Research map / Large language models / Retrieval and knowledge
Knowledge in Superposition: Unveiling the Failures of Lifelong Knowledge Editing for Large Language Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Shows that lifelong knowledge editing via the ROME/MEMIT closed-form solution yields an interference term caused by knowledge superposition in whitening space, and measures this superposition as universal across GPT-2, Pythia, GPT-J and Llama models with a scaling law.
Problem
Locate-then-edit methods (ROME, MEMIT, PMET, WilKE) degrade severely after many sequential edits, but the underlying reason for lifelong knowledge editing failure has rarely been explored.
Approach
Extends the linear-associative-memory closed-form editing solution to lifelong editing, derives an interference term proportional to a whitening-space coefficient p(.,.), and interprets that coefficient as knowledge superposition; lossless editing requires orthogonal knowledge representations in whitening space.
Key results
- Knowledge superposition (non-zero off-diagonal p values in whitening space) appears in all layers of all 11 models examined, with heavy-tailed, zero-mean distributions.
- Kurtosis rises with model size within GPT-2, Pythia and Llama-2 families, indicating weaker superposition in larger models.
- Angular distributions in whitening space concentrate around 90 degrees for Llama3-8B and Llama3.1-8B, while activation-space angles show no orthogonal tendency.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Continue applying ROME edits | GPT-J · Unrelated test knowledge during lifelong editing | confidence in correct knowledge | Hurt |
| Increase model size within a model family | Language models · Knowledge superposition across layers | kurtosis of the superposition distribution | Helped |
| Use later rather than earlier layers | Language models · Knowledge superposition across layers | kurtosis | Hurt |
| Measure representations in whitening space rather than activation space | Llama3-8B and Llama-3.1-8B · Knowledge representation angles | orthogonal tendency | Helped |
| Make knowledge representations orthogonal in whitening space | Lifelong knowledge editing · Original and previously edited knowledge | interference | Helped |
| Increase model size | GPT2, Pythia, and Llama-2 families · Knowledge superposition | degree of superposition | Helped |
| Use Llama2 architecture rather than GPT2 or Pythia architecture | Language models · Knowledge superposition | degree of superposition | Hurt |
| Use larger models to represent knowledge | Language models · CounterFact knowledge representations | superposition | Helped |
Surprising findings
- Pairs such as 'Windows 8.1' and 'Mac OS X 10.1' reach p=1.00 in GPT-J layer 0, so editing one may impact the other.
- Earlier layers show higher kurtosis (less superposition) than later layers.
Limitations
- Some appendix experiments were removed due to arXiv restrictions.
- The lifelong-editing failure is demonstrated mainly with a single ROME/GPT-J experiment in the appendix rather than a broad benchmark.
- Analysis uses 128 CounterFact subject items per model, so distribution estimates depend on a small sample.
Experimental setup
Datasets: CounterFact Wikipedia 20220301.en
Benchmarks: CounterFact
Metrics: superposition coefficient p(.,.) kurtosis kernel density estimate angular distribution
Baselines: ROME MEMIT PMET WilKE
Builds on: ROME MEMIT linear associative memory toy models of superposition
Closest papers by meaning
- WISE: Rethinking the Knowledge Memory for Lifelong Model Editing of Large Language Models (NeurIPS 2024 · similarity 0.91)
- Unveiling the Pitfalls of Knowledge Editing for Large Language Models (ICLR 2024 · similarity 0.90)
- Uncovering Overfitting in Large Language Model Editing (ICLR 2025 · similarity 0.90)
- Towards Lifelong Model Editing via Simulating Ideal Editor (ICML 2025 · similarity 0.89)
- History Matters: Temporal Knowledge Editing in Large Language Model (AAAI 2024 · similarity 0.89)
- Revealing and Mitigating Over-Attention in Knowledge Editing (ICLR 2025 · 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.