Research map / Image generation / Editing and style transfer
Resolving Lexical Bias in Model Editing
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.
Proposes PENME, a contrastively trained projection network over LLM layer activations plus a key-value codebook, to disentangle lexical and semantic similarity so model edits fire on paraphrases but not on lexically overlapping irrelevant prompts.
Problem
Adapter-based weight-preserving model editors (GRACE, MELO) use semantic similarity in representation space to scope edits, but lexical bias causes misfires on irrelevant prompts with overlapping words while failing to generalize to paraphrases.
Approach
Learn a 2-layer MLP projection over one layer's feed-forward activations with contrastive loss that pulls edits toward paraphrases and pushes away lexically similar irrelevant prompts and unrelated edits, then store one key per edit with a data-driven threshold in a codebook.
Key results
- On Counterfact, PENME reaches Edit Success 1.000 with Loc/Para of 0.847-0.869 and 0.875-0.906 across T5-small, Llama-2-7b and GPT2-XL, versus GRACE Para of 0.002-0.140.
- On zsRE, PENME scores 0.951 (T5-small), 0.984 (Llama-2-7b) and 0.966 (GPT2-XL) overall.
- Projector reduces the fraction of irrelevant prompts closer to edits than paraphrases from 46% to 6.4% for T5-small and to 0% for Llama-2-7b.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use a learned projection network to disentangle lexical and semantic representations | PENME model editing · Counterfact | Score | Helped |
| Use a learned projection network to disentangle lexical and semantic representations | PENME model editing · zsRE | Score | Helped |
| Apply PENME editing | Llama-2-7b · NLI | score: 0.6476 -> 0.6428 | Hurt |
| Apply PENME editing | Llama-2-7b · Sentiment Classification | score: 0.6573 -> 0.6573 | No clear effect |
| Apply PENME editing | Llama-2-7b · Summarization | score: 0.1865 -> 0.1865 | No clear effect |
| Apply PENME editing | GPT2-XL · NLI | score: 0.5128 -> 0.5128 | No clear effect |
| Apply PENME editing | GPT2-XL · Sentiment Classification | score: 0.4630 -> 0.4630 | No clear effect |
| Apply PENME editing | GPT2-XL · Summarization | score: 0.0936 -> 0.0936 | No clear effect |
| Use PENME retrieval with IKE for long-form generation | Llama-2-7b · Long-form generation | Fluency: 611.54 -> 622.36 | Helped |
| Use PENME retrieval with IKE for long-form generation | Llama-2-7b · Long-form generation | Reference Score: 16.57 -> 21.98 | Helped |
| Increase the contrastive-loss margin m from 10 to 40 (ablation) | GPT2-XL projector · GPT2-XL projector ablation | Generalization: 0.634 -> 0.967 | Helped |
| Increase the contrastive-loss margin m from 40 to 50 (ablation) | GPT2-XL projector · GPT2-XL projector ablation | Generalization: 0.967 -> 0.978 | Helped |
Surprising findings
- GRACE and MELO trade locality against paraphrase generalization sharply, e.g. GRACE on Llama-2-7b gets Loc 0.997 but Para 0.002 on Counterfact.
- Downstream NLI and summarisation performance of GPT2-XL is unchanged after PENME editing, and Llama-2-7b drops only on NLI.
Limitations
- Results on Counterfact are below perfect for locality and generalization (0.79-0.87), and misfires persist due to varying per-edit thresholds.
- Stores only exact strings as edit values, so long-form generation requires additional prompting via IKE-style retrieval.
- Projector trained on a relatively small number of edits (1000-5000) and needs to be retrained for new data; performance on T5-small lags other models.
Experimental setup
Datasets: Counterfact zsRE NQ DAiR-Emotions CNN/DailyMail RTE
Benchmarks: Counterfact zsRE DAiR-Emotions CNN/DailyMail RTE
Metrics: edit success locality generalization score ROUGE fluency reference score
Baselines: GRACE MELO MEMIT SERAC FT IKE
Builds on: GRACE MELO MEMIT SERAC IKE
Closest papers by meaning
- AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models (ICLR 2025 · similarity 0.89)
- Spectral Editing of Activations for Large Language Model Alignment (NeurIPS 2024 · similarity 0.89)
- Rethinking Residual Distribution in Locate-then-Edit Model Editing (NeurIPS 2025 · similarity 0.88)
- Should We Really Edit Language Models? On the Evaluation of Edited Language Models (NeurIPS 2024 · similarity 0.88)
- Uncovering Overfitting in Large Language Model Editing (ICLR 2025 · similarity 0.88)
- Editing Memories Through Few Targeted Neurons (AAAI 2025 · similarity 0.88)
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.