Upload folder using huggingface_hub

icml26/31f6f2e8-0fb2-46ff-ab65-f3408612f6e1/appendix_chunks.jsonl

@@ -0,0 +1,7 @@
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0062", "section": "A.1. Parameter-Efficient Fine-Tuning", "page_start": 11, "page_end": 11, "type": "Text", "text": "Parameter-efficient fine-tuning (PEFT) approaches adjust pretrained LLMs by incorporating lightweight modules, while keeping the base model frozen. Optimization is performed solely through updates to these lightweight components. Representative approaches include Adapters (Houlsby et al., 2019; Zaken et al., 2021) , Prompts (Li & Liang, 2021; Jia et al., 2022) , and Low-Rank Adaptation (LoRA) (Hu et al., 2022; Valipour et al., 2022) . Adapters are compact bottleneck modules located after transformer blocks; Prompts are learnable vectors prepended to the input sequence; and LoRA uses low-rank decomposition matrices to modify model weights during training. Recent progress explores applying PEFT to lifelong model editing, where the base model remains unchanged and new parameter modules are introduced to manage sequential edits over time.", "source": "marker_v2", "marker_block_id": "/page/10/Text/5"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0063", "section": "A.2. Routing Mechanisms in Modular Language Models", "page_start": 11, "page_end": 11, "type": "Text", "text": "When multiple modular components (e.g., LoRA modules) are present, a key challenge lies in routing input queries to the appropriate subset of modules during inference. Recent work explore various strategies to support dynamic module selection. MELO (Yu et al., 2024) introduces a clustering mechanism to assign edits to clusters based on input hidden states. During inference, the nearest cluster center is used to retrieve relevant LoRA modules. However, cluster centers are inevitably updated during the sequential editing, which may lead to semantic drift and incorrect matching. As shown in Fig .2, the number of incorrect matches increases significantly as the number of cluster centre updating increases, which leads to a deterioration in model performance. EL-DER (Li et al., 2025) employs MoE approach in which a learnable neural network scores a set of LoRA modules, activating the top-k modules. This strategy offers parameter efficiency but introduces soft entanglement in routing, and due to shared module usage, new edits may interfere with or overwrite previously edits. Furthermore, both strategies rely on auxiliary routing networks, adding architectural complexity and computational overhead. They also offer limited control over edit selection, making rollback and deletion of edits challenging.", "source": "marker_v2", "marker_block_id": "/page/10/Text/7"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0064", "section": "B. Additional Experiment Details", "page_start": 11, "page_end": 11, "type": "Text", "text": "Benchmarks We evaluate the performance of the model in three representative tasks: Document Classification, Question Answering(QA), and Hallucination Correction. Among them, the document classification task is performed based", "source": "marker_v2", "marker_block_id": "/page/10/Text/9"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0065", "section": "B. Additional Experiment Details", "page_start": 11, "page_end": 11, "type": "Text", "text": "on the SCOTUS dataset (Chalkidis et al., 2022) . SCOTUS is a subset of Fairlex, which collects documents from the U.S. Supreme Court and classifies them into 11 topics. In practice, the topics corresponding to the documents will change, so it is necessary to update the model knowledge to classify the documents into new topics. We divide SCO-TUS into an edit set and an upstream set, where models are edited on the edit set and the upstream set is not involved in editing, only testing. The question answering task is then conducted on the zsRE dataset (Levy et al., 2017) , and NQ datasets (Kwiatkowski et al., 2019) serves as upstream dataset. Specifically, we perform editing in zsRE, and the NQ dataset is again not involved in editing, only testing. For hallucination correction, we adopted the framework introduced by (Manakul et al., 2023) , aiming to address the tendency of GPT models to generate factually inconsistent outputs. Following GRACE (Hartvigsen et al., 2023) , we employ SelfCheckGPT (Manakul et al., 2023) for editing. SelfCheckGPT is a Wikipedia-style biography that is first generated using GPT3 on 256 topics, e.g., \"Bon Jovi\", and then checked to see which of the generated biographies are hallucinations. SelfCheckGPT contains 1392 hallucinatory sentences to be edited, which serves as the editing set, and 516 already correct sentences as the retention set, which is used to measure model's post-edit perplexity on the already correct sentences. WebText (Nakano et al., 2021) is used as the upstream set, remaining unedited and reserved for testing. In addition to this, in order to test the performance of the model