AI & MACHINE LEARNING

AMR-SD consistently outperforms state-of-the-art baselines across diverse benchmarks, demonstrating superior generalization and problem-solving capabilities in scientific, mathematical, and tool-use tasks.

AMR-SD’s architectural design, including temporal annealing and asymmetric modulation, ensures unparalleled training stability and robust performance even in late-stage learning, preventing the common pitfalls of self-distillation methods.

Unlike baselines that exhibit early gains followed by plateaus or regression, AMR-SD achieves superior stability and avoids late-stage performance degradation across different model architectures (Figure 2). Its fine-grained credit assignment, aided by the CIG threshold and temporal annealing, not only stabilizes the RL training process but also significantly enhances the model's true generalization capacity for complex mathematical reasoning. This robust approach ensures that models do not suffer from severe performance collapse post-annealing, a common issue observed in methods like RLSD (Figure 3).

The system maintains a stable, monotonically improving reward signal and successfully bridges in-domain optimization with out-of-domain generalization. Crucially, by introducing a reflection bottleneck, AMR-SD mitigates vulnerabilities inherent in direct answer conditioning, effectively translating training stability to superior out-of-domain accuracy and enhancing reasoning efficiency by reducing unnecessary verbosity while successfully preserving critical cognitive explorations.