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. Code available at https://github.com/technion-cs-nlp/OldHabitsDieHard.

• Publication year

  2026

• Venue

  Unknown venue

• Publication date

  2026-02-08

• Fields of study

  Computer Science

• Identifiers
  arXiv 2603.03308
• External record

  Open on Semantic Scholar

• Source metadata

  Semantic Scholar

• No claims are published for this paper.

• No concepts are published for this paper.

• Sycophancy Hides Linearly in the Attention Heads

  2026cited by this paper
• Belief Dynamics Reveal the Dual Nature of In-Context Learning and Activation Steering

  2025cited by this paper
• Qwen3 Technical Report

  2025cited by this paper
• The LLM Already Knows: Estimating LLM-Perceived Question Difficulty via Hidden Representations

  2025cited by this paper
• gpt-oss-120b&gpt-oss-20b Model Card

  2025cited by this paper
• Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models

  2025cited by this paper
• HACK: Hallucinations Along Certainty and Knowledge Axes

  2025cited by this paper
• Safety Alignment Should Be Made More Than Just a Few Tokens Deep

  2024cited by this paper
• Do-Not-Answer: Evaluating Safeguards in LLMs

  2024cited by this paper
• Refusal in Language Models Is Mediated by a Single Direction

  2024influential reference
• SORRY-Bench: Systematically Evaluating Large Language Model Safety Refusal Behaviors

  2024cited by this paper
• The Remarkable Robustness of LLMs: Stages of Inference?

  2024cited by this paper
• The Llama 3 Herd of Models

  2024cited by this paper
• Large Language Models as Markov Chains

  2024cited by this paper
• Markov Chain of Thought for Efficient Mathematical Reasoning

  2024cited by this paper
• Layer by Layer: Uncovering Where Multi-Task Learning Happens in Instruction-Tuned Large Language Models

  2024cited by this paper
• Distinguishing Ignorance from Error in LLM Hallucinations

  2024cited by this paper
• Frontier Models are Capable of In-context Scheming

  2024cited by this paper
• LoRA Fine-tuning Efficiently Undoes Safety Training in Llama 2-Chat 70B

  2023cited by this paper
• How Language Model Hallucinations Can Snowball

  2023cited by this paper
• Inference-Time Intervention: Eliciting Truthful Answers from a Language Model

  2023cited by this paper
• Universal and Transferable Adversarial Attacks on Aligned Language Models

  2023cited by this paper
• Large Language Models Cannot Self-Correct Reasoning Yet

  2023cited by this paper
• The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasets

  2023cited by this paper
• Towards Understanding Sycophancy in Language Models

  2023cited by this paper
• Survey of Hallucination in Natural Language Generation

  2022cited by this paper
• Extracting automata from recurrent neural networks using queries and counterexamples (extended version)

  2022cited by this paper
• What Makes Good In-Context Examples for GPT-3?

  2021cited by this paper
• Similarity Analysis of Contextual Word Representation Models

  2020cited by this paper
• Similarity of Neural Network Representations Revisited

  2019cited by this paper
• Natural Questions: A Benchmark for Question Answering Research

  2019influential reference
• TriviaQA: A Large Scale Distantly Supervised Challenge Dataset for Reading Comprehension

  2017influential reference
• A Mathematical Theory of Communication

  2006cited by this paper
• Interpolated estimation of Markov source parameters from sparse data

  1980cited by this paper
• Verbal context and the recall of meaningful material.

  1950cited by this paper

• No citing papers are available for this paper.