Explaining Hallucinations from a TCS perspective: Interview with Santosh Vempala

Abstract

After sketching the landscape of how TCS can or does contribute to ML in the last Bulletin, we propose here to dive into one of the most recent attempts to shed a TCS light on some ML phenomenon – namely, using statistics to explain hallucination in language generation. We conducted for this an interview with Santosh Vempala, who pioneered this direction with Adam Tauman Kalai. Santosh tells us here the story behind the maths: the motivation for the paper, how they proceed with modeling and what are the implications of the result.