RapidUn: Influence-Driven Parameter Reweighting for Efficient Large Language Model Unlearning

A new framework for aligning healthcare AI assistants has been introduced, focusing on balancing safety and helpfulness through iterative preference alignment. This approach utilizes Kahneman-Tversky Optimization and Direct Preference Optimization to refine large language models (LLMs) against specific safety signals, resulting in significant improvements in harmful query detection metrics.

A recent study presents the Universal Weight Subspace Hypothesis, revealing that deep neural networks trained on various tasks converge to similar low-dimensional parametric subspaces. This research analyzed over 1,100 models, including Mistral-7B, Vision Transformers, and LLaMA-8B, demonstrating that these networks exploit shared spectral subspaces regardless of initialization or task.

A large-scale mixture-of-experts (MoE) pretraining study has been conducted using pure AMD hardware, specifically the MI300X GPUs and Pollara networking. This research provides practical guidance on system and model design, including microbenchmarks for core collectives and kernel sizing to optimize training throughput and inference latency.

CoGraM (Contextual Granular Merging) is a newly introduced optimization method designed to enhance the merging of neural networks without retraining, addressing issues of accuracy and stability that are prevalent in existing methods like Fisher merging. This multi-stage, context-sensitive approach utilizes rollback mechanisms to prevent harmful updates, thereby improving the robustness of the merged network.