Research map / Large language models / Retrieval and knowledge
KnowledgeSmith: Uncovering Knowledge Updating in LLMs with Model Editing and Unlearning
Yinyi Luo, Zhexian Zhou, Hao Chen, Kai Qiu, Marios Savvides, Sharon Li, Jindong Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use editing instead of unlearning | LLM knowledge update · KnowledgeSmith over-spreading (CCR) | share of failures: 15 -> 35 | Hurt |
| use unlearning instead of editing | LLM knowledge update · KnowledgeSmith under-forgetting (RR) | share of failures: 20 -> 35 | Hurt |
| use editing instead of unlearning | LLM knowledge update · KnowledgeSmith conflict emergence | share of failures: 12 -> 30 | Hurt |
| use editing instead of LoRA at k=10 samples | Llama3-8B-Instruct · ID accuracy | accuracy (%): 16.7 -> 25 | Improved |
| increase LoRA training samples from 1 to 1000 | Llama3-8B-Instruct LoRA · OOD accuracy (MMLU) | accuracy (%): 63.0 -> 61.6 | Hurt |
| increase LoRA training samples to 1000 | Llama3-8B-Instruct LoRA · ID accuracy | accuracy (%): 12.5 | Hurt |
| use editing instead of unlearning | LLM knowledge update · training time on 1000 samples (H100) | hours: 0.2 -> 6 | Hurt |
| increase training samples from 1 to 10000 for unlearning | ReLearn · KL representation similarity | normalized KL similarity: 0.014 -> 0.883 | Improved |
| increase training samples from 1 to 10000 for editing | AlphaEdit · KL representation similarity | normalized KL similarity: 0.140 -> 0.652 | Improved |
| use larger models for updates | LLaMA3 family · plasticity under few-shot data | immediate plasticity | Hurt |
| apply update at root rather than leaf node | editing and unlearning · KnowledgeSmith hierarchy | achievable in-domain accuracy | Hurt |
| update history-subject knowledge instead of other subjects | editing and unlearning · KnowledgeSmith subjects | update accuracy | Hurt |
Closest papers by meaning
- Mitigating Heterogeneous Token Overfitting in LLM Knowledge Editing (ICML 2025 · similarity 0.92)
- UniEdit: A Unified Knowledge Editing Benchmark for Large Language Models (NeurIPS 2025 · similarity 0.91)
- Unveiling the Pitfalls of Knowledge Editing for Large Language Models (ICLR 2024 · similarity 0.91)
- AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models (ICLR 2025 · similarity 0.90)
- Uncovering Overfitting in Large Language Model Editing (ICLR 2025 · similarity 0.90)
- Golden Layers and Where to Find Them: Improved Knowledge Editing for Large Language Models Via Layer Gradient Analysis (NeurIPS 2026 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
