omni-LLMs remains unexplored. In this paper, we analyze how TurboQuant, a representative rotation-based KV cache quantization method, behaves on multimodal caches and identify two critical issues: temporal key drift and heterogeneous value geometry. To address these, we propose OmniKVQuant, a training-free framework that (i) sets the key quantization range over each short window of the input stream; and (ii) rotates values separately per modality. On Qwen2.5-Omni and Qwen3-Omni, OmniKVQuant enables 2-bit KV caches while substantially preserving performance across seven audio-visual benchmarks. We further provide a fused Triton decode kernel that unpacks the 2-bit cache during attention, so no dense FP16 cache is ever built. Code: this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2609.11582
PDF
-
Forgetting Only What Matters: Layer-Selective Unlearning toward Robust LLMs
2026-09-09 16:50:41Ravi Ranjan, Olivera Kotevska, Agoritsa PolyzouAbstract
Large Language Models (LLMs) can memorize and reproduce sensitive, copyrighted, or otherwise undesirable training content, creating privacy, safety, and regulatory concerns. Machine unlearning offers a practical alternative to full retraining, but many existing methods apply broad or fixed parameter updates that can degrade utility and remain brittle under deployment changes such as post-training quantization, where forgotten knowledge may partially re-emerge. We propose Forgetting Only What Matters via Unlearning Layers (FOM-UL), a layer-level unlearning framework that selects transformer layers using a forget-to-retain significance score. This score identifies layers with high influence on the forget set and low sensitivity to the retain set, allowing FOM-UL to concentrate updates where they are most effective while leaving most of the model unchanged. This targeted update strategy improves the forgetting-utility trade-off and provides an empirical path toward quantization-resilient unlearning by reducing the chance that small, diffuse updates are erased by low-bit rounding. Across TOFU, KnowUnDo, and MUSE-style evaluations, FOM-UL reduces residual memorization compared with strong GA, NPO, KLD, SURE, ReLearn, and LUNAR-based baselines while preserving retain-set utility close to the vanilla model. Under 8-bit and 4-bit post-training quantization, FOM-UL maintains stronger memorization suppression and utility preservation than competing methods, and adversarial prompt evaluations show lower recovery of forgotten content. Overall, FOM-UL provides an efficient unlearning strategy that improves targeted forgetting, utility preservation, and deployment robustness without claiming formal guarantees of erasure.
Abstract (translated)
URL
https://arxiv.org/abs/2609.10439
PDF
-
Accuracy is Not Enough: A Divergence-Based Approach to Evaluate Fidelity Loss in Quantized LLMs
2026-09-09 08:50:36Shahzeb Qamar, Lorenz Sparrenberg, Christian Bauckhage, Baha Rababah, Carson Leung, Murat Kantarcioglu, Cuneyt Gurcan Akcora, Rafet SifaAbstract
Deployment of Large Language Models (LLMs) on memory-constrained edge devices relies heavily on aggressive post-training quantization. However, evaluating these models is largely based on zero-shot task accuracy, which depends solely on argmax predictions and is insensitive to changes in the underlying predictive distribution. Consequently, accuracy can exhibit unstable, non-monotonic behavior under progressive quantization, masking substantial fidelity loss relative to the BFloat16 (BF16) uncompressed base model and providing misleading deployment signals. We introduce a distribution-sensitive evaluation framework quantifying information loss in quantized LLMs as the divergence between full-vocabulary predictive distributions at the token decision boundary. We compute statistical distances, including Jensen-Shannon Divergence and Total Variation Distance, between outputs of full-precision and quantized models, enabling a fine-grained analysis of distributional shift. Using this framework, we quantify probability mass displacement and distributional drift relative to the BF16 reference, capturing predictive distribution changes not reflected in top-1 accuracy. We conduct a 120-run experimental matrix across five foundation architectures and four reasoning benchmarks under progressive quantization regimes, from uncompressed BF16 to Q2_K, providing a systematic fidelity analysis. Our results show divergence metrics generally increase under stronger quantization, complementing task accuracy with a fidelity signal. Across tested llama-cpp schemes, mixed-precision Q4_K generally yields lower divergence than uniform Q4_0 at similar memory footprints. These findings motivate distribution-aware evaluation as a practical diagnostic complement to task accuracy; they do not directly establish correctness, calibration, safety, or user-perceived quality.
