A New research paper From OpenAI asks why large language models such as GPT-5 and chatbots such as Chatgpt still hallucinate, and whether something can be done to reduce those hallucinations.
In A blog post in which the newspaper is summarizedOpenAI defines hallucinations as “plausible but false statements generated by language models”, and it acknowledges that despite improvements “hallucinations” remain a fundamental challenge for all major language models ” – one that will never be completely eliminated.
To illustrate the point, researchers say that when they asked ‘a commonly used chatbot’ about the title of Adam Tauman Kalai’s Ph.D. Thesis, they got three different answers, all wrong. (Kalai is one of the authors of the article.) They then asked for his birthday and received three different dates. Again, they were all wrong.
How can a chatbot be so wrong – and so self -confident sounds in his mistakes? The researchers suggest that partial hallucinations arise, due to a pre -road process that focuses on getting models to correctly predict the next word, without true or false labels associated with the training statements: “The model only sees positive examples of flowing language and must approach the overall distribution.”
“Spelling and hooks follow consistent patterns, so mistakes there disappear with scale,” they write. “But random low -frequency facts, such as the birthday of a pet, can not only be predicted from patterns and therefore lead to hallucinations.”
The proposed solution of the article, however, focuses less on the initial pretrainment process and more on how large language models are evaluated. It argues that the current evaluation models themselves do not cause hallucinations, but they ‘determine the wrong stimuli’.
The researchers compare these evaluations with the type of multiple choice tests random gamble makes sense, because “you may be lucky and right”, while the answer leaves empty “guarantees a zero”.
Techcrunch event
San Francisco
|
27-29 October 2025
“In the same way, when models are only assessed for accuracy, the percentage of the questions they get just right are encouraged to guess instead of saying:” I don’t know, “they say.
The proposed solution is therefore comparable to tests (such as the SAT) that “include negative [scoring] For wrong answers or partial credit to leave questions to discourage blind councils. “Likewise, OpenAI says that model evaluations” must punish self -assured errors than you punish and partially give credit for appropriate expressions of uncertainty. “
And the researchers claim that it is not enough to “introduce a few new uncertainty tests on the side.” Instead, “the commonly used, accuracy -based evals must be updated, so that their score discourages guidance.”
“If the most important scoreboards continue to reward happy guesses, models continue to guess,” the researchers say.
#bad #stimuli #fault #Hallucinations #Techcrunch