The accuracy of GPT-4.5 on troubleshooting questions, particularly in complex domains like multimodal troubleshooting and virology, is not explicitly detailed in the available information. However, GPT-4.5 has shown improvements in certain areas compared to its predecessors. For instance, it performs better in recognizing patterns and drawing connections, which could potentially enhance its troubleshooting capabilities[5].
In terms of specific evaluations, GPT-4.5 generally performs on par with GPT-4o and o1 in refusing unsafe content and in some safety-related evaluations[1][4]. However, when it comes to accuracy in answering questions, GPT-4.5 has shown significant improvements in certain datasets. For example, in the PersonQA dataset, which evaluates hallucinations and accuracy, GPT-4.5 achieves an accuracy of 0.78 and a hallucination rate of 0.19, outperforming GPT-4o and o1[1][4].
In fairness and bias evaluations, GPT-4.5 performs similarly to GPT-4o but is outperformed by o1 in providing unbiased answers to unambiguous questions[1][4]. While GPT-4.5 has demonstrated improvements in handling conflicting messages and adhering to safety instructions, its overall accuracy in troubleshooting questions specifically is not directly compared to GPT-4o and o1 in the available data[1][4].
In general, GPT-4.5's ability to recognize patterns and draw connections may enhance its performance in troubleshooting tasks, but specific comparisons in this domain are limited. Industry observers have noted improvements in certain tasks, such as data extraction, where GPT-4.5 outperforms GPT-4o by extracting fields more accurately[7]. However, detailed comparisons specifically for troubleshooting questions are not provided in the available information.
Citations:
[1] https://cdn.openai.com/gpt-4-5-system-card.pdf
[2] https://www.reddit.com/r/ChatGPTPro/comments/1ggnm90/gpt4o_and_others_provide_more_incorrect_than/
[3] https://www.reddit.com/r/OpenAI/comments/1izp6tt/gpt_45_released_heres_benchmarks/
[4] https://assets.ctfassets.net/kftzwdyauwt9/7EaDv6OaWHhXLAehUYu7Db/64e9f7916d3581ba4b5d0f0a6c5098d1/GPT-4-5_System_Card_2272025.pdf
[5] https://www.theverge.com/news/620021/openai-gpt-4-5-orion-ai-model-release
[6] https://neoteric.eu/blog/gpt-4o-vs-gpt-4-vs-gpt-3-5-comparison-in-real-world-scenarios/
[7] https://venturebeat.com/ai/industry-observers-say-gpt-4-5-is-an-odd-model-question-its-price/
[8] https://community.openai.com/t/gpt-4-vs-gpt-4o-which-is-the-better/746991