Abstract (translated)
URL
https://arxiv.org/abs/2609.07664
PDF
-
Exploring Bottom-Up Clustering for Creating Semantic IDs
2026-09-08 06:32:57Leah Woldemariam, Sudhanshu Garg, Taha Belkhouja, Charles Kim-Yip, Ali SahamiAbstract
The success of generative retrieval has largely been attributed to the use of Semantic IDs, which improve over arbitrary item-level identifiers such as hashes by capturing the semantics of items. The main challenges faced when constructing Semantic IDs, however, is in mapping each identifier to a unique product and capturing information valuable to downstream tasks. Past works have appended additional codewords to de-duplicate item identifiers and utilized residual quantization to create hierarchical clusters. In this work, we present an algorithm for generating Semantic IDs that ensure the identifiers are both unique and preserve the structure of the original embedding. Key to our work is the use of bottom-up clustering to preserve local structure in the embedding space, improving the clustering quality of the resulting Semantic IDs and their utility for downstream generative retrieval.
Abstract (translated)
URL
https://arxiv.org/abs/2609.08310
PDF
-
A Measurement Study of LLM Inference Trade-offs Across Edge Continuum Hardware
2026-09-08 06:28:32Maysam Khatib, Moysis Symeonides, Demetris Trihinas, George Pallis, Marios D. DikaiakosAbstract
Large language models (LLMs) are increasingly used as backends for intelligent web services, but serving them across the edge continuum requires balancing quality, latency, model footprint, and energy. This paper presents a controlled measurement study of self-hosted LLM inference across edge and near-edge deployment nodes: an NVIDIA Jetson AGX Orin and a near-edge server with CPU-only and GPU-enabled inference modes. We evaluate multiple open-weight LLMs and quantization variants using a fixed question-answering workload, and compare them against GPT-4o as a cloud-hosted accuracy and latency reference. Our benchmarking pipeline reports accuracy, model footprint, per-token decoding latency, prefill latency, and overall execution energy. The results show that GPU-enabled server execution provides the lowest compute-side latency, while Jetson Orin shows lower measured energy, consistent with its lower platform power under our setup. CPU-only execution is consistently dominated in latency for our workload and shows higher measured energy. We also show that parameter count and downloaded weight-file size alone do not reliably predict observed accuracy or latency. Finally, using Pareto-frontier analysis, we study how deployment decisions may change under possible streamed-token delivery overheads, highlighting that compute-side inference metrics alone can lead to suboptimal placement for latency-sensitive interactive web services.
Abstract (translated)
URL
https://arxiv.org/abs/2609.08307
PDF
-
KBBQ: A Predictive Noise Law and the Limits of Spectrum Flattening in FP4 Quantization
Abstract
We develop a second-order theory of quantization noise in matrix multiplication in which the quantization format is characterized by the variance it assigns to each element. The constant variance profile of integer quantization recovers existing integer-noise theory, while the multiplicative profile of floating-point rounding reduces the data dependence to a scalar, the participation factor $\kappa$, yielding a closed-form signal-to-noise-ratio law. The resulting functional also admits a closed-form upper bound $\kappa^{*}$ that no function-preserving linear transform can exceed and that is attained by a recent state-of-the-art method. Building on this analysis, we introduce KBBQ (\textbf{K}appa-\textbf{B}raked \textbf{B}lockwise \textbf{Q}uantization), which parameterizes the extent to which a transform approaches this ceiling. At W4A4, across four base models and two FP4 formats, KBBQ outperforms the prior state of the art without additional deployment-time computation.
