Elamur Research⌕
☰
home / papers / 2026 / EvoEdit: Evolving Null-space Alignment for Robust and Efficient Knowledge Editing

EvoEdit: Evolving Null-space Alignment for Robust and Efficient Knowledge Editing

Paper · 2026 · Sicheng Lyu, Yu Gu, Xinyu Wang et al.

Summary

Locate-then-edit methods for updating facts in LLMs, such as ROME, MEMIT and AlphaEdit, break down when many edits are applied one after another. Each new change disturbs earlier ones, and the model can eventually collapse. EvoEdit targets this sequential setting. AlphaEdit projects updates into the null space of preserved knowledge using a fixed projector. EvoEdit instead refreshes the projector after every step, so it covers both the original knowledge and all facts edited so far. The authors call this sequential null-space alignment. They prove that outputs for already-stored keys stay unchanged across long edit sequences. For efficiency, they use the Woodbury matrix identity, which replaces a large solver with a small inner system. In experiments on CounterFact with Llama-3-8B and other LLMs, EvoEdit matches or beats the state of the art at 2K edits and stays stable up to 10K edits. It also runs faster and uses less memory. This makes continual fact maintenance in deployed models more practical.

Key points

  • Retention of early edits (Llama-3-8B, CounterFact): the first 100 edits were checked after 2,000 total edits.
    • AlphaEdit lost 53% rewrite accuracy and 34% paraphrase accuracy.
    • EvoEdit lost only 2% and 7%, respectively.
  • General abilities over 2,000 single edits: scores were tracked on SST, MRPC, MMLU and NLI.
    • ROME degraded sharply by about 400 edits, and MEMIT by about 800.
    • EvoEdit stayed largely stable throughout.
  • Large-scale editing: in a 10K-edit run (batch size 100), EvoEdit led the baselines on most or all reported metrics. The table caption and the main text disagree on whether this was four or five metrics.
  • Speed: on an NVIDIA H100 with 500 edits, EvoEdit's wall-clock time was 1.98× to 3.53× faster than AlphaEdit's.
  • Memory: over 1,000 edits, peak allocated GPU memory fell by up to 14% and peak reserved memory by up to 15%.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Université de MontréalMontreal · Canada68 works · 79,382 citations🎓McGill UniversityMontreal · Canada87 works · 11,027 citations🔬Mila - Quebec Artificial Intelligence InstituteMontreal · Canada54 works · 13,565 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

Cite

Cite as Sicheng Lyu, Yu Gu, Xinyu Wang, Jerry Huang, Sitao Luan, Yufei Cui, Xiao-Wen Chang and Peng Lu. 2026. EvoEdit: Evolving Null-space Alignment for Robust and Efficient Knowledge Editing. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.75

@inproceedings{lyu2026evoedit,
  title = {{EvoEdit: Evolving Null-space Alignment for Robust and Efficient Knowledge Editing}},
  author = {Sicheng Lyu, Yu Gu, Xinyu Wang, Jerry Huang, Sitao Luan, Yufei Cui, Xiao-Wen Chang and Peng Lu},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.75},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.75}
}

References

3 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. EvoEdit (s) Solve ↓ Total ↓ Solve ↓ Proj ↓ Total ↓ Llama-3-8B BS=1 BS=10 BS=100 1313.6 224.5 22.0 1313.6 224.5 22.0 1.4 0.3 0.1 661.0 141.7 8.8 662.4 (x1.98) 142.0 (x1.58) 8.9 (x2.47) Qwen2.5-7B BS=1 BS=10 BS=100 3251.3 461.6 39.9 3251.3 461.6 39.9 2.3 0.4 0.1 1491.4 169.0 11.2 1493.8 (x2.18) 169.4… find
  2. Anthropic. 2023. Model card: Claude 3. [Online]. Available: https://www.anthropic.com/ model-card-claude-3. *Table 3: Runtime on the MCF dataset across batch sizes (BS) for Llama-3-8B and Qwen2.5-7B-Instruct for 500 total edits. Solve: time to compute the update matrix; Proj: time to update the proj… ↗
  3. Peak Reserved (GB) MEMIT AlphaEdit EvoEdit 36.87 34.79 31.73 (-14%) 38.45 35.36 32.74 (-15%) *Table 4: GPU memory usage over 1000 edits on Llama-3-8B. EvoEdit reduces peak allocated and reserved memory compared to prior methods.* \#\# Conclusion find

Updated

This page in Markdown → https://md.research.elamur.ai/free/en/papers/lyu-2026-evoedit-evolving-null-space-alignment/