Abstracts: January 25, 2024

Get ready to explore the latest breakthroughs in technology research with Abstracts. This episode features Senior Researchers Jordan Ash and Dipendra Misra in conversation with host Gretchen Huizinga on their accepted paper, "The Truth is in There: Improving Reasoning in Language Models with Layer-Selective Rank Reduction," at the 2024 International Conference on Learning Representations (ICLR).

The team's new technique, Layer-Selective Rank reduction (LASER), involves targeted reduction of parameters in transformer-based models. By removing certain parameters, the method retains and even improves model performance without additional training. This research demonstrates the value of parameter-reduction methods in natural language processing and how they can be optimized for specific use cases.

To learn more about the paper, visit their publication from December 2023 or check out the LASER code on GitHub. You can also register for the Microsoft Research Forum, featuring panel discussions and lightning talks on science and technology research in the age of general AI.

Stay up-to-date on the latest advances in technology by tuning in to Abstracts.