mechanistic analysis of synthetic-data-induced model collapse
investigating how model collapse manifests inside language model internals and whether lost capabilities are gone or merely inaccessible
Quick Navigation: Overview • Research Progress
overview
Investigating how model collapse manifests inside language model internals — tracing where degradation originates, what drives it, and whether collapsed capabilities are destroyed or merely unreachable.
Timeline: August 2026 – Present
Advisor: Dr. Maria Leonor Pacheco, BLAST Lab
Status: In progress
research progress
- Investigating how model collapse manifests inside language model internals, showing recursively trained model generations are linearly decodable from activations with near-perfect accuracy.
- Comparing different training regimes to disentangle pure data drift from compounding parameter drift.
- Using neuron ablation and activation patching to causally localize where collapse-related information concentrates in the network.
- Testing whether activation steering can repair collapsed behavior, probing whether lost capabilities are gone or merely inaccessible.