事後学習で新たな知識の獲得は難しい [Gekhman+ 24] Test Loss – 計算資源量︓学習に⽤いるデータ量 × パラメータ数 Compute Dataset Size Parameters PF-days, non-embedding tokens non-embedding [Kaplan+ 20] より 7 Figure 1 Language modeling performance improves smoothly as we increase the model size, datasetset
[Azizi+ 26] Looped Language Models: A Survey and Meta-Analysis of Recurrent-Depth Transformers [Takase and Kiyono] Lessons on Parameter Sharing across Layers in Transformers [Gekhman+ 24] Does Fine-Tuning LLMs on New Knowledge Encourage Hallucinations? [Kaplan+ 20] Scaling Laws for Neural Language Models [Le Scao+ 22] What Language Model to Train if You Have One Million GPU Hours? [Zeng+ 23] GLM-130B: An Open Bilingual Pre-trained Model [Nishida+ 24] Initialization of Large Language Models via Reparameterization to Mitigate Loss Spikes [Takase+ 25] Spike No More: Stabilizing the Pre-training of Large Language Models [Zhang+ 22] OPT: Open Pre-trained Transformer Language Models 29
22] PaLM: Scaling Language Modeling with Pathways [Tian+ 26] Simulating Training Spikes of Deep Neural Networks by Layer-wise Random Feature Models [Li+ 22] The Stability-Efficiency Dilemma: Investigating Sequence Length Warmup for Training GPT Models [Glorot+ 10] Understanding the Difficulty of Training Deep Feedforward Neural Networks [He+ 15] Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification [Yang+ 21] Tensor Programs V: Tuning Large Neural Networks via Zero-Shot Hyperparameter Transfer [Yano+ 26] Pre-training LLM without Learning Rate Decay Enhances Supervised FineTuning [Shibata+ 26] Suppressing Final Layer Hidden State Jumps in Transformer Pretraining 30