The paper investigates loss spikes that occur late in language model pretraining, attributing them to delayed, terminal divergences in output logits—termed time‑bomb spikes. It shows that these spikes are preceded by rapidly growing quantization error in the LM head, and that reducing numerical precision speeds up the spikes while increasing precision slows them. The authors propose replacing AdamW with SGD with momentum on the LM head to suppress logit growth, and find that various mitigation techniques reduce quantization error, indicating a common numerical cause for stability.
By Zaiqian Chen, Aditya Somasundaram, Sophie Li, Sean McLeish, John Kirchenbauer, Tom Goldstein, Micah Goldblum
arXiv:2606. 05516v1 Announce Type: new Abstract: Zeroth-order (ZO) optimization enables memory-efficient fine-tuning of large language models (LLMs) using only forward passes, but it remains unclear how useful adaptation is distributed across layers.
By Wanhao Yu, Ziyan Wang, Zheng Wang, Abeer Matar Almalky, Yihang Zuo, Shuteng Niu, Sen Lin, Adnan Siraj Rakin, Deliang Fan, Li Yang
arXiv:2608.23358v1 Announce Type: new
Abstract: The performance gap between low- and high-resource languages in LLMs is widely known, but it remains unclear which internal model factors drive these d...
By Francois Meyer, Jan Buys
The paper proposes Embedding Space Separation (ES2), a fine‑tuning technique that increases the distance between harmful and harmless embeddings in large language models. ES2 uses two regularization losses—a embedding anchor loss and a KL‑divergence loss—to preserve the model’s general capabilities while enhancing safety. Experiments show that ES2 effectively defends against various attack methods with only modest performance loss on some tasks.
By Xu Zhao, Xiting Wang, Weiran Shen
The study demonstrates that large language models can be compressed by clustering weight matrices into only 16–64 shared values without retraining, preserving strong accuracy. Fine‑tuning the cluster centroids recovers most of the remaining accuracy loss. Analysis reveals that maintaining the relative rank of weights and their overall variance is crucial for performance, while early layers and spatial geometry of cluster assignments also play significant roles.
By Borja Aizpurua, Sukhbinder Singh, Rom\'an Or\'us
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: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
The paper investigates how the optimizer Muon compares to AdamW during fine‑tuning of large language models. While Muon speeds up pretraining loss minimization, the study finds that when pretraining data is sufficiently large, AdamW actually outperforms Muon after fine‑tuning. The authors explain this reversal by introducing the margin‑to‑spread ratio, showing that Muon has a higher ratio at small data budgets but AdamW surpasses it as data grows, a trend confirmed both analytically and in simulations.
The paper investigates how the optimizer Muon, which accelerates pretraining loss minimization compared to AdamW, performs during fine‑tuning. It finds that when pretraining data is sufficiently large, AdamW actually outperforms Muon after fine‑tuning, and this advantage grows as model size increases. The authors explain this reversal by introducing the margin‑to‑spread ratio, showing that Muon has a higher ratio at small data budgets while AdamW dominates at larger budgets, a trend supported by theoretical analysis and simulations.
By Catherine Li, Jingchu Gai, Andrej Risteski, Jacob Mitchell Springer, Aditi Raghunathan
arXiv:2606. 06888v1 Announce Type: new Abstract: Classical scaling laws for language model pretraining balance model size against training dataset size under a fixed compute budget, assuming abundant data and a single pass over the corpus.
By Zhiwei Xu, Shihao Wu, Hanseul Cho, Wei Hu, Yixin Wang
The paper introduces MSign, an optimizer designed to prevent training instability in large language models by restoring the stable rank of weight matrices. It identifies two precursors to gradient explosions—rapid stable rank decline and increased Jacobian alignment—and proves that these jointly cause exponential gradient growth. Experiments on models ranging from 5 M to 3 B parameters show that MSign stops training failures while adding less than 7.0% computational overhead.
By Lianhai Ren, Yucheng Ding, Xiao Liu, Peng Cheng, Yeyun Gong
The paper introduces Latent Space Refusal Anchoring (LSR‑Anchoring), a training‑free technique that extracts a refusal direction from English prompts and applies it to the residual stream of instruction‑tuned models at inference time. The primary variant, Mean‑Activation Steering (MAS), works across several architectures (Llama‑3‑8B, Llama‑3.1‑70B, Mistral‑7B‑Instruct, Qwen2.5‑7B), restoring safety for low‑resource African languages with minimal performance loss, while a refined SAE‑Derived Steering (SDS) further reduces KL divergence without degrading legitimate prompt performance. The method shows positive transfer for Yoruba, Igbo, Igala, and Hausa, but fails for Arabic, suggesting a geometric mismatch rather than a data scarcity issue.
By Godwin Abuh Faruna