Improving Reasoning in Language Models with LASER: Layer-Selective Rank Reduction
Episode 1
Duration:
Talk
Speakers:
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)
Session Code: E1 D