MathGPT breaks records, beats ChatGPT, Microsoft AI models.

MathGPT breaks records, beats ChatGPT, Microsoft AI models

MathGPT has achieved a groundbreaking feat by surpassing Microsoft's "ToRA 13B," the previous record-holder, in benchmarks evaluating mathematical proficiency. Developed by Mathpresso, the creators of QANDA, the leading AI-driven learning platform in Asia, MathGPT has outperformed both OpenAI's GPT-4 and Microsoft's models in assessments like 'MATH' (comprising 12,500 challenging math questions) and "GSM8K" (featuring 8,500 elementary school arithmetic problems), as reported by Interesting Engineering.

The remarkable achievement comes after a collaborative effort between QANDA, Upstage, and KT, with MathGPT emerging as the new #1 in mathematical aptitude benchmarks. The development initiative began in November of the previous year, leveraging a strategic cooperation between QANDA and Upstage. Upstage, armed with learning data from 10 million daily searches on QANDA, integrated factors such as learning level, context, and interaction to refine MathGPT.

In a significant show of confidence, KT invested $8 million in Mathpresso in September of the same year to support the growth of their large language model (LLM). The collaboration aimed to harness the power of QANDA's extensive learning data to enhance MathGPT's capabilities.

One distinctive feature of MathGPT is its ability to perform logical inference, a refinement introduced by Upstage to the natural language-based model. This enhancement addresses potential hallucinations, preventing the generation of responses that may inadvertently convey inaccurate information.

In contrast, models like ChatGPT, which relies on large-scale textual data for training, may exhibit the phenomenon of hallucinations. This occurs when the model produces responses that, while plausible, might convey inaccurate information, highlighting the specialized nature of MathGPT in the realm of mathematical aptitude.