Large models 'know' they are about to hallucinate before they generate even a single token.
April 16, 2026
Original Paper
Before the First Token: Scale-Dependent Emergence of Hallucination Signals in Autoregressive Language Models
arXiv · 2604.13068
The Takeaway
In models larger than 1B parameters, researchers found a detectable signal in the internal latent state that predicts a hallucination before the first word is ever spoken. This contradicts the belief that hallucinations are just a cumulative result of bad autoregressive choices. It suggests the model has an internal sense of 'un-groundedness' from the start. This discovery unlocks a 'hallucination circuit breaker' that can stop a model from answering if the latent signal looks suspicious. By monitoring this signal, practitioners can build significantly more reliable RAG systems and chatbots that refuse to answer when they're 'about to lie.' It turns hallucination detection into a pre-emptive security layer rather than a post-generation cleanup task.
From the abstract
When do large language models decide to hallucinate? Despite serious consequences in healthcare, law, and finance, few formal answers exist. Recent work shows autoregressive models maintain internal representations distinguishing factual from fictional outputs, but when these representations peak as a function of model scale remains poorly understood.We study the temporal dynamics of hallucination-indicative internal representations across 7 autoregressive transformers (117M--7B parameters) usin