'Old Habits' paper: chat history geometrically traps LLMs

A new paper argues that once a large language model has said something inside a conversation, it becomes measurably harder for that model to say something different later on. The authors call the effect a 'geometric trap,' and they think it helps explain why an early hallucination inside a session tends to stick around instead of getting corrected.

The work, posted to arXiv as 'Old Habits Die Hard: How Conversational History Geometrically Traps LLMs,' introduces a framework called History-Echoes that looks at conversation drift from two angles at once. Probabilistically, the authors model a running chat as a Markov chain and quantify how consistent the model's state stays across turns. Geometrically, they measure the consistency of the hidden representations the model produces at consecutive exchanges. Across three model families and six datasets, the two views line up: gaps in the latent space appear to confine the model's trajectory toward what it has already committed to.

Why this matters outside the interpretability crowd is that a lot of real product surface area now runs on long conversations. Agents that plan across many turns, coding copilots that carry file context, chat products that remember an entire session, all of them quietly assume a wrong turn early on can be talked out of later. If the geometry actually says otherwise, the honest workaround is not always 'try again in the same thread,' it is start a fresh one.

The honest caveat is that this is one paper with one framework, tested on a specific set of models and datasets, and the authors do not claim to have solved the underlying problem, only measured it. What the write-up does not give you is whether the same trapping effect shows up in the closed frontier systems most people actually use, or whether editing turns out of the history would break the trap or just move it. Still, if the geometric account holds up, it points at a cleaner design principle for anyone shipping long-running assistants: treat conversation history as load-bearing state, not free context.