One Mask to Rule Them All: On Hidden Facts after Editing and How to Find Them (2026.findings-acl)
Copied to clipboard
| Challenge: | Knowledge editing methods such as ROME and MEMIT update factual associations by modifying MLP weights. |
| Approach: | They propose to use a mask to reverse edits by eliminating overattention in later layers . they also show that injecting the mask during editing drops editing success from 98% to 38% . |
| Outcome: | The proposed method reverses edits by eliminating overattention in later layers and drops editing success from 98% to 38%. |
Similar Papers
Editing Common Sense in Transformers (2023.emnlp-main)
Copied to clipboard
Anshita Gupta, Debanjan Mondal, Akshay Sheshadri, Wenlong Zhao, Xiang Li, Sarah Wiegreffe, Niket Tandon
| Challenge: | Currently, the performance of transformer-based model editing methods is limited to statements about encyclopedic knowledge with a single correct answer. |
| Approach: | They propose to improve MEMIT's model editing algorithm by varying edit tokens and improving the layer selection strategy to improve commonsense knowledge. |
| Outcome: | The MEMIT editing algorithm outperforms baseline models on PEP3k and 20Q datasets while fine-tuning baselines shows significant trade-offs. |
Model Editing at Scale leads to Gradual and Catastrophic Forgetting (2024.findings-acl)
Copied to clipboard
| Challenge: | Existing model editing methods are evaluated using metrics for reliability, specificity and generalization over one or few edits. |
| Approach: | They evaluate model editing methods for three crucial properties - editing proficiency, fact forgetting and downstream performance. |
| Outcome: | The proposed methods are based on two state-of-the-art models - ROME and MEMIT. |
Editing the Mind of Giants: An In-Depth Exploration of Pitfalls of Knowledge Editing in Large Language Models (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Knowledge editing is a promising technique for updating factual knowledge in large language models (LLMs) but studies have identified side effects such as knowledge distortion and the deterioration of general abilities that have emerged after editing. |
| Approach: | They propose to evaluate the side effects of knowledge editing in large language models using metrics and benchmarks. |
| Outcome: | The results of the study highlight the limitations of current knowledge editing methods and outline potential research directions. |
MEMIT-Merge: Addressing MEMIT’s Key-Value Conflicts in Same-Subject Batch Editing for LLMs (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing knowledge editing techniques that modify models’ internal knowledge without full model retraining have gained significant attention. |
| Approach: | They propose an enhanced approach that merges value computation processes for facts sharing the same subject to improve editing efficiency. |
| Outcome: | The proposed method maintains a 98% editing success rate on same-subject and distinct-sub subject datasets while the original success rate drops to 46%. |
Revealing the Deceptiveness of Knowledge Editing: A Mechanistic Analysis of Superficial Editing (2025.acl-long)
Copied to clipboard
| Challenge: | Existing knowledge editing algorithms are prone to generating original knowledge . despite the fact that many models achieve near-perfect performance, superficial editing remains a challenge . |
| Approach: | They propose to use "**superficial editing**" to describe the phenomenon . they investigate the internal mechanisms of the attention module and their corresponding left singular vectors . |
| Outcome: | The proposed method can modify specific knowledge in a pretrained large language model while ensuring that unrelated knowledge remains unaffected. |
Detecting Edit Failures In Large Language Models: An Improved Specificity Benchmark (2023.findings-acl)
Copied to clipboard
| Challenge: | Recent model editing techniques can introduce large unwanted side effects, a new study shows . existing specificity benchmarks do not detect these unwanted side-effects . a recent study shows that model edits can cause significant performance drop . |
| Approach: | They extend existing CounterFact benchmark to include a dynamic component and propose a new benchmark to evaluate model editing techniques. |
| Outcome: | The proposed benchmark improves existing benchmarks for specificity and avoids unwanted side effects. |
Can We Edit LLMs for Long-Tail Biomedical Knowledge? (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing knowledge editing methods can enhance LLMs' performance on long-tail biomedical knowledge, but their performance on high-frequency popular knowledge remains inferior to that on high frequency popular knowledge. |
| Approach: | They conduct the first comprehensive study to investigate the effectiveness of knowledge editing methods for editing long-tail biomedical knowledge. |
| Outcome: | The proposed methods improve LLMs' performance on long-tail biomedical knowledge, but their performance on high-frequency popular knowledge remains inferior even after editing. |
Correct When Paired, Wrong When Split: Decoupling and Editing Modality-Specific Neurons in MLLMs (2026.acl-long)
Copied to clipboard
Tingchao Fu, Wenkai Wang, Fanxiao Li, Huadong Zhang, Jinhong Zhang, Dayang Li, Yunyun Dong, Renyang Liu, Wei Zhou
| Challenge: | Existing knowledge editing paradigms suffer from editing decoupling failures . entity knowledge is sequestered into disentangled modality-specific pathways . |
| Approach: | They propose a method that explicitly disentangles and localizes modality-specific neuron groups for targeted knowledge. |
| Outcome: | The proposed method outperforms baselines in reliability and consistency while preserving model locality. |
Do All Autoregressive Transformers Remember Facts the Same Way? A Cross-Architecture Analysis of Recall Mechanisms (2025.emnlp-main)
Copied to clipboard
| Challenge: | Prior work has identified MLP modules in early layers as key contributors to factual recall. |
| Approach: | They conduct a comprehensive evaluation of factual recall across several models . they find that Qwen-based models behave differently from previous patterns . |
| Outcome: | The results show that attention modules in the earliest layers contribute more to factual recall than MLP modules. |
Navigating the Dual Facets: A Comprehensive Evaluation of Sequential Memory Editing in Large Language Models (2024.acl-long)
Copied to clipboard
Zihao Lin, Mohammad Beigi, Hongxuan Li, Yufan Zhou, Yuxiang Zhang, Qifan Wang, Wenpeng Yin, Lifu Huang
| Challenge: | Memory Editing (ME) has emerged as an efficient method to modify erroneous facts or inject new knowledge into Large Language Models (LLMs). |
| Approach: | They propose to evaluate LLMs with single edit only and parameter-modifying ME with parameter-preserving ME. |
| Outcome: | The proposed method can maintain LLMs’ fundamental capabilities but struggles to accurately recall edited knowledge presented in a different format. |