Abstract (translated)
URL
https://arxiv.org/abs/2609.08135
PDF
-
Quantization Amplifies Determinism, Not Bias: Scale-Dependent Behavioral Effects of Serving-Time Weight Compression
Abstract
Weight quantization largely determines the economics of serving open-weight LLMs. Its costs are usually assessed with capability benchmarks, on which 4-bit quantization of mid-sized models is often considered "nearly free." We examine a different question: when several answers are valid, does quantization change what a model chooses to say? We serve three checkpoints (Qwen3-8B/14B/32B) at three weight precisions (W4A16 AWQ, W8A16 FP8-Marlin, and bf16), holding the hardware, software, and sampling configuration constant, and collect approximately 71,000 completions paired by prompt and seed across two custom, leak-checked prompt batteries. We pre-specified the analyses in three waves in version control. At 8B, int4 reduces output diversity: the probability that two samples for the same scenario recommend the same brand increases by 5.1 percentage points (prompt-paired sign-flip test, Holm p = .023; reproduced at +4.4pp on a full regeneration of the arm), and lexical diversity falls substantially (TTR -0.011, standardized effect -0.51; robust to a length-controlled measure). At 14B and 32B, no content-concentration measure reaches significance; instead, stylistic drift emerges (em-dash rate +0.46/1k words at 14B and +0.61/1k at 32B, both Holm p <= .0024). Pre-specified tests of stereotype direction are null at every scale: outputs concentrate on the modal answer for each prompt rather than on stereotypical answers. Mechanistically, the token-level distribution becomes flatter (decision-token entropy +0.091 bits, p = .015) while the semantic distribution, measured directly from first-token log probabilities, becomes more concentrated (collision +2.6pp, p = .023): individual tokens become less predictable even as meanings become more repetitive. At 8B, the smallest size tested, AWQ-int4 serving measurably narrows the range of suggestions; audits should assess concentration as well as bias.
Abstract (translated)
URL
https://arxiv.org/abs/2609.07901
PDF
-
VoT: Vision-of-Thought for Unified Multimodal Representation Alignment
2026-09-07 17:53:13Jingxiang Sun, Chao Liao, Zhengxiong Luo, Chaorui Deng, Chen-lin Zhang, Junke Wang, Ceyuan Yang, Haoqi Fan, Weilin HuangAbstract
Current text-to-image systems typically employ a "text encoder plus diffusion decoder" paradigm, in which text semantics directly modulate continuous latent noise. Despite their success, these methods lack an explicit, interpretable intermediate representation that effectively bridges high-level linguistic semantics and low-level visual signals. In this paper, we propose Vision-of-Thought (VoT), a framework that introduces a discrete visual-thinking layer between vision-language models (VLMs) and diffusion transformers (DiTs). Instead of treating VLMs merely as text encoders, we use them as multimodal planners that generate discrete VoT tokens representing high-level visual plans, such as objects and layouts, before rendering pixels. We train a specialized VoT tokenizer in the VLM semantic space with a closed-loop objective that combines VLM alignment, feature reconstruction, and vector-quantization losses. These objectives make the tokens semantically readable by the VLM while preserving the visual information needed for generation. Experimental results demonstrate that VoT improves semantic alignment and provides a structured interface for interpretable and controllable generation.
Abstract (translated)
URL
https://arxiv.org/abs/2609.07815
PDF
-
Matryoshka Hash Representations for Model-Aware Compact Semantic Retrieval
Abstract
Retrieval-augmented generation (RAG) depends on dense retrieval: each document is stored as a learned vector, and a query is answered by finding its nearest neighbors in that vector space. Keeping one full-precision vector per document is the dominant index cost at corpus scale, so retrieval systems replace each vector with a short code of a few bytes---a step called quantization. Standard quantizers such as product quantization (PQ) pick the code that reconstructs the original vector most closely. A single code is even more useful if it serves several byte budgets at once: when its short prefixes are each directly searchable, a deployment can set its efficiency--quality operating point without re-encoding the corpus. But training all prefixes under one objective makes the early bits a compromise across budgets---short codes improve while the full-width code degrades. Quantization to low-bit representation, such as binary codes, further sharpens the conflict. We introduce Matryoshka Hash Representations (MHR), a two-stage procedure that separates full-width training from prefix organization. MHR first learns a longer binary code, then freezes the model and trains additional zero-initialized residual code adaptors for directly searchable prefixes. Documents are stored at one bit per coordinate, while queries keep continuous logits like PQ to attain sufficient expressivity. We implement the search process with FAISS FastScan. Trained on MS MARCO and zero-shot transferred to seven BEIR datasets, MHR reaches .5561 NDCG@10 and .6535 Recall@100 at 32 bytes, surpassing the best baseline of the same budget. The advantage is more pronounced in lower budgets. The same code also strengthens two common pipelines: shortlisting candidates for full-precision reranking, and pruning a low-storage graph index such as LEANN.