approach with different sized models as well as datasets, we conducted hallucination correction experiments on Uniedit (Chen et al., 2025) and WikiBigEdit (Thede et al., 2025) to evaluate the generalization performance of the models. Uniedit is based on open domain knowledge covering information from 25 common domains in 5 major categories. The WikiBigEdit dataset is based on regular updates of the knowledge graph in Wikipedia and contains 5 months of extensive factual editing and improvement. The Uniedit and WikiBigEdit datasets already have edit set, retention set and upstream set that can be used directly for hallucination correction tasks.", "source": "marker_v2", "marker_block_id": "/page/10/Text/10"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0066", "section": "B. Additional Experiment Details", "page_start": 11, "page_end": 11, "type": "Text", "text": "Training details Following (Hartvigsen et al., 2023) , we employ BERT, T5-small, and GPT2-XL as the base models for the three tasks—document classification, question answering, and hallucination correction, respectively. Among these, BERT and T5-small are the official pre-trained models, while GPT2-XL is obtained from (Hartvigsen et al., 2023) which finetuned from the official release (Hartvigsen et al., 2023) . For training, we adopt Stochastic Gradient Descent (SGD) as the optimizer with a cosine decay learning rate schedule. The learning rates is 0.05, training epochs", "source": "marker_v2", "marker_block_id": "/page/10/Text/11"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0067", "section": "B. Additional Experiment Details", "page_start": 12, "page_end": 12, "type": "Text", "text": "is 40, and LoRA rank is 4. In the three tasks, we edit the models and the edited layers of the individual model edits are shown in the Tab .6.", "source": "marker_v2", "marker_block_id": "/page/11/Text/1"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0068", "section": "B. Additional Experiment Details", "page_start": 12, "page_end": 12, "type": "TableGroup", "text": "Table 6. Edited layers of different model Model Edit layer BERT bert.encoder.layer.9.output.dense bert.encoder.layer.10.output.dense bert.encoder.layer.11.output.dense T5-Small encoder.block.5.layer.1.DenseReluDense.wi 0 encoder.block.5.layer.1.DenseReluDense.wi 1 encoder.block.5.layer.1.DenseReluDense.wo encoder.block.6.layer.1.DenseReluDense.wi 0 encoder.block.6.layer.1.DenseReluDense.wi 1 encoder.block.6.layer.1.DenseReluDense.wo GPT2-XL transformer.h.36.mlp.c fc transformer.h.37.mlp.c fc", "source": "marker_v2", "marker_block_id": "/page/11/TableGroup/239"}

icml26/31f6f2e8-0fb2-46ff-ab65-f3408612f6e1/appendix_text_v3.txt

@@ -0,0 +1,20 @@
+[p. 11 | section: A.1. Parameter-Efficient Fine-Tuning | type: Text]
+Parameter-efficient fine-tuning (PEFT) approaches adjust pretrained LLMs by incorporating lightweight modules, while keeping the base model frozen. Optimization is performed solely through updates to these lightweight components. Representative approaches include Adapters (Houlsby et al., 2019; Zaken et al., 2021) , Prompts (Li & Liang, 2021; Jia et al., 2022) , and Low-Rank Adaptation (LoRA) (Hu et al., 2022; Valipour et al., 2022) . Adapters are compact bottleneck modules located after transformer blocks; Prompts are learnable vectors prepended to the input sequence; and LoRA uses low-rank decomposition matrices to modify model weights during training. Recent progress explores applying PEFT to lifelong model editing, where the base model remains unchanged and new parameter modules are introduced to manage sequential edits over time.
+
+[p. 11 | section: A.2. Routing Mechanisms in Modular Language Models | type: Text]
+When multiple modular components (e.g., LoRA modules) are present, a key challenge lies in routing input queries to the appropriate subset of modules during inference. Recent work explore various strategies to support dynamic module selection. MELO (Yu et al., 2024) introduces a clustering mechanism to assign edits to clusters based on input hidden states. During inference, the nearest cluster center is used to retrieve relevant LoRA modules. However, cluster centers are inevitably updated during the sequential editing, which may lead to semantic drift and incorrect matching. As shown in Fig .2, the number of incorrect matches increases significantly as the number of cluster centre updating increases, which leads to a deterioration in model performance. EL-DER (Li et al., 2025) employs MoE approach in which a learnable neural network scores a set of LoRA modules, activating the top-k modules. This strategy offers parameter efficiency but introduces soft entanglement in routing, and due to shared module usage, new edits may interfere with or overwrite previously edits. Furthermore, both strategies rely on auxiliary routing networks, adding architectural complexity and computational overhead. They also offer limited control over edit selection, making rollback and deletion of edits challenging.
