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.
More from Research
- Medusa from training to inference: a two-part guide to multi-token prediction acceleration — No_Progress_5399 · 2026-08-31
- LayerRecall: layer-wise memory routing for long-horizon consistency in video generation — zju · 2026-08-31
- EASEL benchmark: multimodal agents fail at dexterous, closed-loop visual tool use — EASEL-Bench · 2026-08-31
- StarHarness: evolving fixed-weight agent harnesses for enterprise tool use — ServiceNow-AI · 2026-08-31
- ECCV 2026 Paper: Fourier Self-Supervision for Category Discovery — y_m_asano · 2026-08-31
- RAG poisoning causes 'attention collapse', fooling confidence detectors — rohanpaul_ai · 2026-08-31