Скачать или смотреть Evaluating and Surgically Improving Generalization in Language Models

Aaron Mueller explores challenges in understanding and improving language model (LM) generalization. The talk highlights two projects: evaluating how LMs process language structure, showing pre-training on code enhances robustness, and investigating chain-of-thought reasoning, which often improves performance only in-distribution. Mueller also introduces mechanistic interpretability using feature circuits to reveal how LMs handle tasks like subject-verb agreement and mitigate distractions, concluding with an application of this method to improve classifier generalization.

To find out more, see the Nokia Bell Labs Responsible AI hub: https://www.bell-labs.com/research-in...

#transparency , #ai , #artificialintelligence , #responsibleai , #ResponsibleArtificialIntelligence, #ethicalai , #TrustworthyAI, #RegulatoryActivity, #researchanddevelopment , #trust , #fairness , #safety , #reliability , #security , #privacy , #transparency , #sustainability , #accountability , #innovation , #technology , #BellLabs, #NokiaBellLabs, #nokia