IC-370Knowledge entropy (sparsity of FFN memory coefficients) decreases consistently during pretraining for OLMo 1B, 7B, and Pythia 1.4B, and this decrease strongly correlates with reduced knowledge acquisition and increased forgetting in continual learning

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

The paper measures how broadly a model activates its feed-forward memory vectors (the rows of the second projection matrix V) by computing the entropy of the average memory coefficient distribution across all tokens. For OLMo 1B, 7B, and Pythia 1.4B, this knowledge entropy declines monotonically as pretraining progresses, with the sharpest drop in the last layer. In a continual knowledge learning setup (further training on PubMed or C4 with injected fictional knowledge), models from later pretraining stages show lower acquisition rates and higher forgetting rates. The Pearson correlation between knowledge entropy and acquisition is 0.94 (p=6e-5) and with forgetting is -0.96 (p=1e-5). Mid-stage checkpoints (~50% of the learning rate schedule) achieve the best balance of acquisition, retention, and overall performance.

Evidence
correlational
Key metric
Pearson correlation between knowledge entropy and acquisition: 0.94 (p=6e-5); with forgetting: -0.96 (p=1e-5). Baseline acquisition a(θ) across OLMo 1B pretraining steps 118k/369k/554k/738k: 25.2/24.8/21.4/18.8%; forgetting f(θ): 10.5/12.9/17.0/19.5%. Downstream performance p(θ): 54.0/54.6/52.8/52.0%.
Caveat
The study is limited to OLMo 1B and 7B (and Pythia 1.4B in the appendix) because they are the only models that publicly provide intermediate pretraining checkpoints. The authors note that extending to other models is future work. The mid-point is approximated as 50% of the learning rate schedule, and defining it precisely remains an open question.
Model
OLMo / OLMo base, Pythia
Datasets
Dolma [source], PubMed [train], C4 [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]
Methods
Continual Knowledge Learning [primary]
Related work
Geva et al. 2021 (Transformer feed-forward layers are key-value memories) [builds-on], Jang et al. 2022 (Towards continual knowledge learning of language models) [builds-on], Chang et al. 2024 (How do large language models acquire factual knowledge during pretraining?) [compared-to], Sun & Dredze 2024 (AMURO & CHAR) [context], Dohare et al. 2024 (Loss of plasticity in deep continual learning) [context]
Related findings
IC-371
Extraction
automatic-extraction