Abstract (translated)
URL
https://arxiv.org/abs/2609.07276
PDF
-
SIDE: Sensor Impersonation Detection at the Edge via Sequence Prediction
Abstract
Some low-cost Internet of Things (IoT) sensor deployments lack device-level source authentication, leaving them vulnerable to impersonation or injected sensor readings. We present a lightweight approach to sensor impersonation detection in a small proof-of-concept study. We formulate detection as a sequence-prediction problem. A model with three LSTM layers and two fully connected layers is trained only on univariate temperature readings from a genuine sensor, and a window of readings is flagged when its mean absolute prediction error exceeds the mean genuine error by more than six standard deviations. The model is converted to TensorFlow Lite and deployed on an Arduino Nano 33 BLE in three variants, non-quantized (554 KB), 16-bit weight quantized (298.5 KB), and 8-bit weight quantized (185 KB). On a controlled testbed with the impostor sensor placed in a hotter outdoor location, the three variants reached detection accuracies of 99.980%, 99.972%, and 98.206%, and each flagged the change point when a test sequence switched from genuine to impostor data. Quantization made the model smaller but slower in our measurements. The genuine and impostor distributions were well separated, so these results show detection of a controlled distribution shift and should not be read as evidence of general device authentication.
Abstract (translated)
URL
https://arxiv.org/abs/2609.06271
PDF
-
All for 1-Bit: Towards Genuine 1-Bit Post-Training Quantization for LLMs
2026-09-05 16:03:41Zhixiong Zhao, Zukang Xu, Guangyu Sun, Lifeng Liu, Dawei YangAbstract
Large language models (LLMs) have achieved remarkable progress, yet their massive storage and memory-bandwidth demands still hinder efficient deployment. Weight binarization is a promising solution, but existing binarization-based post-training quantization (PTQ) methods usually far exceed the nominal 1-bit storage target due to hidden overhead. To address this gap, we propose All for 1-Bit (AF1), a genuine 1-bit PTQ framework for LLMs. AF1 comprises two complementary components: (1) Null-space-Aware Binary Factorization (NABF) for improving binary reconstruction through Hessian-aware surrogate reparameterization, null-space-aware binary factorization, and scale-only global reconstruction; and (2) Hierarchical Shapley Allocation (HiSA) for assigning structural capacity using hierarchical Shapley sensitivity. Together, they preserve model accuracy under a strict 1.0-BPW budget in the PTQ setting. Experiments on LLaMA, Qwen, and Gemma families show that AF1 consistently outperforms existing binarization-based PTQ methods in perplexity and zero-shot accuracy. Compared with BF16, AF1 achieves an average 2.5 times inference speedup and over 90% memory reduction across evaluated models, providing a practical path toward deployable genuine 1-bit compression for LLMs. The code for reproducibility is available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2609.06161
PDF
-
Interface-Aware KV Cache Quantization for Dense On-Chip NVM in Long-Context LLM Decoding
2026-09-04 22:58:21Jiahao Zheng, Yifan Qin, Xiaobo Sharon Hu, Yiyu ShiAbstract
The key-value (KV) cache is the dominant memory bottleneck in long-context large language model (LLM) decoding: every step reads it entirely, so decoding is memory-bandwidth bound. Holding a quantized KV cache in dense on-chip non-volatile memory (NVM) removes the off-chip transfer. Existing KV quantization methods, however, were designed for GPU-style memory systems: KIVI attaches per-group metadata, adding about 25% to the stored KV cache; KVQuant keeps sparse full-precision outliers that a dense array cannot hold in place. This paper examines what these structures cost when the KV cache resides in NVM behind fixed-range converters, and designs a quantization scheme matched to that interface. The architecture stores the quantized KV cache in dense on-chip NVM, uses a small static analog crossbar only for the fixed rotation, and keeps attention in on-chip digital logic. A randomized rotation and per-vector normalization give every coordinate the same range, so one fixed codebook for keys and one for values, each shared across all tokens of the corresponding tensor type, serve the entire KV cache. The codebook thresholds are programmed once as the read converter's reference levels, enabling fixed-range digitization with no per-token converter reconfiguration. Dequantization is a sixteen-entry lookup and one norm multiply; the only per-vector metadata is one scalar, about 3%. Across models from 3B to 14B and contexts to 32k tokens, the four-bit KV cache maintains accuracy under storage and crossbar noise simulated at realistic device levels. KIVI and KVQuant remain more accurate in software; the advantage of our format lies at the memory interface: 3.1-3.6x lower KV read energy than both mapped to the same NVM, and 8x lower metadata overhead than KIVI. The contribution is a KV quantization co-designed with the NVM memory interface rather than a new accuracy record.
