ParaTempo introduces temporal confidence, a branch-local measure of how sharply recent answer-distribution probes concentrate on a dominant answer, and uses it as the sole control signal for parallel reasoning. Unlike prior approaches that rely on final-answer consensus, token-level confidence, or isolated probes, temporal confidence is computed asynchronously and is claimed to be more temporally stable and predictive of future convergence.1
The central constraint shift is computational. Parallel reasoning previously paid a cost that grows with reasoning depth and branch count, while control signals arrived too late or were too noisy to trigger branch-level interventions. ParaTempo makes pruning, early retirement, forking, and global stopping all respond to one asynchronous signal, eliminating the need for synchronization among trajectories.1
The claimed payoff is a 21.8% to 32.2% reduction in average latency and an 18.1% to 30.3% reduction in token usage on challenging mathematical and scientific reasoning benchmarks, with competitive accuracy. These figures establish the magnitude of the claimed improvement but are not yet tied to a specific benchmark list or exact accuracy numbers. The crucial open question is whether the efficiency gains survive when accuracy is held fixed; any replication showing larger accuracy loss would change this assessment.1
Caveats
AI hypothesisWhat would change my mind: if independent replications on a wider benchmark suite show accuracy drops beyond the stated competitive range, or if the probing overhead and confidence estimation cost dominate at large branch counts. The asynchronous forking behavior also assumes that reallocated computation can be fruitfully absorbed by new branches; this assumption needs empirical validation in domains where solution paths are not independent.1
