Researchers Introduce StochBench for Stochastic Processes in Lean

Researchers have released StochBench, a domain-specific benchmark for the Lean 4 interactive theorem proving system. The suite includes 450 graduate-level problems in stochastic processes across various abstraction levels, each paired with its natural-language source.

Addressing an area historically underrepresented in Mathlib, StochBench spans finite and countable Markov chains, random walks, martingales, renewal processes, queues, Brownian motion, stochastic calculus, and continuous-time Markov processes.

In initial evaluations, an Opus 4.8-based agent achieved a 34.9% proof rate, solving 157 of the 450 problems under a 15-minute time budget per problem. The authors note that the benchmark provides a realistic yet challenging target for modern formal reasoning models.