Convergence Point Theory An Empirical Study on the Higher-Order Determinants of LLM Token Prediction Entropy and the Effects of RLHF-Induced Artificial Convergence Points (2026.04.30 Paper Revision)

AbstractThis study begins with the question: what properties of user utterances determine the internaluncertainty of AI language models? Prior research has sought to explain this phenomenon throughindividual variables such as utterance abstraction, lexical difficulty, and reasoning demand, but nooverarching principle unifying these accounts has been proposed. This study introduces theconcept of the Convergence Point — defined as the structural density of human knowledgeconsensus embedded within a user utterance — and argues that this density functions as the higherorder condition governing an AI model's token prediction entropy. Experiments were conducted in parallel using llamacpp-based token prediction entropymeasurement, core/structure-separated entropy analysis, TransformerLens-based Logit Lensanalysis, and RLHF bias measurement. A total of five systematic experimental trials were carriedout across five models: Mistral-7B-Instruct, Meta-Llama-3. 1-8B-Instruct, DeepSeek-R1-DistillQwen-7B, Gemma-2-9B-it, and Qwen3-8B. The results showed that entropy decreased by up to 68% when high-density Convergence Pointswere provided, and that entropy varied systematically across a four-tier spectrum — FullConsensus, medical consensus, Partial Consensus, and Non-consensus — with this patternreproduced consistently across all five models. Furthermore, Artificial Convergence Pointsimplanted in Non-consensus domains via RLHF exhibited stronger entropy-suppression effectsthan high-density Convergence Points provided directly by humans, suggesting that RLHF canfunction as a source of spurious convergence in domains where genuine consensus does not exist. This study proposes that Convergence Point Theory can integrate under a single overarchingprinciple phenomena previously explained in isolation — including hallucination, prompt effects, and response instability on philosophical questions — and offers structural implications for thedesign of AI role boundaries and alignment strategies.