Old Habits Die Hard: How Conversational History Geometrically Traps LLMs

How does the conversational past of large language models (LLMs) influence their future performance? Recent work suggests that LLMs are affected by their conversational history in unexpected ways. For instance, hallucinations in prior interactions may influence subsequent model responses. In this work, we introduce History Echoes, a framework that investigates how conversational history biases subsequent generations. The framework explores this bias from two perspectives: probabilistically, we model conversations as Markov chains to quantify state consistency; geometrically, we measure the consistency of consecutive hidden representations. Across three model families and six datasets spanning diverse phenomena, our analysis reveals a strong correlation between the two perspectives. By bridging these perspectives, we demonstrate that behavioral persistence manifests as a geometric trap, where gaps in the latent space confine the model's trajectory.

Large language models (LLMs) sometimes get stuck in a cyclic feedback loop during a conversation. If a model starts hallucinating or agreeing too much with a user, it often continues doing so in future responses. In this work, we investigated why an LLM's conversational past so strongly influences its future behavior. We developed a framework to track how these behaviors persist across a conversation. By analyzing the LLM's internal workings, we discovered that past interactions create a "geometric trap." Previous responses confine the LLM's internal state to a specific region, making it difficult to change direction and break the habit. Our research explains the mechanics behind why LLMs are getting stuck. Understanding how these models get trapped is a crucial step toward building more reliable AI systems that can correct themselves and break out of unhelpful conversational loops.