The paper introduces Trainable Vector Quantization (TVQ), a method that uses straight-through estimation to allow gradient descent optimization of vector-quantized large language models. TVQ reformulates VQ with trainable latent weights, enabling arbitrary objective minimization and is applied in a block-wise knowledge distillation framework to balance quality and training efficiency. Experiments show TVQ achieves state‑of‑the‑art low‑bit post‑training quantization, achieving near‑lossless 2‑bit quantization for LLaMA2‑70B with only a 1% accuracy drop on question‑answering benchmarks.
By Anastasiia Prutianova, Alina Bogdanova, Mikhail Shutov, Dmitrii Poimanov, Anatoly Gorkov, Ilia Zharikov
The paper proposes replacing dense output projection in large language models with an HNSW-based vector index to perform maximum inner product search over token embeddings. This approach reduces memory bandwidth usage by retrieving only a small set of high-scoring tokens and can be integrated into existing decoding pipelines via sparse logits scattering. Experiments on Gemma 3, Llama 3.2, and Qwen 3 show up to 82% speed‑up in batch‑size‑one decoding while maintaining generation quality.
By Martin Loretz, Sepp Hochreiter
The paper introduces Message-Passing Quantization (MPQ), a new approach to post‑training quantization (PTQ) that treats row‑wise PTQ as posterior inference over discrete codebooks. Unlike existing solvers that commit each weight block to a single codeword early, MPQ maintains soft posterior information across blocks using vector approximate message passing and only hard‑codes decisions at the final stage. Experiments on 2‑bit lattice and 3‑/4‑bit scalar quantization for dense language models from the Llama, Qwen, and Mistral families show that MPQ improves perplexity and zero‑shot accuracy compared to strong hard‑decision baselines and works well with preprocessing techniques such as activation‑aware scaling and incoherence processing.
By Yanshuo Chen, Chenxi Liu, Reza Shirkavand, Heng Huang
arXiv:2506. 20040v3 Announce Type: replace-cross Abstract: Interpreting language models remains challenging due to the existence of residual stream, which linearly mixes and duplicates features across adjacent layers, causing single-layer analyses to miss this cross-layer structure.
By Ankur Garg, Xuemin Yu, Hassan Sajjad, Samira Ebrahimi Kahou
The paper investigates how to quantize recurrent language models, which use linear recurrences instead of softmax attention. It distinguishes between extra‑state quantization (weights and activations outside the recurrent state) and intra‑state quantization (the recurrent state itself during decoding). The authors introduce LoReS, a quantization pipeline that applies value‑axis scale grouping, decay‑based mixed precision, and stochastic rounding, enabling robust 8‑bit state quantization for large models such as Qwen‑3.5, Nemotron‑H, and Kimi‑Linear, and achieving decode speedups when integrated into vLLM.
By Philipp Nazari, Neehal Tumma, Daniela Rus, T. Konstantin Rusch
arXiv:2601. 21461v3 Announce Type: replace-cross Abstract: Modern sparse language models typically achieve sparsity through Mixture-of-Experts (MoE) layers, which dynamically route tokens to dense MLP "experts.
By Albert Tseng, Christopher De Sa
arXiv:2606. 04620v1 Announce Type: cross Abstract: LLMs have become the state-of-the-art algorithms for solving NLP tasks.
By Pasindu Wickramasinghe, Achyuta Muthuvelan, Rachmad Vidya Wicaksana Putra, Minghao Shao, Muhammad Shafique
FastE is a training‑free, plug‑and‑play method that compresses token prefixes in large language model (LLM) embedding inference. It uses a shared fixed threshold on batch‑mean readout‑prefix alignment to decide when to compress and ranks prefix states by readout attention scores to keep the most important ones. Experiments on Qwen3‑Embedding models show that FastE can cut decoder‑backbone FLOPs by over 40% while preserving more than 99% of the original ranking quality across multiple benchmarks and tasks.
By Jinsong Shu, Jinyong Wen, Baokun Wang, Zhongle Xie, Lidan Shou, Weiqiang Wang, Gang Chen
The study benchmarks how hierarchical text embeddings and quantization affect performance across 38 tasks using four transformer models. It evaluates 168 configurations combining 1D/2D hierarchical representations with 32‑bit, 8‑bit, and binary quantization, reporting storage and evaluation time separately. Results show int8 quantization is nearly as effective as float32, binary quantization degrades performance, and dimensionality reduction paired with int8 can cut storage by 32× with minimal loss.
By Kai Ponel, Marco Polignano, Erasmo Purificato, Aylin Julia Temizkan, Marco Levantesi, Giovanni Semeraro, Ernesto William De Luca
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.
By Asim Kadav, Christian Flores, Chirag Arora, Varun Kotte, Hongbo Zheng, Lan Yan, Priya Shanmugasundaram, Tracy Holloway King
arXiv:2607. 16339v1 Announce Type: new Abstract: Diffusion-based Large Language Models(DLLMs) enable parallel generation via Semi-Autoregressive (SAR) decoding in text generation.
By Xingru Chen, Zelang Liang, Yongjia Ma, Jiqing Zhan, Shuling Yang, Lian Wen, Kun Zhan
arXiv:2607. 16213v1 Announce Type: new Abstract: Large Language Models (LLMs) generate text autoregressively, relying on a key-value (KV) cache whose memory footprint grows linearly with context length, creating a major bottleneck.
By Soumia Bouyahiaoui, Manel Kara laouar, Aicha Boutorh, Mohamed Hadj Ameur