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.
By Michael S. Zhang, Rishi A. Ruia, Arnav Kewalram, Saathvik Dharmapuram, Utkarsh Sharma, Kevin Zhu
The paper addresses training instabilities in large language model pretraining, specifically output logit divergence that occurs near the end of training. By analyzing the geometry of output embeddings, the authors identify anisotropic embeddings as the root cause and propose Output Embedding Centering (OEC) as a mitigation strategy. OEC can be applied deterministically as μ‑centering or as a regularization loss μ‑loss, and experiments show both variants outperform the existing z‑loss method while matching logit soft‑capping in stability, even without weight tying. Additionally, μ‑loss is less sensitive to hyperparameter tuning than z‑loss.
By Felix Stollenwerk, Anna Lokrantz, Niclas Hertzberg
Post‑training quantization compresses large language models by storing weights at reduced precision, introducing errors into hidden states that could accumulate with depth. However, pretrained models accumulate far less hidden‑state error than randomly initialized ones, largely preserving downstream performance. The study identifies two key mechanisms: (1) each layer’s new error tends to oppose inherited error, partially canceling it, and (2) the LM‑head geometry preserves high‑rank token scores, mitigating output changes.
By Yuxiang Chen, Michael Beyer, Jun Zhu, Jianfei Chen
InfoQuant is a train‑free method that reshapes activation distributions to be more compatible with low‑bit quantizers. It uses Peak Suppression Orthogonal Transformation (PSOT) and adaptive outlier‑token selection to reduce numerical range while maintaining dispersion, thereby lowering quantization error. Across several large language model families, InfoQuant outperforms existing post‑training quantization and end‑to‑end training baselines, preserving 97% of floating‑point accuracy on average and narrowing the LLaMA‑2 13B performance gap by 42%.
By Ke Li, Dong An, Xiaoling Zang, Can Ye, Liang Xie, Qibo Qiu, Chen Shen, Xiaofei He, Wenxiao Wang
arXiv:2502. 11034v3 Announce Type: replace Abstract: Loss spikes remain a persistent obstacle in large-scale language model pretraining.
By Guoxia Wang, Shuai Li, Congliang Chen, Jinle Zeng, Jiabin Yang, Dianhai Yu, Yanjun Ma, Li Shen
arXiv:2505. 01043v2 Announce Type: replace Abstract: Large language models (LLMs) have achieved impressive performance across various domains.
By Zhiwei Hao, Jianyuan Guo, Li Shen, Yong Luo, Han Hu, Guoxia Wang, Dianhai Yu, Yonggang Wen, Dacheng Tao
arXiv:2605. 09825v4 Announce Type: replace-cross Abstract: Why does full-pipeline FP4 training of large language models often diverge, even when forward activations and activation gradients remain stable?
By Musa Cim, Sarthak Arora, Poovaiah Palangappa, Miro Hodak, Ravi Dwivedula, Meena Arunachalam, Mahmut Taylan Kandemir
arXiv:2608. 08188v1 Announce Type: new Abstract: Post-training quantization reduces the deployment cost of large language models, yet how severely a quantized model degrades is not determined by bit-width alone.
By Chenxi Zhou, Pengfei Cao, Jinyu Ye, Bohan Yu, Haida Yu, Jiang Li, Jun Zhao, Kang Liu
The paper demonstrates that layer dropout, also known as stochastic depth, can be effectively used in state‑of‑the‑art large language model (LLM) training. By optimizing the layer distribution, schedule, and optimizer settings, the authors show that layer dropout can reduce training loss while saving up to 25 % of training FLOPs. Additionally, layer dropout enables post‑training optimizations such as early exit and self‑speculative decoding, achieving up to 1.5× inference speedup with negligible accuracy loss across models ranging from 271 M to 8.2 B parameters and datasets up to 160 B tokens.
By Mostafa Elhoushi, Alex Pretko, Nolan Dey, Bin Claire Zhang, Gavia Gray, Gurpreet Gosal, Abdulrahman Mahmoud, Shane Bergsma, Joel Hestness
The paper introduces Calibrated Clipping, a dynamic method to align FP8 quantization bounds with high‑precision BF16 distributions, thereby mitigating training instability in full‑pipeline FP8 reinforcement learning for large language models. It identifies that compounded FP8 noise distorts importance ratios, causing entropy surges and garbled outputs. Experiments across GRPO and DAPO algorithms on 8B‑32B models show the technique restores performance to BF16 levels.
By Fanchao Chen, Ziheng Jiang, Ziyun Wei, Zheng Zhong, Du Li, Chi Zhang, Haibin Lin, Shivaram Venkataraman
arXiv:2606. 02288v1 Announce Type: new Abstract: Massive activation spikes in Large Language Models (LLMs) severely degrade quantization by stretching dynamic ranges.
By Yung-Chin Chen, Chung Peng Lee, Ze-Wei Liou, Naveen Verma
Deep neural networks can exhibit periodic loss spikes known as the "Slingshot Mechanism" during long‑term training. This paper shows that such spikes arise from finite‑precision floating‑point arithmetic: when the correct‑class logit becomes much larger than the others, the gradient for that class is rounded to zero while others remain nonzero, breaking the zero‑sum constraint and creating a positive feedback loop that inflates feature and classifier norms. The authors term this Numerical Feature Inflation (NFI) and demonstrate that it explains the rapid norm growth, reappearance of gradients, and loss spikes, while also showing that partial absorption can still cause hidden norm growth without visible spikes.
By Hanqing Liu, Jianjun Cao, Yuanze Li, Zijian Zhou