Loading...

Improving Reasoning in Language Models with LASER: Layer-Selective Rank Reduction

Episode 1

Duration:

Talk

Talk

Speakers:
Talk

Talk

Details

Dipendra Misra presented a surprising discovery that by merely replacing selected weight matrices in an LLM with their suitable low-rank approximation, you can significantly improve the performance of the LLM, at times by 20-30 percentage points.

Speakers (1)
 Dipendra   Misra
Senior Researcher Microsoft Research AI Frontiers

Session Code: E1 D