An audio LLM that grades its own reasoning kept improving where a frozen external judge plateaued, and it topped MMAU test-mini at 76.1 percent
AspireAI community · arXiv (cs.SD)
Reinforcement learning for audio models normally trains against a reward that stays put - a final-answer check, a hand-written process score, or an external judge frozen before training starts. AudioRSI removes that fixed reference: the same Audio LLM plays actor and critic, generating several reasoning paths per audio question and then scoring them pairwise, and the critic is periodically replaced by the current policy so the training signal improves alongside the model.
The payoff is visible in the training curve. On out-of-distribution MMAU test-mini accuracy, the run trained against a frozen external judge plateaus while the self-critic run keeps climbing and stays ahead from step 2k onward. Because the critic hears the same recording as the actor, it can reject reasoning that cites acoustic evidence the recording does not contain instead of rewarding reasoning that merely sounds plausible. Across four out-of-distribution benchmarks disjoint from the AVQA training distribution, AudioRSI reports 76.10 percent on MMAU test-mini, ahead of its own reproduced baseline Ke-Omni-R at 74.60 and of Audio Flamingo 3 at 73.30, and it also ranks first on MMSU and leads the 7B open-source tier on MMAR among methods with no external reward model. No reward specified how to reason, yet structured reasoning behaviours - logical reasoning, perception-aware analysis, reduced hallucination - emerge, corroborated by an external LLM-as-Judge pass and a keyword analysis over 1,000 reasoning paths per model.
The honest caveat is in the paper's own comparison table. AudioRSI and the reproduced Ke-Omni-R are the only two rows the authors evaluated themselves; every other row is a number reported by that system's authors, so this is context rather than a controlled comparison, and the authors note a 2 percent gap between their base-model reproduction and its reported score and warn that small margins should not be over-read. They also note that RL systems trained on other data or with an external reward model report higher scores than AudioRSI.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.