+
+[p. 11 | section: B. Additional Experiment Details | type: Text]
+Benchmarks We evaluate the performance of the model in three representative tasks: Document Classification, Question Answering(QA), and Hallucination Correction. Among them, the document classification task is performed based
+
+[p. 11 | section: B. Additional Experiment Details | type: Text]
+on the SCOTUS dataset (Chalkidis et al., 2022) . SCOTUS is a subset of Fairlex, which collects documents from the U.S. Supreme Court and classifies them into 11 topics. In practice, the topics corresponding to the documents will change, so it is necessary to update the model knowledge to classify the documents into new topics. We divide SCO-TUS into an edit set and an upstream set, where models are edited on the edit set and the upstream set is not involved in editing, only testing. The question answering task is then conducted on the zsRE dataset (Levy et al., 2017) , and NQ datasets (Kwiatkowski et al., 2019) serves as upstream dataset. Specifically, we perform editing in zsRE, and the NQ dataset is again not involved in editing, only testing. For hallucination correction, we adopted the framework introduced by (Manakul et al., 2023) , aiming to address the tendency of GPT models to generate factually inconsistent outputs. Following GRACE (Hartvigsen et al., 2023) , we employ SelfCheckGPT (Manakul et al., 2023) for editing. SelfCheckGPT is a Wikipedia-style biography that is first generated using GPT3 on 256 topics, e.g., "Bon Jovi", and then checked to see which of the generated biographies are hallucinations. SelfCheckGPT contains 1392 hallucinatory sentences to be edited, which serves as the editing set, and 516 already correct sentences as the retention set, which is used to measure model's post-edit perplexity on the already correct sentences. WebText (Nakano et al., 2021) is used as the upstream set, remaining unedited and reserved for testing. In addition to this, in order to test the performance of the model approach with different sized models as well as datasets, we conducted hallucination correction experiments on Uniedit (Chen et al., 2025) and WikiBigEdit (Thede et al., 2025) to evaluate the generalization performance of the models. Uniedit is based on open domain knowledge covering information from 25 common domains in 5 major categories. The WikiBigEdit dataset is based on regular updates of the knowledge graph in Wikipedia and contains 5 months of extensive factual editing and improvement. The Uniedit and WikiBigEdit datasets already have edit set, retention set and upstream set that can be used directly for hallucination correction tasks.
+
+[p. 11 | section: B. Additional Experiment Details | type: Text]
+Training details Following (Hartvigsen et al., 2023) , we employ BERT, T5-small, and GPT2-XL as the base models for the three tasks—document classification, question answering, and hallucination correction, respectively. Among these, BERT and T5-small are the official pre-trained models, while GPT2-XL is obtained from (Hartvigsen et al., 2023) which finetuned from the official release (Hartvigsen et al., 2023) . For training, we adopt Stochastic Gradient Descent (SGD) as the optimizer with a cosine decay learning rate schedule. The learning rates is 0.05, training epochs
+
+[p. 12 | section: B. Additional Experiment Details | type: Text]
+is 40, and LoRA rank is 4. In the three tasks, we edit the models and the edited layers of the individual model edits are shown in the Tab .6.
+
+[p. 12 | section: B. Additional Experiment Details | type: TableGroup]
+Table 6. Edited layers of different model Model Edit layer BERT bert.encoder.layer.9.output.dense bert.encoder.layer.10.output.dense bert.encoder.layer.11.output.dense T5-Small encoder.block.5.layer.1.DenseReluDense.wi 0 encoder.block.5.layer.1.DenseReluDense.wi 1 encoder.block.5.layer.1.DenseReluDense.wo encoder.block.6.layer.1.DenseReluDense.wi 0 encoder.block.6.layer.1.DenseReluDense.wi 1 encoder.block.6.layer.1.DenseReluDense.wo GPT2-XL transformer.h.36.mlp.c fc transformer.h.37.mlp.c fc

icml26/31f6f2e8-0fb2-46ff-ab65-f3408612f6e1/assets.json

@@ -0,0 +1,57 @@
+[
+  {
+    "filename": "_page_0_Figure_9.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/31f6f2e8-0fb2-46ff-ab65-f3408612f6e1/marker_markdown/31f6f2e8-0fb2-46ff-ab65-f3408612f6e1/_page_0_Figure_9.jpeg",
+    "bytes": 17695,
+    "width": 670,
+    "height": 402,
+    "aspect_ratio": 1.6666666666666667,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_0_Figure_9.jpeg"
+  },
+  {
+    "filename": "_page_1_Figure_3.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/31f6f2e8-0fb2-46ff-ab65-f3408612f6e1/marker_markdown/31f6f2e8-0fb2-46ff-ab65-f3408612f6e1/_page_1_Figure_3.jpeg",
+    "bytes": 27018,
+    "width": 660,
+    "height": 379,
+    "aspect_ratio": 1.7414248021108178,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_1_Figure_3.jpeg"
+  },
+  {
+    "filename": "_page_2_Figure_1.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/31f6f2e8-0fb2-46ff-ab65-f3408612f6e1/marker_markdown/31f6f2e8-0fb2-46ff-ab65-f3408612f6e1/_page_2_Figure_1.jpeg",
