Compare_Hotblaz: Comparison-Driven Multidimensional Reasoning for Large Language Models

Large Language Models (LLMs) have achieved remarkable performance across reasoning, scientific analysis, software engineering, and knowledge-intensive tasks. Despite these advances, most existing inference frameworks remain fundamentally constrained by next-token prediction and single-path reasoning paradigms. Such approaches naturally favor responses that remain semantically close to the original prompt, potentially limiting the discovery of hidden facts, latent associations, and alternative reasoning trajectories.

In this paper we introduce Compare_Hotblaz, a comparison-driven inference framework designed to augment language models through multidimensional answer exploration. Instead of generating a single reasoning chain, Compare_Hotblaz produces multiple parallel comparison samples, incorporates memory-guided associative mechanisms, and performs statistical answer selection in embedding space.

We further propose the Mean Closeness Hypothesis, which suggests that correct answers frequently emerge near the statistical center of answer clusters rather than at the shortest semantic distance to the original prompt. Experimental evaluation on GPQA Diamond demonstrates that Compare_Hotblaz improves the performance of DeepSeek-V3.2-thinking from 85.86% to 92.6%.

Our findings suggest that comparison-based reasoning may provide a practical inference-time alternative to purely sequential reasoning paradigms and may serve as a foundation for future multidimensional reasoning architectures.