From the source
Lead story
Ask your AI
Top stories
Models & availability
Latest
Pruning LLMs Like a Physicist: Block Removal as an Ising Optimization Problem
Reformulates block removal as Ising optimization, achieving 23-point MMLU gain at 50% compression.
Hugging Face and Multiverse Computing published a research paper proposing a physics-inspired approach to pruning large language models by reformulating block removal as a constrained binary optimization problem mapped to an Ising glass.
The method achieves significant compression gains, reaching 23 percentage points improvement on MMLU over competing block-removal methods at 50% compression of Llama-3.3-70B-Instruct.
From the source
at 50% compression of Llama-3.3-70B-Instruct, we gain almost 23 percentage points on MMLU over the best competing block-removal method.
huggingface.co