+    "bytes": 105408,
+    "width": 1363,
+    "height": 649,
+    "aspect_ratio": 2.100154083204931,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_2_Figure_1.jpeg"
+  },
+  {
+    "filename": "_page_5_Figure_2.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/31f6f2e8-0fb2-46ff-ab65-f3408612f6e1/marker_markdown/31f6f2e8-0fb2-46ff-ab65-f3408612f6e1/_page_5_Figure_2.jpeg",
+    "bytes": 57599,
+    "width": 1263,
+    "height": 346,
+    "aspect_ratio": 3.6502890173410405,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_5_Figure_2.jpeg"
+  },
+  {
+    "filename": "_page_5_Figure_4.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/31f6f2e8-0fb2-46ff-ab65-f3408612f6e1/marker_markdown/31f6f2e8-0fb2-46ff-ab65-f3408612f6e1/_page_5_Figure_4.jpeg",
+    "bytes": 17086,
+    "width": 668,
+    "height": 520,
+    "aspect_ratio": 1.2846153846153847,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_5_Figure_4.jpeg"
+  }
+]

icml26/31f6f2e8-0fb2-46ff-ab65-f3408612f6e1/chunks_v3_anonymized.jsonl

@@ -0,0 +1,69 @@
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0000", "section": "Abstract", "page_start": 1, "page_end": 1, "type": "Text", "text": "The dynamic evolution of real-world necessitates model editing within Large Language Models. While existing methods explore modular isolation or parameter-efficient strategies, they still suffer from semantic drift or knowledge forgetting due to continual updating. To address these challenges, we propose SoLA, a Semantic routingbased LoRA framework for lifelong model editing. In SoLA, each edit is encapsulated as an independent LoRA module, which is frozen after training and mapped to input by semantic routing, allowing dynamic activation of LoRA modules via semantic matching. This mechanism avoids semantic drift caused by cluster updating and mitigates catastrophic forgetting from parameter sharing. More importantly, SoLA supports precise revocation of specific edits by removing key from semantic routing, which restores model's original behavior. To our knowledge, this reversible rollback editing capability is the first to be achieved in existing literature. Furthermore, SoLA integrates decision-making process into edited layer, eliminating the need for auxiliary routing networks and enabling end-to-end decision-making process. Extensive experiments demonstrate that SoLA effectively learns and retains edited knowledge, achieving accurate, efficient, and reversible lifelong model editing.", "source": "marker_v2", "marker_block_id": "/page/0/Text/4"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0001", "section": "1. Introduction", "page_start": 1, "page_end": 1, "type": "Text", "text": "Large language models (LLMs) have demonstrated remarkable capabilities in the field of natural language processing, attracting the attention of numerous researchers (Radford et al., 2019; Brown et al., 2020; Achiam et al., 2023; Tou vron et al., 2023) . However, LLMs also face severe challenges, including hallucinations (Huang et al., 2025) , bi-", "source": "marker_v2", "marker_block_id": "/page/0/Text/6"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0002", "section": "1. Introduction", "page_start": 1, "page_end": 1, "type": "FigureGroup", "text": "Figure 1. Comparison of trainable parameters and ERR accuracy between SoLA (Ours) and other methods. All experiments are conducted on Scotus datasets with the same backbone.", "source": "marker_v2", "marker_block_id": "/page/0/FigureGroup/467"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0003", "section": "1. Introduction", "page_start": 1, "page_end": 1, "type": "Text", "text": "ases (Hartvigsen et al., 2022) , and the generation of harmful content (De Cao et al., 2021) . Moreover, the dynamic nature of real-world knowledge requires the continuous updating of specific information in LLMs (Lazaridou et al., 2021) . Re-training these models from scratch is both expensive and time-consuming (Biderman et al., 2023) . In this scenario, the demand for model editing is increasing with aims to modify specific knowledge in the training model continuously without re-training the model or affecting its performance on unrelated inputs, which is known as lifelong model editing (Hartvigsen et al., 2022; Yao et al., 2023; Yu et al., 2024) .", "source": "marker_v2", "marker_block_id": "/page/0/Text/11"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0004", "section": "1. Introduction", "page_start": 1, "page_end": 1, "type": "Text", "text": "The lifelong model editing aims to continuously update the model to adapt to the constantly changing world. However, most of the existing model editing methods are designed for single, isolated editing (Mitchell et al., 2021; Meng et al., 2022a) . When applied to a continuous environment, these methods often lead to catastrophic forgetting of the previously learned knowledge in the LLMs, resulting in performance degradation and reduced model reliability (Yao et al., 2023; Gu et al., 2024) . To solve this problem, recent methods propose freezing the original model parameters and