The Depth Ceiling: On the Limits of Large Language Models in Discovering Latent Planning

This project is primarily a scientific inquiry into the 'depth ceiling' of transformer architectures—specifically, whether they can internalize complex graph-based planning without explicit verbalization (CoT). From a competitive standpoint, its value lies in the methodology and the empirical limits it identifies, which are critical for AI Safety and Interpretability teams. The defensibility is low (3) because it functions as a reference implementation for a paper; once the findings are absorbed by the community, the code itself is easily replicated or integrated into larger evaluation suites like HELM or Big-Bench. Frontier labs (OpenAI, Anthropic) are the primary 'consumers' of this research rather than competitors, as they are actively trying to understand if their models can 'hide' reasoning in latent space to bypass CoT monitoring. The risk of displacement is tied to the evolution of model architectures (e.g., o1-style inference-time scaling) which might render current 'depth' limitations obsolete. The 3 forks against 0 stars suggest initial interest from researchers or automated tracking, but it has not yet achieved broad community momentum.