reasoning-limits

This paper demonstrates that extended chain-of-thought reasoning degrades performance on deterministic state-tracking tasks due to information-theoretic limits of decoder-only attention, and proposes tool delegation when the reasoning horizon exceeds a threshold.