introducing lightweight trainable modules to inject new knowledge. These methods usually retain most of the model's knowledge by keeping the base parameters fixed and combining the learnable modules for specific task updates. For example, Melo (Yu et al., 2024) determines the clustering centres of editors by building a neuron index", "source": "marker_v2", "marker_block_id": "/page/0/Text/12"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0005", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "Text", "text": "and dynamically updates its semantic representation based on the distance from the clustering centres. ELDER (Li et al., 2025) adopts the Mixture-of-Experts (MoE) approach, assigning scores to each LoRA expert through a set of learnable LoRA allocation codes and selecting the top-k LoRA for calculation.", "source": "marker_v2", "marker_block_id": "/page/1/Text/2"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0006", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "FigureGroup", "text": "Figure 2. Number of mismatches in LoRA allocation with ERR/TRR accuracy in MELO. All experiments are conducted on Scotus datasets with the same backbone.", "source": "marker_v2", "marker_block_id": "/page/1/FigureGroup/464"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0007", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "Text", "text": "Although these methods show potential in improving parameter efficiency and reducing interference, they face inherent limitations when applied to lifelong model editing. Specifically, MELO (Yu et al., 2024) improves module separability through semantic clustering, but its clustering centers are updated during the editing process, which may lead to semantic drift and module matching errors. ELDER (Li et al., 2025) employ MoE and selects the top-k routes to activate each editing subset, although this strategy has high parameter efficiency, shared and continuously updated parameters can lead to catastrophic forgetting - new editing may overwrite or interfere with existing editing.", "source": "marker_v2", "marker_block_id": "/page/1/Text/5"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0008", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "Text", "text": "To address these limitations, we propose SoLA, a Semantic routing-based LoRA framework for reversible lifelong model editing. In SoLA, we allocate an independent LoRA module for each edit and establish semantic routing to record the mapping relationship between LoRA module and the input semantic representation. During editing, the LoRA module fully learns for the current task and then remains frozen to retain the learned knowledge. The corresponding key is also not updated. During inference, the semantic representation of the input is calculated, and the corresponding LoRA module is dynamically activated through semantic routing. Since the LoRA module and the semantic representation have not been updated, we fundamentally avoid catastrophic forgetting and semantic drift problems caused by continual updating. At the same time, each edit only requires training for the current LoRA module, significantly reducing the computational resource consumption. As shown in Fig .1, SoLA achieves optimal performance with only 0.08M additional parameters, sig-", "source": "marker_v2", "marker_block_id": "/page/1/Text/6"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0009", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "Text", "text": "nificantly outperforming previous methods, highlighting SoLA's exceptional parameter efficiency and editing accuracy.", "source": "marker_v2", "marker_block_id": "/page/1/Text/7"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0010", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "Text", "text": "More importantly, by removing the semantic key from the mapping memory table, we can precisely revoke specific edits, allowing the model to recover its original behavior without re-training. This means we can freely add and delete edits, and truly achieve controllable rollback and restoration of editing. To our knowledge, this is the first time that controllable rollback of editing has been achieved in existing literature. Moreover, existing works usually require introducing auxiliary routing network outside the edited layer to determine whether to enable the LoRA module. In this paper, we propose a master decision-making mechanism that aggregates the decision-making process into the edited layer, avoiding the reliance on auxiliary routing network and achieving an end-to-end decision-making process.", "source": "marker_v2", "marker_block_id": "/page/1/Text/8"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0011", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "Text", "text": "Extensive experiments validate the effectiveness of our approach. Our main contributions are summarized as follows:", "source": "marker_v2", "marker_block_id": "/page/1/Text/9"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0012", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "ListGroup", "text": "We propose SoLA, a novel framework for reversible lifelong model editing, which incorporates a controllable LoRA mechanism guided by semantic routing. After each edit, both the LoRA modules and the corresponding keys in the mapping memory are frozen, effectively mitigating catastrophic forgetting and semantic drift. Moreover, only the currently active LoRA modules are trained during editing, significantly reducing computational overhead.