IC-371Artificially resuscitating inactive memory vectors by scaling the up-projection matrix K improves knowledge acquisition and reduces forgetting, with the effect more pronounced for later-stage OLMo models

Jiyeon Kim, Hyunji Lee, Hyowon Cho, Joel Jang, Hyeonbin Hwang, Seungpil Won, Youbin Ahn, Dohaeng Lee, Minjoon Seo

SourceKnowledge Entropy Decay during Language Model Pretraining Hinders New Knowledge Acquisition

To test whether low knowledge entropy causally limits acquisition, the authors modify the up-projection matrix K of the final-stage OLMo checkpoint: they identify the lowest p% of memory coefficients and multiply the corresponding rows of K by a factor u that scales them toward the layer mean, controlled by an amplifying factor q. With p fixed at 50, setting q to 1, 2, or 3 generally improves both acquisition and retention over the unmodified model, while q=0.5 (further suppressing inactive coefficients) worsens both. The benefit grows as the original model is from a later pretraining stage. However, even after resuscitation, performance remains below that of a mid-stage model with equivalent knowledge entropy, suggesting linear scaling of a subset of parameters is insufficient to fully restore plasticity.

Evidence
interventional
Caveat
The resuscitation method arbitrarily modifies model parameters and showed performance decline when applied to initial- or mid-stage models. The authors note that more refined methods preserving language modeling capabilities are needed. Performance after resuscitation remains lower than a mid-stage model with similar knowledge entropy.
Model
OLMo / OLMo base
Datasets
PubMed [train], Fictional Knowledge Dataset / Perez et al. (2022b) self-knowledge dataset [eval], SciQ / SciQA [eval], Winogrande [eval], PIQA [eval], OBQA [eval], HellaSwag [eval], ARC Easy [eval]
Related work
Geva et al. 2021 (Transformer feed-forward layers are key-value memories) [builds-on]
Related findings
IC-370
Extraction
automatic-extraction