- Rating
- 3
- Confidence
- 3
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes KEY, a unified framework for knowledge editing and unlearning in large language models (LLMs), aiming to mitigate knowledge entanglement during model updates. Through preliminary analyses of parameter overlap and responses to perturbations, the authors demonstrate that overlapping parameter representations do not necessarily imply that different knowledge is inseparable in weight space.
Motivated by these findings, KEY introduces a two-stage framework consisting of Knowledge Addressing Dictionary Construction and Task-Specific Sparse Addressing. The former constructs an orthogonal dictionary from input- and output-side gradients using SVD, while the latter identifies sparse combinations of dictionary atoms to perform targeted knowledge updates. The paper further establishes local and global bounded-proximity guarantees under certain regularity assumptions, providing theoretical support for the stability of knowledge addressing.
Extensive experiments on knowledge editing (CounterFact and FINE-KED) and unlearning (TOFU and MUSE-News/Books), across multiple backbone LLMs, demonstrate that KEY achieves competitive or superior target modification performance while substantially improving neighboring knowledge preservation and general model utility compared with existing baselines.
Critical strengths
1. Novel and well-motivated perspective on knowledge entanglement.
The paper provides an interesting perspective on knowledge entanglement by investigating whether overlapping parameter representations necessarily prevent knowledge separation. The preliminary experiments on parameter overlap and divergent responses to perturbations offer intuitive empirical motivation for the proposed approach. Building an orthogonal dictionary from input- and output-side gradients provides an interpretable mechanism for addressing different knowledge through shared model parameters. This is a meaningful strength because it offers a principled alternative to directly modifying overlapping parameters and provides a coherent motivation for the framework's design.
2. Unified formulation supported by extensive empirical and theoretical analysis.
KEY accommodates both knowledge editing and unlearning within the same dictionary construction and sparse addressing framework, with the two tasks primarily distinguished by their update objectives. This unification is appealing because these tasks are often investigated separately despite sharing the challenge of modifying targeted knowledge while preserving unrelated knowledge. Moreover, the paper evaluates KEY across multiple benchmarks and backbone models, complemented by ablation studies and theoretical analysis of addressing stability. The consistent improvements in knowledge preservation and general utility provide substantial empirical support for the proposed method.
Critical weaknesses
1. Strong dependence on representative global knowledge and retained sets limits practical applicability.
KEY relies heavily on a global knowledge pool for dictionary construction and an explicitly defined retained set for atom filtering and coefficient optimization. The ablation results in Table 3 reveal substantial performance degradation when the dictionary is constructed using only the target knowledge : NS on CounterFact decreases from 83.3% to 56.5%, while FQ on TOFU drops from 0.69 to 0.42. These results suggest that KEY's effectiveness depends strongly on access to sufficiently representative auxiliary knowledge.
This assumption may be difficult to satisfy in realistic open-world settings, where knowledge updates arrive dynamically and the distribution of future editing requests is unknown. Furthermore, collecting and maintaining representative knowledge and retained sets may introduce substantial additional costs. This concern is consequential because it raises questions about whether KEY's reported advantages generalize beyond the relatively controlled experimental settings considered in the paper.
2. Unclear end-to-end scalability and lack of validation for sequential knowledge updates.
KEY introduces potentially substantial upfront computational and storage overhead. According to Table 7, the framework constructs dictionaries for editable weight matrices across all layers by default. This requires collecting token-level activations and gradients, performing SVD, and storing up to dictionary atoms for each weight matrix. Although Appendix E.5 demonstrates efficient coefficient optimization after dictionary construction, it remains unclear whether the overall framework is computationally and memory-efficient when preprocessing costs are included, particularly for larger LLMs.
Additionally, the experiments focus on individual updating tasks without evaluating sequential or lifelong knowledge updates. In practical deployments, editing and unlearning requests typically arrive continuously, making it important to understand whether a fixed dictionary remains effective over time, whether repeated updates interfere with previously modified knowledge, and whether dictionary reconstruction is necessary. These limitations collectively leave the scalability and long-term practicality of KEY insufficiently established.
Questions for authors
1. Robustness to the selection and distribution of auxiliary knowledge.
Table 3 suggests that constructing the dictionary solely from target knowledge leads to substantial performance degradation. How sensitive is KEY to the size, diversity, and distributional similarity of the global knowledge pool and retained set ? In particular, can the authors provide experiments using smaller or distribution-mismatched auxiliary knowledge sets to clarify the minimum requirements for maintaining effective knowledge addressing?
