LLMs Exhibit 'Zero-Visible-Byte' Output in Cross-Vendor Study
TL;DR. A new preprint details a cross-vendor evaluation revealing significant instances of LLMs producing successful zero-visible-byte outputs, termed "Voids." - The study analyzed 31,430 single-turn trials across 11 models from OpenAI, Anthropic, Google, and Moonshot. - Researchers observed Voids in 37.09% of trials, establishing a reproducible class of unexpected model behavior. - This finding indicates a need for deeper understanding of LLM termination and alignment mechanisms.
- 37.09% of LLM executions across major vendors resulted in 'Voids' (zero visible output bytes).
- Study included 31,430 trials from OpenAI, Anthropic, Google, and Moonshot models.
- The phenomenon is distinct from refusals, safety blocks, or operational failures.
- Findings highlight a reproducible class of successful zero-visible-byte executions needing further investigation.