'Zero-Byte' Output Phenomenon Discovered in Major LLMs

As a result of conducting 31,430 experiments on 11 large language models (LLMs) including GPT, Claude, Gemini, and Kimi, a 'Zero-Byte' phenomenon in which 11,658 successful executions had no visible UTF-8 output bytes whatsoever was discovered.

This phenomenon is analyzed as an issue related to the model's semantic structure, rather than a simple refusal, safety block, rate limit, or transmission error. In particular, among 4,290 strictly matched semantic pairs, 2,505 'Voids' occurred under certain conditions.

The research suggests a structural flaw occurring in the model's output control and semantic processing, and the related data and verification code have been made public.