Audio reasoning is crucial for machines to understand the acoustic world, but current reinforcement learning (RL) reward designs have complementary limitations. Outcome-based rewards only check the final answer, allowing models to ignore the audio, while process-based rewards use fixed, hand-crafted criteria that are not tailored to each question or grounded in the acoustic evidence. As policies improve, static criteria become less effective.
To address this, researchers from Hugging Face introduce AudioRubrics, a novel RL framework that supervises audio reasoning with self-evolving, audio-grounded rubric rewards. The framework uses a single audio-capable model that consumes raw waveforms and acts as both rubric generator and judge, ensuring every criterion is anchored in the actual audio content.
AudioRubrics synthesizes per-sample rubrics from the waveform and, based on the model's own rollouts, regenerates and reweights criteria per group. It also prunes rubrics with constant verdicts and includes negatively phrased criteria to capture recurring failure modes. A linear length penalty prevents unbounded reasoning trace growth, and the rubric machinery is discarded after training, leaving a self-contained reasoner.
Comprehensive evaluations across three audio reasoning benchmarks show that AudioRubrics substantially outperforms a wide range of open-source and training-based baselines. The gains scale with the capability of the rubric generator and judge, and the framework converges to a stable reasoning length. Improvements in audio perception further demonstrate the effectiveness of anchoring supervision in acoustic evidence.