The Surprising Effectiveness of Deleting Weights in LLM Post-Training
0 repliesarxiv.org ↗
The Surprising Effectiveness of Deleting Weights in LLM Post-Training
arxiv.org↗
The Surprising Effectiveness of Deleting Weights in LLM Post-TrainingHow little of a pretrained large language model (LLM) has to change for it to acquire new capabilities? We find that zeroing fewer than 0.05% of its weights, with no other modification, matches full…
An arXiv preprint tests how little of a pretrained large language model (LLM) has to change to learn new capabilities.
- The method, Deletion Tuning, sets fewer than 0.05% of a model's weights (the numbers that store what it learned) to zero. Across three training methods it matched full fine-tuning, which retrains every weight.
- The three were reinforcement learning with verifiable rewards, supervised fine-tuning, and on-policy distillation.
- It learns a binary mask (a keep or delete switch) over the gate projection of transformer blocks, and the trained mask is hundreds to thousands of times smaller than a full checkpoint.
- Under reinforcement learning, accuracy on earlier tasks stayed near the base level, while full fine-tuning, LoRA (a cheaper fine-tuning method) and random sparse fine-tuning lost up to 5.4 points.
AI LLMs that see every letter: Most LLMs never see the individual letters of a word directly. A new method changes that by retrofitting existing models at a fraction of the usual training cost
EmbeddingGemma 2: a best-in-class open model for natively multimodal embeddings
mistralai/Mistral-Large-4-1T-A52B
Cerebras Deploys Trillion-Parameter AI Models
Muse Spark Open