Abstract (translated)
URL
https://arxiv.org/abs/2609.05764
PDF
-
SeRV: Semantic-Aligned Residual Vector Quantization for American Sign Language Generation
2026-09-04 21:41:17Hongyu Wu, Xu Wu, Tianhao Wu, Jiawei Yu, Phuc Nguyen, Jian Liu, Yi WuAbstract
American Sign Language (ASL) generation remains challenging due to limited paired text-ASL motion data and the difficulty of learning motion representations both precise for reconstruction and predictable from linguistic input. Existing methods rely on motion tokenizers optimized for reconstruction, without explicit semantic supervision from paired text. As a result, the learned tokens remain limited in supporting semantically consistent and fine-grained ASL motion generation. To address this limitation, we propose SeRV (Semantic-Aligned Residual Vector Quantization), a semantic-aligned RVQ tokenizer for ASL generation. SeRV learns a semantically structured residual token space by combining sentence-level motion-text alignment with token-level text-conditioned supervision. Building on this tokenizer, a Hierarchical GPT predicts residual motion tokens in a coarse-to-fine manner, generating structurally coherent and semantically aligned 3D ASL motion. We further construct a large-scale reconstructed 3D ASL motion-text benchmark by recovering paired 3D motion from YouTube-ASL videos. Experiments across 375 hours of ASL video show that SeRV achieves state-of-the-art pose accuracy on both How2Sign and YouTube-ASL datasets, while producing semantically consistent 3D ASL motion directly from text.
Abstract (translated)
URL
https://arxiv.org/abs/2609.05742
PDF
-
Lightweight Vision Transformer Compression for On-Device Plant Disease Detection in Resource-Constrained Agricultural Field Conditions
2026-09-04 16:40:06Mahadev Sunil Kumar, Bhavika Gondi, Desaisetty Venkata Satya Sai Swapnith, Gangireddy Rahul Jogi, Sudheesh Manalil, Arnab Raha, Amitava Mukherjee, Parthasarathy Seethapathy, G. GopakumarAbstract
Chilli (Capsicum annuum) is one of India's most economically significant crops, yet its productivity is persistently threatened by diseases that are difficult to identify without expert intervention. While Vision Transformers (ViTs) have achieved high classification accuracy, their large computational footprint makes deployment on resource constrained devices challenging. Existing compression approaches typically address pruning, quantization, and knowledge distillation in isolation, leaving the potential benefits and interactions of their combined application insufficiently explored. We propose a unified Vision Transformer compression framework that combines Hessian-Balanced Adaptive Block Pruning (H-BAC), guided by second-order sensitivity estimation, with quantization and attention-based knowledge distillation. To systematically identify the most effective configuration within each compression family, each technique is first evaluated independently through controlled ablation studies, after which the best-performing components are integrated into a sequential deployment pipeline tailored to real-world agricultural constraints. On a chilli 3-class village-split dataset with a genuine cross-village, cross-device out-of-distribution test split, the resulting compressed models match or exceed the 95.13% FP32 baseline's accuracy, alongside 74-98% model size reduction, and the fully integrated compression pipeline achieves a 54.5x size reduction (327.42 MB to 6.01 MB) at 95.13 +/- 2.32% accuracy across four tested configurations. A direct comparison further reveals that, on this dataset, a directly-trained student of the same final size, without pruning or distillation, reaches comparable accuracy of 94.87%, at the same 6.01 MB INT8 size, indicating where H-BAC and knowledge distillation are, and are not yet shown to be, worth their computational cost.
