Hierarchical Grading in Large Language Models

Researchers introduced Graded Large Language Models (GLLMs), an algebraic framework applying grading to transformer representation spaces. This construction utilizes geometric invariant theory to maintain expressive power and inference cost. Separately, a study on 570 MMLU questions analyzed how prompt tone affects LLM accuracy and output-token consumption. Results indicate that output-token length varied by up to 44.3% across seven tones, significantly exceeding accuracy variation, particularly within ChatGPT 4o and 5-nano models during the reasoning process.