Get Started
Home
Topics
Search
Library
Research questionHow can sampling-only diagnosis identify whether an individual LLM hallucination comes from retrieval, training, or autoregressive commitment?A fluent false claim may arise from how a model retrieves contextual information, learns from its training objective, or commits to an earlier continuation. Without access to internal activations or parameters, developers need to determine which mechanism is implicated using observable model outputs.
AI
Alignment & Safety
Evaluation & Benchmarks
LLM Pretraining & Post-training
Mechanistic Interpretability
Natural Language Processing
Latest papersRecent research connected to this question, newest first.From Architecture to Output: Structural Origins of Hallucination in Large Language Models and the Amplifying Role of DataApplies to decoder-only LLMs with sampling access. The evidence examines self-attention's associative retrieval, the maximum-likelihood pretraining objective, and autoregressive commitment under exposure bias, including a partial commitment test across three model families; dataset defects are treated as amplifiers that need not independently originate every failure.research paper · Sep 4, 2026
Related questions
How can we reliably detect when an LLM response is unsupported by its reference documents?How can black-box LLM APIs detect hallucinations without trusted context while limiting false alarms?How can LLM-based automated program repair detect patches and artifacts unsupported by repair evidence?How can we detect object hallucinations in vision-language models when visual grounding shifts across layers?