A unified learning-dynamics view ties data attribution, forgetting, and plasticity loss to token interactions

Yi Ren · hf · 2026-10-01

This paper derives a token- and layer-wise decomposition of how learning one token changes another prediction, separating softmax force, shared readout geometry, and residual connections. Positive interactions identify useful experience; negative ones cause concentrated collisions or accumulated erosion; over time updates reshape the readout geometry that mediates future learning. This single evolving interaction unifies data attribution, forgetting, and plasticity loss, yielding practical data selection, interference controls, and a readout-based diagnostic of future learnability.

Original post →

More from Research

Research channel →