Research map / Large language models / Retrieval and knowledge
Precise and Interpretable Editing of Code Knowledge in Large Language Models
Min Xue, Nikolai Bolik, Lennart Stöpler, Erik Imgrund, Janik Schmid, Artur Andrzejak
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 |
|---|---|---|---|
| replace MLP with sparse TransCoder and update only active neurons (TCPE LTC4) instead of ROME | ROME · G4GD multi-error editing | Score: 109.60 -> 171.45 | Improved |
| TCPE with TransCoder instead of ROME on MLP | ROME · G4GD multi-error editing | post-edit accuracy: 57.33 -> 64.00 | Improved |
| use TCPE instead of MEMIT/PMET multi-layer edits | MEMIT · G4GD multi-error editing | Score: 14.96 -> 171.45 | Improved |
| widen TransCoder from 4x to 8x (ablation) | TCPE · G4GD multi-error editing | Score: 171.45 -> 174.82 | Improved |
| use TCPE instead of ROME, HumanEval reliability | ROME · HumanEval after G4GD edits | reliability (RE): 93.61 -> 100.00 | Improved |
| use TCPE instead of MEMIT, number of updated neurons | MEMIT · HumanEval after G4GD edits | updated neurons: 55,040 -> 69 | Improved |
| replace MLP layer with TransCoder (layer 19, LTC4) (ablation) | CodeLlama · G4GD overall accuracy | accuracy: 57.50% -> 58.33% | No clear effect |
| update only highly active neurons (acv>0.2) instead of all active neurons (ablation) | TCPE LTC4 · G4GD multi-error editing | Generalization (GN): 24.29 -> 32.86 | No clear effect |
| restrict update to low-activation or inactive neurons only (acv<=0.1) (ablation) | TCPE LTC4 · G4GD multi-error editing | Generalization (GN): 32.86 -> 7.14 | Hurt |
| use TransCoder features instead of MLP neurons | CodeLlama · Interpretable features out of 50 | count interpretable: 4 -> 33 | Improved |
| lower activation threshold to include all neurons (acv>0 vs acv>0.08) (ablation) | TCPE LTC4 C8 single-error · G4GD single-error editing C8 | Generalization (GN): 34.62 -> 28.85 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- WISE — WISE: Rethinking the Knowledge Memory for Lifelong Model Editing of Large Language Models (NeurIPS 2024) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Editing Memories Through Few Targeted Neurons (AAAI 2025 · similarity 0.89)
- Should We Really Edit Language Models? On the Evaluation of Edited Language Models (NeurIPS 2024 · similarity 0.88)
- Precise Localization of Memories: A Fine-grained Neuron-level Knowledge Editing Technique for LLMs (ICLR 2025 · similarity 0.88)
- AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models (ICLR 2025 · similarity 0.88)
- Unveiling the Pitfalls of Knowledge Editing for Large Language Models (ICLR 2024 · similarity 0.87)
- Guess & Sketch: Language Model Guided Transpilation (ICLR 2024 · similarity 0.87)
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.
