Paper: Deterministic Horizon Limits Pure Neural Reasoning

aronchick · x · 2026-08-31

This paper investigates why extended chain-of-thought reasoning degrades performance on deterministic state-tracking tasks. It introduces an "Attention Bottleneck" analysis showing capacity limits in decoder-only attention and defines a "Deterministic Horizon" (19-31 steps) where unaided accuracy drops below 50%. Experiments on SWE-Bench and WebArena show tool-integrated reasoning achieves 76-94% accuracy versus 17-42% for neural CoT. The authors argue state should be external and deterministic.

Original post →

More from Research

Research channel →