Researchers discover that AI language model hallucinations stem from early commitment to incorrect trajectories governed by asymmetric neural dynamics.

Study uses 'same-prompt bifurcation' technique on Qwen2.5-1.5B model to observe how identical inputs spontaneously diverge into factual or hallucinated outputs

44.3% of test prompts (27 out of 61) showed hallucinations diverging from correct answers as early as the first generated token

Activation patching experiments reveal strong asymmetry: injecting hallucinated activations corrupts correct outputs 87.5% of the time, but reversing this only recovers correct outputs 33.3% of the time

Findings suggest hallucinations result from irreversible early trajectory commitments rather than random errors, requiring sustained multi-step intervention to correct