Research map / Large language models / Language, safety and interpretability
Why and How LLMs Hallucinate: Connecting the Dots with Subsequence Associations
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a subsequence association framework that explains LLM hallucinations as dominant non-contiguous input subsequences, with theory on transformers as subsequence embedders and a beam-search tracing algorithm (SAT) that finds reproducible hallucination triggers.
Problem
Hallucinations in LLMs arise from many entangled causes and are hard to trace; attribution methods give token-wise scores confined to one context and fail to identify reproducible causal triggers.
Approach
Hallucinations occur when dominant hallucinatory subsequence associations outweigh faithful ones; the authors define a PMI-style subsequence association and trace triggers by measuring hallucination reproducibility across randomized input contexts (masking/refill with embedding-similar tokens) using beam search.
Key results
- SAT achieves average S_rep of 46.5 on Llama-70B versus 19.8 for attention, across 7 test domains.
- On Olmo-7B, SAT reaches average S_rep 42.5 versus 27.5 for the best baseline (grad-shap).
- Reproducibility rates correlate with training-corpus conditional probabilities on Dolma with Pearson rho=0.72.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SAT instead of the strongest reported attribution baseline | Llama-70B · HALoGEN average | S_rep: 19.8 -> 46.5 | Helped |
| Use SAT instead of the strongest reported attribution baseline | Olmo-7B · HALoGEN average | S_rep: 27.5 -> 42.5 | Helped |
Surprising findings
- Hallucinated outputs can be reproduced at >20% rates even when 75% of the input tokens are randomly replaced, suggesting specific subsequences drive hallucinations.
Limitations
- Theory relies on simplifying assumptions such as linear attention and independence of subsequence associations.
- Tracing is evaluated only on six programmatically verifiable HALoGEN domains.
- Approximate input distribution via masking/refill may not match real-world input distributions.
Experimental setup
Datasets: HALoGEN Dolma-1.7
Benchmarks: HALoGEN
Metrics: reproducibility score S_rep Pearson correlation with Dolma co-occurrence
Baselines: attention LIME gradient SHAP ReAGent
Builds on: pointwise mutual information HALoGEN Dolma
Closest papers by meaning
- LLM-Check: Investigating Detection of Hallucinations in Large Language Models (NeurIPS 2024 · similarity 0.94)
- Attributive Reasoning for Hallucination Diagnosis of Large Language Models (AAAI 2025 · similarity 0.92)
- In-Context Sharpness as Alerts: An Inner Representation Perspective for Hallucination Mitigation (ICML 2024 · similarity 0.92)
- Scalable Token-Level Hallucination Detection in Large Language Models (NeurIPS 2026 · similarity 0.92)
- HaDeMiF: Hallucination Detection and Mitigation in Large Language Models (ICLR 2025 · similarity 0.92)
- Auditing Meta-Cognitive Hallucinations in Reasoning Large Language Models (NeurIPS 2025 · similarity 0.91)
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.