One wrong word at the start of a sentence traps an AI in a mathematical hole it can never leave.
April 20, 2026
Original Paper
Hallucination as Trajectory Commitment: Causal Evidence for Asymmetric Attractor Dynamics in Transformer Generation
arXiv · 2604.15400
The Takeaway
Hallucinations are the result of asymmetric attractor dynamics where a model commits to a false path from the very first step. Once the transformer enters this stable attractor basin, the internal logic becomes corrupted and makes correction nearly impossible. People often assume AI forgets facts in the middle of a sentence, but the error is actually baked into the trajectory from the beginning. This causal evidence shows that the model is physically pulled toward a specific conclusion by its own initial output. Fixing AI errors requires disrupting these mathematical basins rather than just providing better training data.
From the abstract
We present causal evidence that hallucination in autoregressive language models is an early trajectory commitment governed by asymmetric attractor dynamics. Using same-prompt bifurcation, in which we repeatedly sample identical inputs to observe spontaneous divergence, we isolate trajectory dynamics from prompt-level confounds. On Qwen2.5-1.5B across 61 prompts spanning six categories, 27 prompts (44.3%) bifurcate with factual and hallucinated trajectories diverging at the first generated token