In mechanistic interpretability, tracking how weight spaces evolve during training usually presents severe computational barriers. FGRF addresses this by utilizing scale-dependent spectral estimators to map out structural integrity signatures on a single-pass budget.
In my accompanying preprints, I show how these trained-vs-untrained topological signatures reproduce predictably across diverse architectures, including Qwen2.5, SmolLM2, and GPT-2.
The complete codebase, reproduction scripts, and methodology documentation are now fully open to the research community.
GitHub Repository: https://github.com/RicPini/fgrf Preprint & Mathematical Framework reachable from the link above.
Full publications on SSRN: https://papers.ssrn.com/sol3/cf_dev/AbsByAuth.cfm?per_id=7758764
I welcome feedback, architectural critiques, or collaboration inquiries from fellow researchers and ML engineers working on network topology and interpretability.
#MechanisticInterpretability #DeepLearning #LLMs #GraphTheory #OpenSource