(Fig .1) . We employs semantic routing to establish a precise mapping between LoRA modules and semantic representations. By removing the associated key, any specific edit can be precisely revoked, enabling flexible addition and deletion of edits. we propose a master decision-making mechanism that aggregates the decision-making process into the edited layer, avoiding the reliance on an auxiliary routing network and achieving an end-to-end decision-making process.", "source": "marker_v2", "marker_block_id": "/page/1/ListGroup/465"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0013", "section": "2.1. Model Editing", "page_start": 2, "page_end": 2, "type": "Text", "text": "Model Editing aims to update specific knowledge within LLMs while preserving their previous knowledge. Current model editing methods can be broadly categorized into three paradigms: Meta-Learning methods, Locate-then-Edit methods, and Memory-Based methods. Meta-Learning Methods employ a hypernetwork to predict the necessary gradients for editing and apply these gradients to the base model to achieve the update (Mitchell et al., 2021) . However, this", "source": "marker_v2", "marker_block_id": "/page/1/Text/15"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0014", "section": "2.1. Model Editing", "page_start": 3, "page_end": 3, "type": "FigureGroup", "text": "Figure 3. Main framework of our method. a) indicates the edited layer in model, where we retain the transformer block of base model frozen and add trainable LoRA module to the edited layer of base model. b) shows the editing process, where every edit will be assigned LoRA module and the query of input will be mapped to assigned LoRA module. c) presents the matching process between query of input and LoRA module in reference. Colour green indicates trainable, while color grey is frozen.", "source": "marker_v2", "marker_block_id": "/page/2/FigureGroup/530"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0015", "section": "2.1. Model Editing", "page_start": 3, "page_end": 3, "type": "Text", "text": "approach typically requires additional data for training the hypernetwork. Locate-then-Edit methods first identify the neurons most critical for the current input via minor perturbations and then directly modify these neurons to update the model's knowledge (Meng et al., 2022a; b) . While effective, these methods are often cumbersome and exhibit limitations when handling a large volume of simultaneous knowledge updates. Memory-Based methods store the information associated with edits in an buffer, which acts as a patch model augmenting the original model (Mitchell et al., 2022) . A significant drawback is that each edit necessitates retraining, making it difficult to adapt to continuous editing scenarios, and the memory bank grows incrementally with each edit, leading to escalating storage overhead. Recent studies have also explored editing constraints to maintain model behavior stability. AlphaEdit (Fang et al., 2024) introduces a null-space constrained approach. It aims to apply edits more precisely by constraining changes to the null space of irrelevant knowledge.", "source": "marker_v2", "marker_block_id": "/page/2/Text/3"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0016", "section": "2.1. Model Editing", "page_start": 3, "page_end": 3, "type": "Text", "text": "However, these methods primarily address static editing, suitable for one-time modifications, and struggle to accommodate sequential editing demands over time. To address these limitations, GRACE (Hartvigsen et al., 2023) replace hidden states with vectors retrieved from a learned codebook. MELO (Yu et al., 2024) introduces a vector database and leverages a clustering mechanism to dynamically assign LoRA modules. ELDER (Li et al., 2025) adopts a MoE framework, where a learnable neural network weights shared LoRA modules. However, these approaches in-", "source": "marker_v2", "marker_block_id": "/page/2/Text/4"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0017", "section": "2.1. Model Editing", "page_start": 3, "page_end": 3, "type": "Text", "text": "evitably update the semantic representation in sequential edits, leading to semantic drift or knowledge forgetting.", "source": "marker_v2", "marker_block_id": "/page/2/Text/5"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0018", "section": "3.1. Preliminaries", "page_start": 3, "page_end": 3, "type": "Text", "text": "Lifelong Model Editing The purpose of lifelong model editing, as described by Grace (Hartvigsen et al., 2023) , is to make continual edits to the knowledge of the initial base model without degrading the performance of the base model and without