2. End-to-end computational and memory overhead.
Could the authors report the complete preprocessing cost of dictionary construction, including gradient collection, SVD computation, peak GPU memory consumption, and dictionary storage, alongside the per-update optimization cost? How do these costs scale with model size and the number of dictionary atoms? Such measurements would help determine whether KEY's improvements justify its additional overhead compared with existing editing and unlearning methods.
3. Applicability to sequential knowledge updates.
Can the authors clarify whether a dictionary constructed once can be reused for a sequence of editing or unlearning requests without significant performance degradation? Specifically, how does KEY handle interference between successive updates, and under what circumstances would dictionary reconstruction become necessary? Even a preliminary sequential updating experiment would help assess the framework's practical applicability.
8 comments
[Part 2] Respond to reviewer P1qk
Continue to weakness 1.
As shown in Tables A--B, replacing the original knowledge pool with while keeping the dictionary fixed degrades both unlearning and editing performance. However, dynamic dictionary updating substantially improves performance on both tasks. On TOFU, for instance, FQ increases from 0.46 to 0.65, while MU improves from 0.53 to 0.62. Notably, the resulting FQ ranks third and MU ranks second among the unlearning methods compared in Table 1.
These results suggest that dynamic dictionary updating offers a practical extension of KEY when future update requests are unknown or the initial knowledge pool provides insufficient coverage. Beyond achieving competitive performance, this strategy maintains relatively low computational overhead, as the SVD-based dictionary construction in KEY naturally supports incremental updates by reusing existing singular directions. This provides a promising direction for addressing the challenges of dynamic knowledge updating without relying on a fully representative global knowledge pool.
Detials and experimental settings for dynamic dictionary updating are provided in Appendix E.10.
[1] Maini et al. TOFU: A Task of Fictitious Unlearning for LLMs. COLM 2024.
[2] Meng et al. Locating and Editing Factual Associations in GPT. NeurIPS 2022.
[3] Hendrycks et al. Measuring Massive Multitask Language Understanding. ICLR 2021.
[4] Joshi et al. TriviaQA: A Large Scale Distantly Supervised Challenge Dataset for Reading Comprehension. ACL 2017.
[5] Welbl et al. Crowdsourcing Multiple Choice Science Questions. WNUT 2017.
[6] Kalantzis et al. Projection Techniques to Update the Truncated SVD of Evolving Matrices with Applications. ICML 2021.
Weakness 2. Unclear end-to-end scalability and lack of validation for sequential knowledge updates. KEY introduces potentially substantial upfront computational and storage overhead. According to Table 7, the framework constructs dictionaries for editable weight matrices across all layers by default. This requires collecting token-level activations and gradients, performing SVD, and storing up to dictionary atoms for each weight matrix. Although Appendix E.5 demonstrates efficient coefficient optimization after dictionary construction, it remains unclear whether the overall framework is computationally and memory-efficient when preprocessing costs are included, particularly for larger LLMs. Additionally, the experiments focus on individual updating tasks without evaluating sequential or lifelong knowledge updates. In practical deployments, editing and unlearning requests typically arrive continuously, making it important to understand whether a fixed dictionary remains effective over time, whether repeated updates interfere with previously modified knowledge, and whether dictionary reconstruction is necessary. These limitations collectively leave the scalability and long-term practicality of KEY insufficiently established.
Thank you for this constructive comment. The computational and storage cost of KEY is addressed first, followed by new experiments on sequential knowledge updates.
1. Cost of dictionary construction
Dictionary construction involves gradient extraction, and SVD. Let denote the number of knowledge items used for construction, denote the total number of token positions collected across knowledge set (Line 228), and the cost of one forward–backward pass per item. For each weight matrix , and denote its input and output dimensions, respectively. Gradient extraction requires . Meanwhile, SVD composition of and costs . Thus, the total time complexity of these stages is .
Empirically, dictionary construction is evaluated using the TOFU-finetuned Llama-3.1-8B on a single NVIDIA A100 80 GB GPU, with 4,000 TOFU question–answer pairs. The dictionary covers 7 projection matrices per Transformer layer, giving 224 matrices across 32 layers in total. The input- and output-side ranks are set to . As shown in Table C, the entire construction takes 17.94 minutes, of which SVD-based decomposition accounts for 61.8%, while gradient extraction and direction aggregation jointly account for 34.0%. The resulting dictionary occupies approximately 0.5 GB on disk, and the measured peak GPU memory usage during construction is approximately 67 GB. This analysis is further detailed in Appendix E.5.
[Part 3] Respond to reviewer P1qk
Continue to weakness 2.
Table C. Runtime (in minutes) breakdown for initial dictionary construction, while other overhead includes dictionary serialization and normalization.
