Interesting new work on fine-tuning by resetting a small number of weights.
The surprising effectiveness of weight deletion in LLM after training.
Jack Lu, Zhenbang Yang, Mike Lasby, Texas Futa, Yanni Yuan, Mengia Ren.
paper:.
review:.
project:.
Code: .
model:.
Key: Pre-trained language models can learn new thinking and tool skills without writing new weights. On a large scale, it is enough to reset less than 0.05% of the parameters, and the previous capabilities will not be affected.
What they did: The authors proposed a post-training method called deletion tuning, which fits a frozen model using binary mask training for MLP gate projections. The selected weights remain untouched, while the rest are reduced to zero.
What we got: On the largest models tested it was equal to full fine tuning. With reinforcement learning, the accuracy on old problems drops by no more than 1 point from the base model, although small models sometimes achieve only 90% of the accuracy of dense updating.
Why it matters: Inference systems can switch tasks with megabyte bitmasks on one common model without losing old skills. However, further sequential training has not been tested in action; Class 1 uncompressed bf16 adapters are often smaller in size, and the training itself requires much more memory.
Zero here: