arXiv:2606. 05682v2 Announce Type: replace Abstract: Demand for low-precision inference, including NVFP4-based approaches, has grown as large language models are increasingly deployed in latency and cost constrained production environments.
arXiv:2606. 05682v1 Announce Type: cross Abstract: Demand for low-precision inference, including NVFP4-based approaches, has grown as large language models are increasingly deployed in latency and cost constrained production environments.
arXiv:2601. 22709v5 Announce Type: replace-cross Abstract: Vision-Language Models (VLMs) achieve strong multimodal performance but are costly to deploy, and post-training quantization often causes significant accuracy loss.
The paper argues that relying solely on zero‑shot task accuracy is insufficient for evaluating quantized large language models (LLMs) because accuracy ignores changes in the full predictive distribution. It proposes a distribution‑sensitive framework that measures fidelity loss by computing statistical distances—such as Jensen‑Shannon Divergence and Total Variation Distance—between the full‑vocabulary output distributions of a full‑precision BF16 reference and its quantized counterparts. Experiments across five foundation architectures and four reasoning benchmarks show that these divergence metrics increase with stronger quantization, revealing distributional drift that top‑1 accuracy fails to capture, and suggest that mixed‑precision Q4_K schemes can offer lower divergence than uniform Q4_0 at comparable memory usage.
arXiv:2606. 04238v1 Announce Type: cross Abstract: Aggressive weight quantization to 2-bit precision offers substantial throughput and memory gains for large language model (LLM) inference, but typically incurs severe accuracy degradation.
arXiv:2607. 24953v1 Announce Type: cross Abstract: Reducing training precision is a key lever for improving the e ciency of large language model (LLM) training, but pushing beyond FP8 to 4-bit oating point (FP4) remains challenging due to instability during optimization.
The paper proposes Self‑Distillation Fine‑Tuning (SDFT) as a method to recover performance in Large Language Models that has been degraded by catastrophic forgetting, quantization, or pruning. It shows that SDFT restores model capabilities by aligning the high‑dimensional manifold of the student model’s hidden layers with that of a teacher model, as measured by Centered Kernel Alignment (CKA). The authors provide both empirical evidence of strong correlation between manifold alignment and performance recovery and a theoretical explanation linking generative capability to the structure of these manifolds.
arXiv:2609.39223v2 Announce Type: new Abstract: Large language model (LLM) inference is increasingly moving toward lower precision to realize the throughput of hardware accelerators, but aggressive p...
arXiv:2608.20953v1 Announce Type: cross Abstract: Serving large language models cheaply increasingly means shipping models that are both structurally compressed to a fraction of their parameters and...
The paper investigates low‑bit quantization for Multimodal Large Language Models (MLLMs), showing that MXFP8 retains near‑lossless performance while 4‑bit formats like MXFP4 and HiF4 cause significant degradation. It identifies activation quantization as the main source of this loss and introduces Residual Fallback Quantization (RFQ), a lightweight framework that adds a quantized residual pathway to improve activation fidelity without architectural changes. Experiments on Wan2.2 and Qwen3‑VL demonstrate that RFQ recovers much of the performance gap to BF16 baselines across generation and reasoning tasks.
The study evaluates 4‑bit quantization and low‑rank adapter fine‑tuning (QLoRA) on several large protein language models, finding that many model‑task pairs retain over 90% of full fine‑tuning performance while achieving up to 90% GPU memory savings. QLoRA preserves early‑layer representations and induces task‑specific changes in later layers, closely resembling full fine‑tuning with smaller representational shifts. For generative models, 4‑bit quantization largely maintains structural and sequence‑level properties, though token‑level analysis reveals model‑dependent changes in autoregressive output distributions.
The paper introduces a post‑training softmax reparameterization technique that selects a functionally equivalent output head before quantization. By subtracting a scalar multiple of the vocabulary‑row mean from each output row and tuning this coefficient via validation KL, the method preserves the full‑precision softmax distribution while enabling efficient W4 quantization. Experiments on seven heads show significant error reductions and latency improvements, with the approach remaining complementary to other quantization strategies and transferable across datasets.
The paper introduces Engram Adapter, a conditional memory adapter that selectively activates on in-domain inputs while suppressing out-of-domain (OOD) retrievals. It leverages multi‑channel matching over local n‑gram patterns with occupancy tracking and a learned scalar gate to inject residuals only when appropriate. Experiments on Qwen3-4B/8B with AG‑News and MedMCQA show improved in‑domain accuracy while preserving nearly all OOD performance, outperforming always‑on baselines on LegalBench.