Research map / Architectures and efficiency / RNNs and normalization
Reasons and Solutions for the Decline in Model Performance after 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.
Investigates why knowledge-edited LLMs degrade, attributing it to editing-target diversity/token length (via a new MQD dataset) and L1-norm growth of edited layers, and proposes D4S, an O(1)-space sequence editing method that accumulates edit statistics to suppress norm growth.
Problem
Edited language models often show performance degradation after knowledge editing, and the causes are unexplained and no mitigation exists; sequence editing also hits a bottleneck in the number of successful edits.
Approach
Attributes post-edit degradation to diversity/perplexity of editing targets and to explosive L1-norm growth of edited MLP layers, then proposes D4S, which stores running sums R K^T and K K^T so each new edit is merged into a single batch update in O(1) space, keeping norms from exploding.
Key results
- On ZsRE with 1000 sequential edits on Llama2, D4S reaches 81.88 average vs 15.53 for ROME and 2.82 for MEMIT; on GPT-J it reaches 88.77 average.
- On CounterFact, D4S attains 69.93 average on GPT-J vs 59.00 for MEMIT; on MQuAKE it attains 64.93 vs 0.00 for MEMIT and ROME.
- Llama2 downstream performance after 10,000 D4S edits drops only slightly (e.g., HellaSwag 57.14 to 52.99, MMLU 41.35 to 39.76).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use D4S instead of the strongest baseline by Avg. | GPT-J (6B) and Llama2 (7B) · ZsRE, 500 and 1000 edits | Avg.: 42.99 -> 85.34; 15.53 -> 81.88; 73.60 -> 88.40; 66.69 -> 88.77 | Helped |
| Use D4S instead of the strongest baseline by Avg. | GPT-J (6B) and Llama2 (7B) · Counterfact | Avg.: 66.83 -> 69.93; 66.71 -> 71.93 | Helped |
| Use D4S instead of the strongest baseline by Avg. | GPT-J (6B) and Llama2 (7B) · Mquake | Avg.: 7.74 -> 64.93; 52.97 -> 62.05 | Helped |
| Edit the model with D4S for 10,000 edits | Llama2 (7B) · Downstream tasks | accuracy: 43.34 -> 42.41 (↓0.93); 57.14 -> 52.99 (↓4.15); 41.35 -> 39.76 (↓1.59); 38.97 -> 38.56 (↓0.41); 69.14 -> 68.75 (↓0.39) | Hurt |
| Use direct-generation objectives rather than multiple-choice or true/false objectives | ROME-edited model · MQD | model performance | Hurt |
| Use editing objectives with higher perplexity | ROME-edited model · MQD | model performance: 12.3, 43.3, and 297.4 | Hurt |
| Increase the number of sequence edits | ROME · ZsRE | editing probability | Hurt |
| Increase the number of sequence edits beyond 850 | MEMIT · ZsRE | editing probability | Hurt |
| Use D4S sequence editing to consolidate edit history | MEMIT-style sequence editing · Sequence editing | L1-norm growth | Helped |
Surprising findings
- MEMIT's editing collapses to near zero after about 850 sequential edits, and ROME's effective edits stop around 100, despite both working well for single or few edits.
- Editing the 20Q dataset (binary 0/1 targets) causes less degradation for ROME and MEMIT than other datasets.
Limitations
- Only GPT-J (6B) and Llama2 (7B) were tested due to compute limits.
- D4S training was limited to 2000 steps; larger edit counts were not explored.
- MEMIT and PMET were run with batch size 1 only, so batch editing was not fully assessed.
Experimental setup
Datasets: ZsRE ELKEN 20Q CKD ATOMIC CounterFact MQuAKE
Benchmarks: ZsRE CounterFact MQuAKE ARC-challenge HellaSwag MMLU TruthfulQA WinoGrande GSM8K
Metrics: Efficacy Paraphrase Specificity Average accuracy probability value L1-norm
Baselines: FT ROME MEMIT PMET GRACE
Builds on: ROME MEMIT PMET GRACE
Closest papers by meaning
- Should We Really Edit Language Models? On the Evaluation of Edited Language Models (NeurIPS 2024 · similarity 0.92)
- Unveiling the Pitfalls of Knowledge Editing for Large Language Models (ICLR 2024 · similarity 0.90)
- Rethinking Residual Distribution in Locate-then-Edit Model Editing (NeurIPS 2025 · similarity 0.89)
- Uncovering Overfitting in Large Language Model Editing (ICLR 2025 · similarity 0.88)
- UniEdit: A Unified Knowledge Editing Benchmark for Large Language Models (NeurIPS 2025 · similarity 0.88)
- History Matters: Temporal Knowledge Editing in Large Language Model (AAAI 2024 · 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.