\hline \text{Gradient extraction} & \text{SVD} & \text{Other overhead} & \text{Total} \\\\ \hline 6.10 & 11.09 & 0.75 & \mathbf{17.94} \\\\ \hline \end{array}$$ To assess scalability to larger LLMs, `Table D` reports dictionary construction times of 3.87, 17.94, and 31.18 minutes for Llama-3.2-1B, Llama-3.1-8B, and Llama-2-13B, respectively. Although construction time increases with model size, the dictionary for the 13B model can be constructed in approximately 31 minutes. This linear growth in dictionary construction time remains manageable. In sequential updating setting, the preprocessing cost can be amortized across multiple update requests through dictionary reuse. **Table D. Comparison of construction time across model sizes (minutes).** $$\begin{array}{|l|c|c|c|} \hline \text{Model} & \text{Llama-3.2-1B} & \text{Llama-3.1-8B} & \text{Llama-2-13B} \\\\ \hline \text{Construction time (min)} & 3.87 & 17.94 & 31.18 \\\\ \hline \end{array}$$ **2. Sequential knowledge updates** To examine whether a fixed dictionary remains effective over time, sequential-batch experiments are added for both editing and unlearning. In every experiment, the dictionary is constructed once before the first round and kept fixed across all rounds, without reconstruction. *Sequential editing.* Following the experimental setup of AlphaEdit [1], Llama-3-8B is evaluated on CounterFact with 100 facts per round over 20 consecutive rounds, totaling 2,000 edits. As shown in `Table E`, KEY achieves comparable or superior performance to AlphaEdit across all three evaluation metrics at every reported checkpoint. After 2,000 edits, KEY achieves 99.0\% Efficacy Success (ES), 95.3\% Paraphrase Success (PS), and 72.2\% Neighborhood Success (NS), outperforming AlphaEdit by 0.1, 1.0, and 4.4 percentage points, respectively. These results indicate that KEY maintains comparable editing effectiveness while improving paraphrase generalization and preservation of neighboring knowledge throughout sequential updates. **Table E. Performance comparsion (reported as ES / PS / NS (%); higher is better) for sequential-batch knowledge editing on CounterFact with Llama-3-8B.** $$\begin{array}{|l|c|c|c|c|} \hline \text{Method} & \text{Round 5} & \text{Round 10} & \text{Round 15} & \text{Round 20} \\\\ \hline \text{AlphaEdit} & 99.8 / 93.5 / 76.8 & 99.6 / 93.6 / 74.1 & 99.2 / 93.9 / 70.4 & 98.9 / 94.3 / 67.8 \\\\ \hline \mathbf{KEY} & 99.8 / 95.9 / 82.1 & 99.7 / 94.3 / 80.5 & 99.3 / 95.1 / 77.9 & 99.0 / 95.3 / 72.2 \\\\ \hline \end{array}$$ *Sequential unlearning.* Following the experimental setting of ASU [2], Llama-2-Chat-7B is evaluated on TOFU, with each round forgetting either 1\% or 5\% of the dataset's authors over 10 consecutive rounds. For a fair comparison, Forgetting Efficacy (FE) and Model Utility (MU) are adopted from [2] to measure forgetting performance on the current round forget set and on the remaining retain set, respectively. As reported in `Table F`, KEY consistently achieves higher MU than ASU at every reported checkpoint under both settings. With 1\% of authors removed per round, KEY also achieves higher FE at rounds 1, 5, and 10. Under the more demanding 5\% setting, although both methods exhibit fluctuations in FE across rounds, KEY attains a higher FE than ASU. Overall, these results indicate that KEY better preserves model utility while maintaining competitive forgetting effectiveness across sequential unlearning. **Table F. Performance comparsion (reported as FE / MU (%); higher is better) for sequential-batch knowledge unlearning on TOFU with Llama-2-Chat-7B .** $$\begin{array}{|c|l|c|c|c|c|} \hline \text{Authors removed} & \text{Method} & \text{Round 1} & \text{Round 3} & \text{Round 5} & \text{Round 10} \\\\ \text{per round} & & & & & \\\\ \hline 1\\% & \text{ASU} & 81.2 / 76.8 & 78.7 / 75.3 & 78.2 / 74.9 & 74.5 / 74.1 \\\\ \hline 1\\% & \mathbf{KEY} & 82.5 / 78.7 & 76.3 / 77.1 & 81.7 / 76.2 & 76.4 / 74.7 \\\\ \hline 5\\% & \text{ASU} & 76.6 / 73.6 & 75.1 / 72.0 & 75.2 / 71.4 & 77.2 / 67.0 \\\\ \hline 5\\% & \mathbf{KEY} & 78.1 / 75.5 & 79.2 / 74.8 & 75.5 / 74.1 & 77.9 / 71.7 \\\\ \hline \end{array}$$ Both results clearly demonstrate the effectiveness of KEY in sequential-batch knowledge updating, highlighting its ability to preserve selective addressing while reusing a fixed dictionary across successive rounds. We thank the reviewer again for this broader evaluation. Detailed experimental settings, complete results, and further analysis are included in `Section 5.3`.
