“Large language model hallucinations are produced by the same underlying mechanism that generates correct outputs.”

Broadly, hallucinations in LLMs can be divided into two primary sources: (1) Prompting-induced hallucinations, where ill-structured, unspecified, or misleading prompts cause inefficient outputs, and (2) Model-internal hallucinations, which [are] caused by the model's architecture, pretraining data distribution, or inference behavior. Distinguishing between these two causes is essential for developing effective mitigation strategies.

Self-correction is generally thought of as a single process, but we decided to break it down into two components, mistake finding and output correction. Generally, we found these state-of-the-art models perform poorly, with the best model achieving 52.9% accuracy overall in mistake identification. Note that knowing the mistake location is different from knowing the right answer: CoT traces can contain logical mistakes even if the final answer is correct, or vice versa.

We argue that language models hallucinate because the training and evaluation procedures reward guessing over acknowledging uncertainty, and we analyze the statistical causes of hallucinations in the modern training pipeline.

As Large Language Models (LLMs) are nondeterministic, the same input can generate different outputs, some of which may be incorrect or hallucinated. If run again, the LLM may correct itself and produce the correct answer. Through a preliminary study, we identified five types of output inconsistencies.

LLMs generate both correct outputs and hallucinations through the same autoregressive next-token prediction mechanism; differences arise from probability distributions over the vocabulary, where low-probability tokens lead to factual errors despite high-probability coherent text.

This work argues that hallucinations in language models are not just occasional errors but an inevitable feature of these systems. We demonstrate that hallucinations stem from the fundamental mathematical and logical structure of LLMs. It is, therefore, impossible to eliminate them through architectural improvements, dataset enhancements, or fact-checking mechanisms.

Hallucination, in this light, should be seen not as a marginal bug but as a structural byproduct of intelligence under open-world conditions. This paper reframes “hallucination” as a manifestation of the generalization problem. When false memorization occurs, an LLM may produce an output inconsistent with facts present in its training data. When false generalization occurs, the model may fail to provide a correct answer absent from its training set but inferable from a human perspective.

Pre-trained LLMs are often good at producing a plausible-sounding answer, even when they cannot predict the right one. This problem is often known as hallucination. Fine-tuning techniques modify pre-trained models and change the types of outputs they are likely to produce, but do not remove harmful model capabilities.

While the model decides what is the most probable output, you can influence those probabilities by turning some model parameter knobs up and down. The temperature parameter controls the creative ability of your model: at lower temperature, the model is more conservative; higher temperature allows lesser probable words, resulting in more unpredictable text. Top-k and Top-p also control randomness of selecting the next token.

Before the 2025 research shift, the standard explanation was fairly simple. Hallucinations were seen as a side effect of three long-known factors: 1. Noisy or biased training data; 2. Model architecture and training quirks; 3. Decoding randomness.

In transformer LLMs, correct factual outputs and hallucinations both emerge from the same autoregressive next-token prediction process during inference; the distinction lies in whether the model's learned probabilities align with ground truth, influenced by training data quality and sampling methods.

LLM hallucination causes: Training data issues - Model limitations - Token size constraints - Nuanced language understanding difficulties. These errors do not result from a programmed process but arise from the limitations and complexities inherent in LLM training and data interpretation.

In the case of AI, these misinterpretations occur due to various factors, including overfitting, training data bias/inaccuracy and high model complexity.

The primary source of LLM hallucinations is the model's lack of training with domain-specific data. During inference, an LLM simply tries to account for knowledge gaps by inventing probable phrases.

LLMs are designed to be non-deterministic, meaning that the same input will not always generate the same output due to sampling from probability distributions.