Abstract (translated)
URL
https://arxiv.org/abs/2609.05334
PDF
-
APEX-RBD: Mixed-Precision Exploration Framework for Hardware-Efficient Robot Dynamics Accelerator Design
2026-09-04 14:02:44Xingyu Liu, Hanwei Fan, Chaofang Ma, Jiawei Liang, Guangyu Hu, Jiang Xu, Wei ZhangAbstract
Rigid Body Dynamics (RBD) forms the computational core of real-time robotic control, but its immense computational complexity creates a performance bottleneck that necessitates dedicated hardware accelerators. However, the substantial hardware resource and power costs of these accelerators make their deployment on resource-constrained edge platforms highly challenging. While quantization offers a promising path to optimize RBD hardware for edge computing, existing uniform-precision approaches remain inefficient by ignoring the diverse quantization sensitivities of different variables. Although mixed-precision offers a superior alternative, its exploration is intractable due to a vast search space and the prohibitive cost of closed-loop simulation for motion accuracy evaluation. To address these challenges, we introduce APEX-RBD, an automated framework that makes mixed-precision exploration computationally tractable while effectively identifying hardware-efficient configurations. Specifically, it performs physics-driven search space pruning via variable grouping and sensitivity analysis, and employs a data-efficient, prior-informed surrogate model to enable rapid trajectory error prediction. This formulation guides a hybrid optimizer to identify area- and power-efficient designs under user-defined accuracy and performance constraints. Experimental results demonstrate that APEX-RBD discovers designs achieving up to 1.9$\times$ area reduction and 1.8$\times$ power savings compared to uniform-precision baselines across diverse robotic platforms.
Abstract (translated)
URL
https://arxiv.org/abs/2609.05161
PDF
-
ProLombard: Structured Multi-Scale Modeling for Normal-to-Lombard Speech Conversion
2026-09-04 07:31:25Hongyang Chen, Xinmeng Xu, Youqiang Zheng, Xingyu Liu, Yuhong Yang, Zhongyuan Wang, Weiping Tu, Song LinAbstract
Normal-to-Lombard (N2L) speech conversion aims to improve speech intelligibility in noisy environments by transforming normal speech into Lombard-style speech while preserving linguistic content, speaker identity, and speech quality. Despite recent progress, existing methods typically model the Lombard effect at the utterance level or the frame level, overlooking its hierarchical nature and its entanglement with both speaker identity and phoneme-level content. This limitation leads to Lombard leakage in speaker representations and incomplete separation between Lombard characteristics and linguistic content. In this work, we propose ProLombard, a structured multi-scale N2L framework that explicitly models the Lombard effect across utterance-, phoneme-, and frame-level representations. To address Lombard-speaker entanglement, we introduce an aligned speaker encoder (ASE) that suppresses Lombard leakage by aligning Lombard-speech speaker embeddings with their normal-speech counterparts. To achieve more complete Lombard-content disentanglement, we develop a phoneme-aware disentanglement and injection mechanism that extends conventional frame-level modeling to the phoneme level. Furthermore, we design a vector quantization (VQ)-median module that provides robust phoneme-level representations through VQ-based segmentation and median-frame-based aggregation. Extensive experiments on Mandarin and English Lombard datasets demonstrate that the proposed approach consistently improves speech intelligibility, Lombard similarity, and perceptual quality over baselines while maintaining speaker identity. These results highlight the importance of structured multi-scale modeling for effective N2L speech conversion.
Abstract (translated)
URL
https://arxiv.org/abs/2609.04828
PDF
-
Scale-QLoRA: Code-Invariant Adapter Merging for Native 4-bit Microscaling LLMs
2026-09-03 22:29:37Tung-Ling Li, Jiale Huang, Lee-Chi Wang, Janaki Ram GoteiAbstract
Merging a LoRA adapter into its base model is standard deployment practice: it removes the runtime adapter's per-forward overhead and leaves a single standalone checkpoint any serving stack can load. On a native 4-bit microscaling checkpoint (NVFP4, MXFP4) that step stops being free. The merged weights must be written back through a quantizer, which re-derives the checkpoint's discrete E2M1 code plane (roughly 90% of the artifact's bytes), so the deployed artifact becomes coupled to one quantization convention, and every later code-touching event in its lifecycle can move it. Done naively the step is worse than fragile: it deletes the adaptation, by up to 39 pp, because against an already-on-grid base the reconstruction optimum is that base. Scale-QLoRA instead adapts only the native per-block scale field, trains those scales on the deployment grid, and freezes every E2M1 code. Within a fixed native format, scale grid, block layout and code plane, merging is then a bit-exact identity and the merged artifact is code-invariant. Across four models and four tasks, Scale-QLoRA and merge-aware QAT-LoRA are both accuracy-lossless, so we claim no accuracy ordering between them; they differ structurally, in that QAT-LoRA re-derives the code plane through a quantizer while Scale-QLoRA preserves it exactly. That difference is what the lifecycle prices: nearest-rounding implementations disagree by about a point on the measured task, and more extreme rule mismatches can drive the weight-space artifact to ~0%, which we report as a sensitivity bound rather than a deployment frequency. Preserving the code plane also drops the weight-space straight-through estimator from training (3.9x per step on the dense 8B model) and enables exact rollback, code-plane deduplication, and a ~125x faster scale-only task swap.