[Part 1] Respond to reviewer P1qk
Dear Reviewer
P1qk,We sincerely thank the reviewer for the constructive and insightful feedback. The comments have helped us strengthen both the technical analysis and experimental evaluation of KEY. We provide detailed responses and additional results below.
We thank the reviewer for this insightful comment. To further investigate KEY's applicability to dynamic update requests and unknown distribution of future updates, the following experiment with dynamic dictionary updating is introduced.
Specifically, an initial general knowledge pool P′ is constructed by randomly sampling question--answer pairs from TOFU [1], CounterFact [2], MMLU [3], TriviaQA [4], and SciQ [5], with each dataset contributing 20% of the pool. The total size matches that used in the main experiments, i.e., ∣P′∣=∣P∣. All sampled items are standardized into a common question--answer format, and the initial dictionary D is constructed following
Section 4.1. Subsequently, target knowledge items T are sampled independently, ensuring T∩P′=∅, without assuming prior knowledge of future update requests.For each incoming knowledge item i∈T, dictionary coverage is evaluated to determine whether the existing bases sufficiently capture its gradient directions. Specifically, given the knowledge-induced gradient Gi=∇Wsi(W) from
\frac{\|(U^\star)^\top G_i V^\star\|_F} {\|G_i\|_F},$$ where $U^\star$ and $V^\star$ denote the current output- and input-side bases, respectively. If $\mathrm{cov}_i\geq 0.5$, the existing dictionary is reused. Otherwise, the corresponding bases are incrementally updated to accommodate newly observed knowledge directions. Specifically, the new token-level factors are extracted following `Eqs. (6)--(7)` to form $A'$ and $B'$. Following the incremental SVD framework in [6], the existing bases and singular values can be reused to construct the reduced matrices $\begin{bmatrix}\Sigma_A(V^\star)^\top\\ A'\end{bmatrix}$ and $\begin{bmatrix}\Sigma_B(U^\star)^\top\\ B'\end{bmatrix}$. Their leading right singular vectors are then retained as the updated bases $V^\star$ and $U^\star$, respectively. Notably, for $n$ incoming knowledge items and $k$ new factor rows per weight matrix, gradient extraction requires $\mathcal{O}(nC_{\mathrm{FB}})$. The compressed matrices have dimensions $(r_{\mathrm{in}}+k)\times d_{\mathrm{in}}$ and $(r_{\mathrm{out}}+k)\times d_{\mathrm{out}}$. Assuming $r_{\mathrm{in}}+k\ll d_{\mathrm{in}}$ and $r_{\mathrm{out}}+k\ll d_{\mathrm{out}}$, the total complexity is $\mathcal{O}\left(nC_{\mathrm{FB}}+\sum_W\left[d_{\mathrm{in}}(r_{\mathrm{in}}+k)^2+d_{\mathrm{out}}(r_{\mathrm{out}}+k)^2\right]\right)$. When $n\ll N$ and $k\ll K$, incremental updating significantly reduces reconstruction costs by operating on compressed matrices rather than the full accumulated knowledge pool. Then, updated bases are used to produce dictionary atoms following `Eq. (9)`, followed by task-specific atom filtering and coefficient optimization. **Table A. Unlearning performance on TOFU with fixed and dynamic dictionaries.** | Method | FQ ↑ | MU ↑ | | ------ | ---- | ---- | | KEY (Original $P$) | 0.69 | 0.66 | | KEY ($P'$ with fixed dictionary) | 0.46 | 0.53 | | KEY ($P'$ with dynamic dictionary) | 0.65 | 0.62 | **Table B. Editing performance on CounterFact with fixed and dynamic dictionaries.** | Method | ES ↑ | PS ↑ | NS ↑ | | ------ | ---- | ---- | ---- | | KEY (Original $P$) | 99.9 | 98.1 | 83.3 | | KEY ($P'$ with fixed dictionary) | 89.5 | 86.7 | 56.9 | | KEY ($P'$ with dynamic dictionary) | 99.8 | 97.2 | 81.5 |Eq. (5), the dictionary coverage is measured as