Reproduce “Induction Head Origins of SLAM-ASR Speech Recognition Models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Artem Fedorchenko, Tanel Alumäe

The authors analyze how SLAM-ASR uses a trained projector to connect a frozen speech encoder to a frozen language model. They identify induction heads by repeated-pattern matching; blocking their access to speech raises WER across three decoders, including Llama 3.2-3B from 3.3–3.6% to 7.1–8.3%. Random-head blocking does not, but effects vary by decoder and seed and do not establish exclusivity.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.