{"id":1190337,"date":"2026-05-11T17:17:41","date_gmt":"2026-05-11T16:17:41","guid":{"rendered":"https:\/\/staging-artefact.digitalseeder.com\/?post_type=blog&p=1190337"},"modified":"2026-05-11T17:32:03","modified_gmt":"2026-05-11T16:32:03","slug":"detecting-hallucinations-in-llms-one-token-at-a-time","status":"publish","type":"blog","link":"https:\/\/staging-artefact.digitalseeder.com\/fr\/blog\/detecting-hallucinations-in-llms-one-token-at-a-time\/","title":{"rendered":"D\u00e9tection des hallucinations dans les LLM, un token \u00e0 la fois"},"content":{"rendered":"

How entropy-based scoring can tell you when your model is making things up \u2014 and where \u2014 wrapped up in artefactual, our Python package.<\/p>\n

Note: This article is a follow-up to the article from our friends at Ardian<\/a>, detailing how crucial accountable AI is for financial institutions. Make sure to check it out!<\/p>\n

The hallucination problem<\/h2>\n

Les grands mod\u00e8les linguistiques sont \u00e9tonnamment capables. Ils r\u00e9sument, traduisent, raisonnent et codent (mieux que moi). Mais contrairement \u00e0 moi, ils sont \u00e9galement devenus c\u00e9l\u00e8bres pour avoir invent\u00e9 des faits avec une confiance d\u00e9concertante.<\/p>\n

In the Natural Language Processing (NLP) literature, a hallucination is any model-generated content that is factually incorrect, nonsensical, or unfaithful to a provided source, while appearing perfectly plausible. The consequences range from benign (a wrong trivia answer) to severe (a fabricated legal citation, an incorrect drug dosage). As organizations integrate LLMs into production systems, the question shifts from \u201ccan this model generate useful text?\u201d<\/em> \u00e0 \u201ccan we trust what it just said?\u201d<\/em><\/p>\n

Consider a concrete example. You work at a financial institution, and you ask your local LLM:<\/p>\n

\u201cWhat was Emerson Electric\u2019s net revenue in 2023?\u201d<\/p><\/blockquote>\n

\"\"<\/p>\n

The model replies: \u201cEmerson Electric reported a net revenue of approximately $15.2 billion for fiscal year 2023.\u201d<\/em> Sounds reasonable. But is it right? You don\u2019t have the annual report open. You don\u2019t have a ground truth to compare against. You just have the model\u2019s output \u2014 and doubt.<\/p>\n

This is the setting we work in. No oracle. No reference answer at inference time. Just an LLM response and the metadata it produces while generating it. The goal: quantify how likely this output is to be hallucinated, from a single generation pass.<\/p>\n

Detecting hallucinations: it\u2019s harder than it sounds<\/h2>\n

The brute-force approach<\/h3>\n

One natural idea is to ask the model the same question several times and check whether the answers agree. If five out of six runs say \u201c$15.2 billion\u201d and one says \u201c$18.7 billion\u201d, the consensus gives you some confidence. This is the principle behind methods like SelfCheckGPT, which measure consistency across multiple sampled outputs \u2014 a \u201cMonte Carlo-style\u201d approach to hallucination detection.<\/p>\n

It works. But it comes with two significant drawbacks:<\/p>\n

    \n
  1. Cost.<\/strong> Each additional generation multiplies your inference budget. For SelfCheckGPT with 10 samples, you pay roughly 10x the compute, plus the cost of a semantic similarity model on top. At scale, this is prohibitive.<\/li>\n
  2. Granularity.<\/strong> Multi-shot methods operate at the sequence level. They tell you \u201cthis answer seems unreliable\u201d, but not which part of the answer is problematic. A response might be 90% accurate with a single hallucinated figure buried in the middle. You\u2019d like to know where.<\/li>\n<\/ol>\n

    These limitations motivated us to look for a different signal \u2014 one that is cheap, single-shot, and works at the token level (the individual pieces of words the LLM manipulates internally).<\/p>\n

    The signal is already there<\/h3>\n

    When an LLM generates text, it doesn\u2019t just output tokens. At each step, it computes a probability distribution over its entire vocabulary: \u201cgiven the prompt and everything I\u2019ve generated so far, how likely is each possible next token?\u201d<\/em>\u00a0The winning token gets sampled. The rest are discarded. But those probabilities (and more specifically, how spread out they are) carry information about the model\u2019s internal confidence.<\/p>\n

    If the model is very sure, most of the probability mass concentrates on a single token. If the model hesitates, the probability spreads across many candidates. This spread is exactly what entropy measures.<\/p>\n

    Entropy: a quick detour<\/h3>\n

    Entropy is an information-theoretic quantity that measures the uncertainty of a probability distribution. The intuition is straightforward. Imagine three boxes. One contains a cookie. You have to guess which one.<\/p>\n

    \"\"<\/p>\n