arXiv:2609.26708v1 Announce Type: new Abstract: Quantization-aware distillation (QAD) restores much of the short-form question-answering performance lost to sub-3-bit quantization, yet leaves mathema...
OnlineQAT introduces a two-stage framework for ultra-low-bit large language models, beginning with block-wise quantization-aware training to create a usable low-bit initialization and then applying on-policy distillation (OPD) using student-generated responses. The method employs a frozen full-precision teacher to provide a sampled reverse-KL training signal at each visited prefix, addressing the mismatch between offline recovery data and the deployed model’s self-generated prefixes. Experiments on Qwen3-1.7B show that OnlineQAT achieves the best average performance among compared quantized methods, outperforming ReasoningQAT by 2.90 and 0.44 points at W3A16 and W2A16 respectively, highlighting the benefit of student-visited states for recovery.
arXiv:2606. 02011v1 Announce Type: new Abstract: Large Reasoning Models (LRMs) rely on long reasoning traces, making inference expensive.
arXiv:2606. 00206v1 Announce Type: new Abstract: Post-training quantization (PTQ) is widely used to deploy large language models efficiently, but its effect on reasoning models is not well understood.
Large Reasoning Models (LRMs) rely on long reasoning traces, making inference expensive. While low-bit quantization reduces per-token decoding cost, we show that aggressive 2-bit inference can fail to deliver end-to-end speedup because instability in the generation process inflates total token count.
arXiv:2512. 18934v2 Announce Type: replace-cross Abstract: Catastrophic forgetting poses a fundamental challenge in continual learning, particularly when models are quantized for deployment efficiency.
The paper introduces RATIO, a framework for improving quantized reasoning models by identifying overthinking tokens and applying token-specific penalties. It uses Quantization-aware Reasoning Behavior Analysis to detect problematic tokens and Token-Specific Penalty Determination to assign penalties without extra training. Experiments show RATIO outperforms existing methods, boosting accuracy by up to 9.8 points and shortening chain-of-thought length by up to 51.3%.
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...
arXiv:2606. 15682v1 Announce Type: new Abstract: Large Reasoning Models (LRMs) achieve strong problem-solving through long chain-of-thought, but their deployment is constrained by the high cost of full-precision inference and growing KV cache footprints.
Post-training quantization (PTQ) has become a widely adopted technique for reducing the memory footprint and inference cost of large language models (LLMs). However, recent studies reveal that when ap...
The paper presents practical training recipes for looped language models, showing that a compute‑efficient pipeline can reduce the training budget from 7.7 T tokens to 310 B tokens while maintaining strong reasoning performance. In controlled experiments, a 1.4 B LoopLM outperforms a parameter‑matched dense model on 12 benchmarks, achieving significant gains on GSM8K, MATH, and DROP, and approaches a 3.9 B dense model at matched inference compute. The authors also provide a minimal recipe to convert pretrained dense models into looped ones, demonstrating improvements on Qwen3‑1.7B‑Base across multiple data regimes.
The paper introduces DAMP, a decay‑aware mixed‑precision quantization scheme for recurrent‑state representations in GDN and KDA language models. By identifying high‑risk channels through quantization‑error energy and decay persistence, DAMP stores these channels at higher precision while compressing the rest to INT8, achieving a 9.9‑bit average precision. Experiments on Qwen3.6‑35B and Kimi‑Linear‑48B show a 69.1% reduction in recurrent‑state storage, up to 2.01× faster state‑update kernels, and up to 10.9% lower full‑model TPOT while preserving accuracy close to the FP32 baseline.