The case for zero-error horizons in trustworthy LLMs
Summary
Zero-Error Horizon (ZEH) is proposed as the maximum error-free capability of an LLM. The paper reports GPT-5.2 failing simple tasks like parity checks and balanced parentheses, highlighting limitations for safety-critical use, and discusses correlations between ZEH and accuracy across models (e.g., Qwen2.5) along with potential speedups to reduce the cost of ZEH evaluation.