The paper introduces Nexus, an optimizer that encourages gradient similarity during pretraining to bring task‑specific minima closer together. Experiments on models from 130M to 3B parameters show that Nexus achieves the same pretraining loss as standard optimizers while significantly improving downstream performance, including a 15% accuracy boost on GSM8k and a 0.012 reduction in out‑of‑distribution loss for the 3B model.
arXiv:2510. 01163v2 Announce Type: replace Abstract: The factors driving the performance of in-context learning (ICL) in large language models (LLMs) remain poorly understood despite ICL's surprising effectiveness, enabling models to adapt to new tasks from only a handful of examples.
arXiv:2602. 14696v2 Announce Type: replace Abstract: Instruction fine-tuning of large language models (LLMs) often involves selecting a subset of instruction training data from a large candidate pool, using a small query set from the target task.
The paper evaluates ten deep learning optimizers across models ranging from 0.1B to 1.2B parameters and varying data-to-model ratios, highlighting that fair comparisons require rigorous hyperparameter tuning and end-of-training evaluations. It finds that claimed speedups over AdamW are lower than advertised, dropping to only 1.1× for the largest models, and that intermediate checkpoint comparisons can mislead due to learning‑rate decay. The study shows that matrix‑based optimizers like Muon and Soap achieve the highest speedups, but these benefits diminish as model size increases.
The paper introduces LARE, a method that adjusts domain weights during large language model pretraining by reweighting losses based on running loss statistics, rather than altering data exposure. Experiments on a 1.5B model trained with 240B tokens show that LARE improves performance across five benchmarks, achieving significant gains on GSM8K (+17.06 pp) and MATH (+9.00 pp). The approach demonstrates that balancing loss scales can enhance joint capabilities without increasing domain exposure.
arXiv:2605.12484v3 Announce Type: replace Abstract: Large language models (LLMs) are trained for downstream tasks by updating their parameters (e.g., via RL). However, updating model parameters force...
The paper investigates federated learning where each client’s data consists of unknown mixtures of distinct tasks, a scenario termed compound heterogeneity. It shows that when tasks share a common feature geometry, the optimal model for a mixed client is a convex combination of task‑specific models, motivating input‑dependent routing to specialized experts. The authors propose FedSEE, a method that recovers task experts via a convex program and achieves better performance than baselines, reducing negative transfer by 2.9 points overall and 3.7 points for the worst‑served quartile.
arXiv:2606. 03217v1 Announce Type: cross Abstract: Chain-of-thought (CoT) reasoning has become a widely used mechanism for eliciting multi-step reasoning in large language models by generating intermediate reasoning steps at inference time.
The paper introduces FedRouter, a personalized federated learning framework that clusters tasks rather than clients to build specialized language models. It employs local and global clustering of adapters to associate them with specific tasks and includes an evaluation router that directs test samples to the most suitable adapter. Experiments on a multitask dataset show FedRouter outperforms existing methods, achieving up to 6.1% relative improvement under task interference and 136% under generalization scenarios.
The paper investigates the often-overlooked scale vectors in large language models, showing that despite their tiny size they are crucial for pre‑training performance. The authors provide theoretical insights that scale vectors mainly aid optimization rather than expressivity, and they analyze how weight decay affects different normalization layers. Building on these findings, they propose lightweight improvements—branch‑specific heterogeneity, better placement, and magnitude‑direction reparameterization—that consistently reduce loss across a range of model sizes and training settings.
arXiv:2505. 20161v2 Announce Type: replace-cross Abstract: Effective generalization in language models depends critically on the diversity of their training data.
The report introduces A.X K2, a 688‑parameter Mixture‑of‑Experts language model designed for agentic applications. Trained on 8.5 trillion tokens, it surpasses its predecessor A.X K1 by over 30 percentage points on several benchmarks, thanks to a higher‑quality data mix and improved token efficiency. Key innovations include Sparse Gated Attention for efficient long‑context handling, Gated Norm for training stability, and a Think‑Fusion recipe that allows switching between thinking and non‑thinking modes within the same model.
arXiv:2607. 01104v1 Announce Type: cross Abstract: In Large Language Model (LLM) training, data mixing plays a pivotal role in determining model performance.