Independent ML researcher.
Recent work on audio-language models: how they internally process audio and when they behave like ASR→LLM cascades. I'm also extending into representation geometry, interpretability, and safety — exploring how representations form and reorganize during training, and when linear interpretability methods can be expected to work. Working papers on arXiv.
Formalizes modality collapse in multimodal LLMs as a mismatched-decoder problem; information-theoretic bound shows accessible information is limited by decoder sensitivity and distributional mismatch.
Shows that speech LLMs often reduce to implicit ASR→LLM cascades, with mechanistic evidence (logit lens, concept erasure) that text representations are causally required for task performance.
Introduces a structured family of local-linearity metrics that predict where linear interpretability tools succeed and how they fail.
Controlled studies of representation geometry during training, revealing task-specific collapse floors, top-down layer reorganization, and RankMe as a precursor to capability acquisition.
Dataset and evaluation showing that audio-LLMs systematically follow conflicting text over spoken audio across 8 languages, even under explicit instructions to follow the audio.
Layer-wise decomposition of how fine-tuning distributes across transformer depth, with implications for parameter-efficient adaptation.
Cross-lingual transfer methods for improving ASR in low-resource languages using non-target language data.
A technical explainer — how audio gets into, through, and out of a language model. Covers representation, rate, alignment, and output mismatches, with interactive figures.
Open to collaborations across any of these areas. If something overlaps with your work, drop me a line.