ParaTempo: Efficient Parallel Reasoning via Temporal Confidence

Parallel reasoning—running multiple solution branches simultaneously and voting on the result—is one of the more reliable ways to improve accuracy on hard reasoning tasks. The catch is that it's expensive: every branch runs to completion whether or not it converged ten steps ago. ParaTempo attacks that waste directly, and its core insight is about when to trust a signal.

The paper opens with a careful diagnosis of existing control signals. Token-level entropy and perplexity fluctuate wildly because they respond to local linguistic texture rather than answer evolution—a branch can have low token uncertainty while still wandering between competing answers. Instantaneous answer confidence (probing the model mid-generation for a tentative answer distribution) is better aligned with the answer space but remains noisy: intermediate reasoning steps temporarily favor wrong hypotheses, so a single probe snapshot is unreliable. The authors measure this formally, showing that both token-level signals and instantaneous answer confidence have weak Spearman correlations and poor AUC when predicting whether a branch will stabilize.

Temporal confidence is the fix. Instead of reading a single probe, ParaTempo maintains a sliding window of recent probe distributions per branch, aggregates them, and computes the exponentiated negative entropy of the result. Intuitively: if the last several probes all concentrate probability on the same answer, the branch is converging; if mass is scattered, it isn't. This aggregation is entirely local to each branch—no cross-branch synchronization required.

From that single signal, four control actions follow. Branches below a dynamically calibrated pruning threshold (set per-problem during a warmup phase using a quantile of observed confidence values) are killed and their compute freed. Branches that have held a dominant answer with sufficient confidence for several consecutive probes are retired—they stop generating but keep their vote. When pruning frees a slot, ParaTempo forks a new branch from the highest-confidence active donor, inheriting its reasoning prefix but using a fresh sampling seed. And globally, generation stops early once the confidence-weighted vote concentrates on a single answer past a threshold.

The experimental results are on four hard benchmarks—competition mathematics (AIME 2026, HMMT November 2025, HMMT February 2026) and GPQA Diamond—using two models at different scales. Against standard self-consistency with a fixed branch budget, ParaTempo cuts average latency by roughly 21–32% and total token usage by roughly 18–30% while staying within about one accuracy point. Against Parallel-Probe, the closest prior method, it gains roughly 3.8–3.9 accuracy points while also reducing latency. The ablation confirms that all three operations contribute: removing forking hurts accuracy most, removing retirement inflates token counts and also costs accuracy, and removing pruning bloats latency with little accuracy gain.

The framework is training-free and runs on standard inference infrastructure. Its main constraint is the probing overhead—every active branch is periodically interrupted to extract a tentative answer distribution—and the hyperparameter surface (window size, probe interval, quantile threshold, retirement patience) is non-trivial, though the paper provides defaults. Whether those defaults transfer cleanly to domains outside mathematics and science remains an open question.