Stopping reasoning early cuts thinking compute up to 73%

TSDS halts a small model's reasoning once its planned action stops shifting. Only calls that still look uncertain escalate to a larger cloud model, per a paper posted on arXiv.

That split matters for anyone running a small-model-first setup with a cloud fallback. TSDS cut per-episode thinking compute 43 to 73 percent on three of four benchmarks, against baselines that only decide whether to escalate.

The system pairs two triggers. A convergence probe watches the model's intended action and stops local reasoning the moment that action quits changing between steps. A separate perplexity check flags calls where the model still looks unsure, and only those get sent up to a bigger model.

The thresholds aren't hand-tuned guesses. The paper calibrates the convergence cutoff and the escalation trigger together, using a method called Learn-Then-Test. That gives finite-sample guarantees on two numbers at once: expected episode reward, and how often the system calls out to the cloud model.

The paper doesn't say how the fourth benchmark, the one where the method didn't beat baselines, differs from the other three.