The Deterministic Horizon: Information-Theoretic Proof That Extended CoT Fails and Tool Use Is Necessary
The paper proves an Attention Bottleneck Theorem establishing information-theoretic limits on how far decoder-only transformers can track state in purely neural chain-of-thought. A Deterministic Horizon exists at approximately 19-31 steps beyond which accuracy collapses super-exponentially. Across 12 models and 8 task domains (SWE-Bench, WebArena, SQL-Multi), tool-integrated reasoning achieves 86-94% accuracy versus 24-42% for neural CoT. Fine-tuning improves performance by less than 5%, confirming the limits are architectural, not training-related. Accepted at ICML 2026.
Why it matters
Provides rigorous theoretical grounding for why agentic tool use is necessary — not just empirically better but provably required past a complexity threshold — setting a principled basis for agent architecture design.
Importance: 3/5
Official arXiv publication; accepted at ICML 2026; theoretical proof with broad empirical validation across 12 models.