Current direction / Ongoing
Learning Without Updating Weights
What can an audio-capable model learn from experience when its weights stay fixed? I am exploring how skills can emerge from feedback, memory, and the way tasks are organized around a model.
Why this interests me
Fine-tuning is only one way to change what a system can do. For audio tasks, I am interested in the space around a fixed model: how we describe the task, what examples or tools it can access, how feedback is recorded, and how the next attempt can use it.
The broader question is whether an agent can develop reusable audio skills through its interaction with the world, rather than only through a new set of model weights. Text optimization is one way to study this: the instructions, strategies, and representations surrounding the model can themselves become objects of search and revision.
This is active work. I am sharing the question and motivation now; methods, results, and publications will be added when they are ready.
What this page will hold
- The working definition of an audio skill and how to evaluate one.
- Experiments on feedback, reusable strategies, and adaptation without weight updates.
- Links to papers, code, or project pages as the work matures.
For related audio research with a clinical focus, see AI for Health Signals & Care.