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

TL;DR

This paper investigates the limits of latent planning in large language models, revealing a maximum of five to seven steps learned during training but generalizing up to eight, highlighting a gap between discovering and executing strategies.

Contribution

It uncovers a fundamental limit in latent planning depth in LLMs, showing that models can discover but not always execute multi-step strategies effectively.

Findings

Tiny transformers discover up to 3 latent steps.

GPT-4o and Qwen3-32B reach 5 steps.

GPT-5.4 generalizes to 8 steps at test-time.

Abstract

The viability of chain-of-thought (CoT) monitoring hinges on models being unable to reason effectively in their latent representations. Yet little is known about the limits of such latent reasoning in LLMs. We test these limits by studying whether models can discover multi-step planning strategies without supervision on intermediate steps and execute them latently, within a single forward pass. Using graph path-finding tasks that precisely control the number of required latent planning steps, we uncover a striking limitation unresolved by massive scaling: tiny transformers trained from scratch discover strategies requiring up to three latent steps, fine-tuned GPT-4o and Qwen3-32B reach five, and GPT-5.4 attains seven under few-shot prompting. Although the maximum latent planning depth models can learn during training is five, the discovered strategy generalizes up to eight latent steps at test-time. This reveals a dissociation between the ability to discover a latent strategy under final-answer supervision alone and the ability to execute it once discovered. If similar limits hold more broadly, strategies requiring multiple coordinated latent planning steps may need to be explicitly taught or externalized, lending credence to CoT monitoring.