Research map / Large language models / Language, safety and interpretability
Old Habits Die Hard: How Conversational History Geometrically Traps LLMs
Adi Simhi, Fazl Barez, Martin Tutek, Yonatan Belinkov, Shay B Cohen
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| shuffle questions randomly (topic-inconsistent) instead of ordering by semantic similarity (ablation) | LLM multi-turn conversations · six datasets, open models | Tr(T) carryover: Tr(T) NaturalQA 1.13 -> 1.00; Sorry 1.57 -> 1.09 | Hurt |
| topic-inconsistent conversations vs consistent (ablation) | LLM multi-turn conversations · Sorry (open models) | P(refusal | refusal): 0.90 -> 0.75 | Hurt |
| topic-inconsistent conversations vs consistent (ablation) | LLM multi-turn conversations · theta_ref, LLaMA-3.1-8B Sorry | theta_ref (geometric separation): 66.52 -> 69.04 | No clear effect |
| use hidden states at 85% depth instead of 30%, 50% or 100% (ablation) | geometric perspective · probabilistic-geometric correlation | Spearman correlation: 0.60 | Improved |
| use refusal datasets instead of hallucination datasets | History-Echoes · open models | Tr(T): 1.13 -> 1.57 | Improved |
| evaluate closed model GPT-5 instead of open models | probabilistic perspective · S-pos | Tr(T): 1.33 -> 1.23 | No clear effect |
| condition on previous state k=2,3 steps back instead of only 1 (higher-order Markov) | probabilistic perspective · six datasets | Delta_k | Mixed |
| fix the question and vary preceding state (200-question subset, LLaMA-3.1) | probabilistic perspective · NaturalQA and Do-not-answer | P(same state | prior state): 74.36 vs 29.06 (NaturalQA); 88.39 vs 16.96 (Do-not-answer) | Improved |
Closest papers by meaning
- Bias Amplification in Language Model Evolution: An Iterated Learning Perspective (NeurIPS 2024 · similarity 0.85)
- ChroKnowledge: Unveiling Chronological Knowledge of Language Models in Multiple Domains (ICLR 2025 · similarity 0.85)
- The Geometry of Forgetting: Temporal Knowledge Drift as an Independent Axis in LLM Representations (NeurIPS 2026 · similarity 0.85)
- When LLMs Play the Telephone Game: Cultural Attractors as Conceptual Tools to Evaluate LLMs in Multi-turn Settings (ICLR 2025 · similarity 0.85)
- Mutable Transcripts: Mitigating Context Pollution through Editable Conversation State (NeurIPS 2026 · similarity 0.84)
- Correlation Dimension of Autoregressive Large Language Models (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
