Autoregressive LLM user-interaction optimization strategies based on the model's probabilistic properties

This paper explains that hallucinations can arise as a default behavior in LLMs because their autoregressive, probabilistic next-token prediction mechanism prioritizes plausibility over truth. It also points out that excessive rebuttals or constraints can contaminate the context and undermine response coherence. As solutions, it proposes an iterative “zeroing-in” approach that gradually narrows conditions, along with observation- and description-focused prompts to strengthen factual anchoring.