sequential-reasoning

Introduces the Complexity Ceiling Benchmark (CCB) that evaluates LLM reasoning decay as the number of sequential steps increases across three domains. Finds a consistent geometric per-step decay and that all models collapse on transitive social logic within 5 steps, even with strong overall accuracy.

This paper argues that large language models struggle with causal reasoning and long-horizon planning due to a mismatch between sequence prediction and reasoning over latent environment dynamics, and introduces the Latent Dynamics Inference perspective along with the Flux environment to study these limitations.