rl-alignment

The paper reveals that exact-match verifiers in reinforcement learning with verifiable rewards introduce language-dependent biases in multilingual mathematical reasoning, proposes a protocol for auditing, and identifies a cross-lingual selection bottleneck.

This paper presents C-Guard, a constitution-grid instrument for data-efficient RL alignment of safety guards, using C-LIM learnability scores to prune, densify, amend, or expand training data. It reduces over-refusal from 22.4% to 12.8% and improves data efficiency.