IN LANGUAGE MODELS

Summary from the TLDR AI newsletter

Language models often hallucinate/confabulate information. This is likely because they are trained to imitate expert sequences which in turn causes them to treat their own (potentially wrong) generations as ground truth. Therefore, as sequence lengths grow, models are more likely to make stuff up. However, if you treat it as a Reinforcement Learning problem and find the next most probable action - while including a backspace action - you can dramatically reduce these hallucination issues.