RuleReasoner: Reinforced Rule-based Reasoning via Domain-aware Dynamic Sampling
This work addresses the problem of improving rule-based reasoning efficiency and generalization for AI systems, representing an incremental advancement through a novel sampling method.
The paper tackles the challenge of enabling small reasoning models to perform rule-based reasoning with robust generalization across diverse tasks and domains, achieving significant performance gains over frontier large reasoning models, such as a 4.1% average improvement on in-distribution tasks and 10.4% on out-of-distribution tasks.
Rule-based reasoning has been acknowledged as one of the fundamental problems in reasoning, while deviations in rule formats, types, and complexity in real-world applications pose severe challenges. Recent studies have shown that large reasoning models (LRMs) have remarkable reasoning capabilities, and their performance is substantially enhanced by reinforcement learning (RL). However, it remains an open question whether small reasoning models (SRMs) can learn rule-based reasoning effectively with robust generalization across diverse tasks and domains. To address this, we introduce Reinforced Rule-based Reasoning, a.k.a. RuleReasoner, a simple yet effective method to conduct rule-based reasoning via a wide collection of curated tasks and a novel domain-aware dynamic sampling approach. Specifically, RuleReasoner resamples each training batch by updating the sampling weights of different domains based on historical rewards. This facilitates domain augmentation and flexible online learning schedules for RL, obviating the need for pre-hoc human-engineered mix-training recipes used in existing methods. Empirical evaluations on in-distribution (ID) and out-of-distribution (OOD) benchmarks reveal that RuleReasoner outperforms frontier LRMs by a significant margin ($Δ$4.1% average points on eight ID tasks and $Δ$10.4% average points on three OOD tasks over OpenAI-o1). Notably, our approach also exhibits higher computational efficiency compared to prior dynamic sampling methods for RL.