1 paper · 1 filter
Qisong He, Jinwei Hu, Xinmiao Huang +3
In a long conversation, an LLM can produce a plausible continuation that rests on premises the conversation has already abandoned. No runtime check ties its output to what the conv…