SPICE: Self-Play In Corpus Environments Improves Reasoning

TL;DR

SPICE introduces a reinforcement learning framework where a model generates and solves reasoning tasks from a large corpus, leading to continuous self-improvement and significant performance gains in reasoning benchmarks.

Contribution

The paper presents SPICE, a novel self-play framework with document grounding that enables models to generate and solve their own reasoning tasks for ongoing improvement.

Findings

Achieved +8.9% in mathematical reasoning benchmarks

Achieved +9.8% in general reasoning benchmarks

Document grounding is crucial for sustained self-improvement

Abstract

Self-improving systems require environmental interaction for continuous adaptation. We introduce SPICE (Self-Play In Corpus Environments), a reinforcement learning framework where a single model acts in two roles: a Challenger that mines documents from a large corpus to generate diverse reasoning tasks, and a Reasoner that solves them. Through adversarial dynamics, the Challenger creates an automatic curriculum at the frontier of the Reasoner's capability, while corpus grounding provides the rich, near-inexhaustible external signal necessary for sustained improvement. Unlike existing ungrounded self-play methods that offer more limited benefits, SPICE achieves consistent gains across mathematical (+8.9%) and general reasoning (+9.8%) benchmarks on multiple model families. Our analysis reveals how document grounding is a key ingredient in SPICE to continuously generate its own increasingly challenging goals and achieve them, enabling sustained self-improvement.