Beyond Text Following: Repairable Arbitration Reversals in Audio-Language Models

The key finding isn't just that models prefer text over audio, it's that the audio signal is correctly encoded and then overridden somewhere in the answer-generation layers, meaning this is an arbitration failure rather than a perception failure. That distinction matters enormously for how you'd fix it: retraining the input pipeline won't help if the problem lives in late-stage token generation.

This connects directly to the financial LLM bias audit covered on June 1st ('Auditing Asset-Specific Preferences in Financial Large Language Models'), which used a similar causal isolation approach to show that internal representations drive outputs in ways that contradict surface-level evidence. Both papers are essentially doing the same diagnostic work: tracing where a model's stated output diverges from what its internal state actually encodes. The multi-domain RL interference paper from the same date adds another angle, showing that shared computational pathways can cause one modality or domain to systematically suppress another during training, which is a plausible mechanism for exactly the text-dominance bias described here.

Watch whether any of the five tested ALMs release updated training documentation or fine-tuning guidance that explicitly addresses modality weighting in conflict scenarios within the next six months. If none do, it signals the field is treating this as an evaluation curiosity rather than a deployment risk.