Subconscious, MIT, Princeton Paper Reframes Speculative Decoding as a Long-Horizon Reasoning Control Lever

A Subconscious / MIT / Princeton / Northeastern paper posted Nov 4 shows that speculative decoding — long treated as a pure throughput optimization — materially shifts frontier open-weight models' output distributions after about 30 steps in long-horizon agent tasks. The authors train draft models to be deliberately rejected at critical reasoning points, letting the target model encode those tokens at full precision. The technique improves accuracy on complex agentic benchmarks without changing the target model weights.