GPT-5.4 is a language model in OpenAI's GPT-5 line. Recent coverage uses it mainly as a benchmark: Ai2 tested it among six LLM judges rating research idea novelty, a NeurIPS 2026 paper found it flipped 44.7% of correct answers when a source was labeled verified, Sitefire had it rewrite 2,250 company blog posts to train an AI-text classifier, and Microsoft and the University of Illinois compared their StudentSim tutors against it.
Ai2 study finds LLM novelty judges unreliable for research ideasAi2 released "Old Ideas, Novel Problems," testing six LLM judges—GPT-5.1, GPT-5.2, GPT-5.4, Claude Sonnet 4.5, Claude Opus 4.5 and Claude Opus 4.6—on rating research idea novelty. Small changes to evaluation setup shifted verdicts sharply; the team released a benchmark, data and code.
Study Reveals Authority Bias in LLMsA NeurIPS 2026 paper finds LLMs that push back on wrong users still accept the same wrong answer from a 'verified source,' an effect called Authority Bias. One source note flipped 45-88% of correct answers in 7 of 8 models; GPT-5.4 flipped 44.7% and Grok-4.20 87.5%.
Sitefire trains classifier to detect AI content by structureYC W26 startup Sitefire ported the StoryScope method to commercial web pages: it collected 2,250 pre-ChatGPT B2B blog posts from 268 company sites, had five models including GPT-5.4 rewrite each, and trained a classifier on 214 structural questions, reaching 98% accuracy separating AI from human posts.
Simulated students help AI tutors learn fasterMicrosoft and the University of Illinois built StudentSim, which replicates individual students from limited data to give AI tutors fast, low-cost feedback. In tests with 60 students across chess, English and math it outperformed GPT-5.4, and a chess tutor trained with it earned the highest expert ratings among three versions.
The A.I. Daily