Jonas Schweisthal

Structured actions like text messages can be improved from observational data by learning a policy over discrete action codes, bypassing the need for expensive retraining of generative models.

LLMs with web access only marginally outperform their offline counterparts, challenging assumptions about the necessity of real-time data for accurate predictions.

LLM development is flying blind by ignoring causal inference, leaving models vulnerable to confounding and distribution shifts throughout pretraining, alignment, and evaluation.