IC-349Larger LLMs (Llama-2 7B, Llama-3 8B) suffer more severe general ability degradation than smaller models (GPT-2 XL 1.5B) under the same number of sequential edits

Jun-Yu Ma, Hong Wang, Hao-Xiang Xu, Zhen-Hua Ling, Jia-Chen Gu

SourcePerturbation-Restrained Sequential Model Editing

When the same editing method (memit) is applied for 300 sequential edits, GPT-2 XL's downstream task performance remains largely unchanged, while Llama-2 and Llama-3 performance drops to nearly zero. The paper notes this as a general trend: 'larger models suffered more pronounced compromises in their general abilities when subjected to the same number of edits.' This scale-dependent susceptibility is observed across multiple downstream tasks and editing methods.

Evidence
correlational
Key metric
with 300 edits, memit's performance on gpt2-xl remained largely unchanged, whereas it dwindled to nearly 0 on llama-2 and llama-3
Caveat
Only three model sizes are tested (1.5B, 7B, 8B), and the two larger models are from the same family (Llama), so the comparison is not fully controlled for architecture.
Model
GPT-2 XL, Llama 2 / Llama 2 base, Llama 3
Concepts
Scale-dependent behaviour
Datasets
CounterFact / Counterfact dataset [eval], GSM8K [eval], SAMSum [eval], Natural Questions / NaturalQA [eval], RTE [eval]
Methods
MEMIT [primary], ROME [primary], MEND [primary]
Related findings
IC-348, IC-350
Extraction
automatic-extraction