Research Statement

Language models are changed after release. Facts are edited, layers are merged, weights are pruned or compressed. A model can pass the usual output tests after such a change while something inside it has broken.

I read a trained model's activations as a graph and compare that graph before and after a change. Delta-Graph Signatures (SDM 2026, accepted) uses this comparison to flag knowledge edits that pass standard checks but damage unrelated knowledge. A separate CIKM 2026 paper gives a formal pre-condition for checking whether a knowledge graph structurally supports a query before a language model generates from it.

Next steps are steering, model stitching, and detecting silent degradation from other model changes: fine-tuning, merging, quantization, pruning and distillation. The colloquium "From Networks to Neural Networks" gives a tutorial-style overview of this line.

Music generation models now produce convincing audio, yet little is known about how they represent style. My first music paper fused two genres into a new one with adversarial dual learning (ICDM 2017, with Alexander Lerch).

Music is now a main and growing line of my work. The questions are how generation models encode style, how far that style can be steered, and how this can serve music for health and hearing. EEG and wearable sensors are being considered for the health side.