negating previous edits. Consider an initial base model fbase, After n edits Dedit = {d1, ..., dn}, some of the model's knowledge is updated and model is transformed into fn, where d i = (x i , yi), For lifetime model editing, the edited model f n should be able to correctly output y within the edited inputs x ∈ Dedit, i.e., fn(xi) = y i . Furthermore, for inputs x ′ ∈/ Dedit that are semantically similar to the edited instances, such as rephrased sentences, the model f n is expected to generalize the updated knowledge appropriately, like fn(x ′ i ) = y i . Moreover, after editing, the model f n should also retain the previous knowledge of fbase, that is, for data samples x j ∈/ Dedit that have not been edited, there should be fn(x j ) = fbase(x j ) = y j .", "source": "marker_v2", "marker_block_id": "/page/2/Text/8"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0019", "section": "3.1. Preliminaries", "page_start": 3, "page_end": 3, "type": "Text", "text": "LoRA for Lifelong Model Editing Low-Rank Adaptation (LoRA) (Hu et al., 2022) is an efficient finetuning method for LLMs that optimizes model outputs by projecting features into a low-rank subspace. In the context of efficient LLM finetuning, LoRA modules are typically inserted into", "source": "marker_v2", "marker_block_id": "/page/2/Text/9"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0020", "section": "3.1. Preliminaries", "page_start": 4, "page_end": 4, "type": "Text", "text": "specific layers of the pre-trained model. During editing, the base LLM parameters W_0 \\in \\mathbb{R}^{d \\times k} remain frozen, while only the LoRA module \\Delta W \\in \\mathbb{R}^{d \\times k} is updated. The parameter update \\Delta W can be represented as a low-rank decomposition of the pretrained weight matrix. Specifically, \\Delta W is factorized into two smaller matrices A \\in \\mathbb{R}^{r \\times k} and B \\in \\mathbb{R}^{d \\times r} , such that \\Delta W = BA , where r \\ll \\min(d,k) and r is the LoRA rank. Given the original h = W_0 x , the forward propagation process is modified as:", "source": "marker_v2", "marker_block_id": "/page/3/Text/1"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0021", "section": "3.1. Preliminaries", "page_start": 4, "page_end": 4, "type": "Equation", "text": "\\boldsymbol{h} = W_0 \\boldsymbol{x} + \\Delta W \\boldsymbol{x} = W_0 \\boldsymbol{x} + B A \\boldsymbol{x} \\tag{1}", "source": "marker_v2", "marker_block_id": "/page/3/Equation/2"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0022", "section": "3.1. Preliminaries", "page_start": 4, "page_end": 4, "type": "Text", "text": "where A is initialized using a zero mean Gaussian distribution and B is initialized as a zero matrix. This approach significantly reduces the number of trainable parameters while maintaining the expressive power of the full-rank update.", "source": "marker_v2", "marker_block_id": "/page/3/Text/3"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0023", "section": "3.2. Cluster Updating in Lifelong Model Editing", "page_start": 4, "page_end": 4, "type": "Text", "text": "In the context of lifelong model editing, when multiple adapter modules are involved, the model needs to dynamically manage the association between edits and the corresponding adapter module. MELO addresses this by employing a clustering mechanism that assigns edits to clusters based on the hidden representations of the inputs, with cluster centres being continuously updated throughout editing. However, this continual updating of cluster centers alters their semantic representation, leading to semantic drift and incorrect module assignments. Moreover, repeated training of LoRA modules can result in forgetting previously acquired knowledge. Varying the cluster radius in MeLO results in a different number of generated cluster centres, thereby affecting the frequency of cluster centre updates. We record MeLO's performance under different numbers of cluster centre updates, as shown in Fig.2. The experiment is conducted on SCOTUS dataset. In Fig.2, the number of updates refers to the number of cluster centre updates during the editing process, and the number of mismatches indicates how many times, during inference, the retrieved LoRA module differs from the one assigned during editing. ERR and TRR represent the model's accuracy on the edited and unedited datasets, respectively, after all editing tasks are completed.", "source": "marker_v2", "marker_block_id": "/page/3/Text/5"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0024", "section": "3.2. Cluster Updating in Lifelong Model Editing", "page_start": 4, "page_end": 4, "type": "Text", "text": "As shown in the Fig.2, for the same datasets SCOTUS of editing, increasing the number of cluster centre updates leads to a rise in mismatches during inference, and a corresponding decline in model performance on both the edited and unedited datasets. This demonstrates that cluster centre updates can cause semantic drift, resulting in incorrect LoRA match, while repeated training of LoRA modules contributes to knowledge forgetting. Therefore, in