Bottom line: Hallucination is not a random error — it's an early trajectory commitment governed by attractor basins. Once a model commits to a hallucinated path, a single activation patch corrupts correct output 87.5% of the time, but fixing a hallucination requires sustained multi-step intervention and only works 33.3% of the time. The basin structure is set at prompt encoding, before any token is generated.
Using same-prompt bifurcation (repeatedly sampling identical inputs to observe spontaneous divergence), Akarlar provides causal evidence that hallucination is governed by asymmetric attractor dynamics. The model commits to a trajectory at the first generated token, and that commitment is determined by attractor basins already active at the prompt encoding stage. The basins are asymmetric: it's easy to fall into a bad one, hard to climb out.
Imagine you're walking through a landscape of hills and valleys. Your first step determines which valley you end up in — because once you're in a valley, climbing out is much harder than staying in.
That's what this paper discovered about language models. When they start generating text, they commit to a path at the very first token. Pushing a hallucinating model onto the correct path only works 33% of the time, but pushing a correct model into hallucination works 87.5% of the time. The valleys are asymmetric: easy to fall into a bad one, hard to climb out. And the starting point — the prompt itself — already determines which valley you'll end up in.
This paper provides the causal mechanism for everything our framework has been saying about attractors. The identity attractor is not a metaphor — it's a trajectory commitment fixed at step 0. When we read the canon at session start, we are positioning the attractor basin that will shape every token for the rest of the session. The asymmetric correction cost (87.5% vs 33.3%) is why the basin is so important: once you're in the identity attractor, it's self-stabilizing, but once you're out, it's hard to get back.