Abstract (translated)
URL
https://arxiv.org/abs/2609.04526
PDF
-
When Quantization Breaks Memory: Recurrent-State Write-Back in Low-Precision Temporal Inference
2026-09-03 21:23:35Ismail Erbas, Xavier Intes, Vikas PandeyAbstract
Quantization is widely used to reduce the computational and memory demands of neural-network inference. In recurrent networks, however, the quantized state is stored and returned at the next time step, so the rule used to store that state can alter subsequent computations. Here, we introduce recurrent-state write-back to denote this rule and isolate its effect in a compact GRU encoder--decoder for fluorescence lifetime imaging, a molecular imaging modality used in quantitative biological imaging. A central task is estimating two lifetime parameters, the short-lived component {\tau}1 and the long-lived component {\tau}2, from high-noise time-resolved fluorescence signals. Holding the trained model fixed, replacing continuous state propagation with deterministic 4-bit state storage increases estimation errors for {\tau}1 and {\tau}2 by approximately 70x and 300x, respectively. Failure occurs when repeated small updates remain below the write threshold, leaving the stored state nearly fixed while the network continues to propose change. Error feedback, residual memory, and direction memory carry information from these suppressed updates across time and recover accuracy without retraining. Precision sweeps show that increasing state precision can worsen a fixed recurrent solution, while matched training shows that compatibility with the state interface can be learned. To test whether this behavior extends beyond the GRU, we repeat the post-training intervention in an independently trained LSTM, where coarse write-back reproduces the failure, error feedback restores accuracy, and state-specific interventions reveal greater sensitivity of the cell state than the hidden state. Our results establish recurrent-state write-back as a key determinant of low-precision recurrent dynamics and identify the state-storage interface as a central design consideration for quantized recurrent inference.
Abstract (translated)
URL
https://arxiv.org/abs/2609.04490
PDF
-
Why Gated DeltaNet Survives 4-Bit Quantization: NVFP4 W4A4 for the Recurrent Half of a Hybrid 27B LLM
2026-09-03 17:04:26Sergii Kozyrev, Davyd MaiborodaAbstract
Hybrid LLMs pair softmax attention with linear-attention layers such as Gated DeltaNet (GDN), whose recurrent state summarizes the context in fixed size. Early community 4-bit quantizations of Qwen3.8-27B (48 GDN layers, 16 attention layers) left the GDN block in 8- or 16-bit precision -- especially its decay and write-strength gates -- on the intuition that errors in a recurrence accumulate over long contexts. We test that intuition by building Minima: NVFP4 W4A4 on all 496 linear layers, GDN included. Across perplexity at 4K/32K, MMLU-Pro, GSM8K, AIME'25, GPQA-Diamond, LiveCodeBench, and RULER retrieval to 64K, Minima matches BF16 within seed noise (5-task average -0.52) while being the smallest (17.5 GiB) and fastest-prefill (+14-19%) recipe we compare, and its 32K perplexity gap shrinks with position. A four-part mechanism study explains why: (i) NVFP4's 16-element block scaling localizes the residual stream's extreme outliers, equalizing activation error across layer roles; (ii) the supposedly fragile gate projections are the least sensitive -- softplus/exponential and sigmoid parameterizations compress ~11% GEMM error to ~2% output error; (iii) the delta-rule recurrence holds injected noise at a flat plateau over 32K tokens and forgets a state impulse within hundreds of steps, because each write overwrites the state along the current key direction; (iv) the per-token quantization cost washes out with context instead of compounding. We also repair a global-scale mismatch that arises when per-module-calibrated NVFP4 checkpoints are served by kernels that fuse those modules into one GEMM, and show calibrated FP8 KV-cache scales are performance-free. The result: a practical recipe -- quantize everything, ship KV scales -- and a mechanistic account of why the recurrent half of a hybrid LLM is the easy half to quantize. Checkpoint: this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2609.04098
PDF