SoLA, we freeze both the LoRA modules and their associated keys once the current task is completed. These components are", "source": "marker_v2", "marker_block_id": "/page/3/Text/6"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0025", "section": "3.2. Cluster Updating in Lifelong Model Editing", "page_start": 4, "page_end": 4, "type": "Text", "text": "no longer updated in subsequent edits, thereby maximizing knowledge retention across editing iterations.", "source": "marker_v2", "marker_block_id": "/page/3/Text/7"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0026", "section": "3.3. Semantic Routing-Based Controllable LoRA", "page_start": 4, "page_end": 4, "type": "Text", "text": "In the layer where the model needs to be edited, we keep the parameters of the original model frozen and insert the learnable LoRA module for editing, as shown in Fig.3 a). In SoLA, each editing operation is encapsulated within an independent LoRA module. During sequential editing, for a given editing task d_i = (\\boldsymbol{x}_i^m, \\boldsymbol{y}_i^m)_{m=1}^n , a dedicated LoRA module LoRA<sub>i</sub> is assigned. For each data instance \\boldsymbol{x}_i^m \\in d_i , the corresponding LoRA id i is recorded, and a semantic routing entry is established by associating LoRA module with a semantic key derived from the input representation, as shown in Fig.3 b). Following prior work(Hartvigsen et al., 2023), we use the hidden representation of the last token in the input sequence as the key e \\in \\mathbb{R}^d vector. During editing, the assigned LoRA module LoRA<sub>i</sub> will fully learn the task d_i , yielding the output:", "source": "marker_v2", "marker_block_id": "/page/3/Text/9"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0027", "section": "3.3. Semantic Routing-Based Controllable LoRA", "page_start": 4, "page_end": 4, "type": "Equation", "text": "h = h_0 + LoRA_i(x) (2)", "source": "marker_v2", "marker_block_id": "/page/3/Equation/10"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0028", "section": "3.3. Semantic Routing-Based Controllable LoRA", "page_start": 4, "page_end": 4, "type": "Text", "text": "where h_0 \\in \\mathbb{R}^{l \\times d} is the frozen base model representation, l is the sequence length, d is the hidden state dimension. At this stage, all other LoRA modules and the stored key vectors remain frozen. Upon completion of the editing process, the trained LoRA<sub>i</sub> module is frozen and will be stored with its associated key e_i as a fixed mapping for future reference. Neither the LoRA module nor the key will be updated in subsequent editing. Since only the current LoRA module is involved in training during each edit, SoLA significantly reduces the number of trainable parameters, as shown in Fig.1. Furthermore, freezing both the LoRA module and its corresponding key after editing prevents semantic drift and mitigates catastrophic forgetting, thereby maximizing knowledge retention. During inference, the hidden representation of the last token in the input sequence will serve as a query vector \\boldsymbol{q} \\in \\mathbb{R}^d from the input \\boldsymbol{x} and matches against the stored keys K \\in \\mathbb{R}^{N \\times d} , as shown in Fig.3 c). If a match is found, the corresponding LoRA module is retrieved from stored LoRA pool R \\in \\mathbb{R}^M and incorporated into the computation.", "source": "marker_v2", "marker_block_id": "/page/3/Text/11"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0029", "section": "3.4. Master Decision Mechanism", "page_start": 4, "page_end": 4, "type": "Text", "text": "During inference, existing approaches usually need to introduce an auxiliary routing network outside the target edited layer to decide whether to activate the LoRA module or not(Yu et al., 2024; Li et al., 2025), which hampers the end-to-end decision-making process. To address this limitation, we propose a master decision-making mechanism that does not require an auxiliary routing component. Specifically, we designate the first edited layer as the master decision layer where input features are dynamically evaluated to activate the relevant LoRA module without the need for an auxil-", "source": "marker_v2", "marker_block_id": "/page/3/Text/13"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0030", "section": "3.4. Master Decision Mechanism", "page_start": 5, "page_end": 5, "type": "TableGroup", "text": "Table 1. Comparison of SoLA to existing methods. All the results are obtained after sequential edits. ERR indicates Edit Reliability Rate, TRR presents Task Retention Rate, and ARR is Accuracy Attention Rate. The best performance is shown in bold. SCOTUS (BERT; Acc ↑) zsRE (T5; F1 ↑) Hallucination (GPT2-XL; PPL ↓) Method ERR TRR Avg. ERR TRR Avg. ERR TRR ARR EWC 0.51 0.82 0.66 0.67 0.50 0.58 1485.7 29.24 109.59 CMR 0.52 0.52 0.52 0.56 0.82 0.69 1449.3 28.14 107.76 CLEAR 0.67 0.83 0.75 0.27 0.99 0.63 2394.3 35.34 195.82 MEND 0.19 0.27 0.23 0.25 0.27 0.26 1369.8 1754.9 2902.5 SERAC 0.33 0.41 0.37 0.72 0.31 0.52 8183.7 133.3 10.04 ROME - - - - - - 30.28 103.82 14.02 GRACE 0.81 0.82 0.82 0.69 0.96 0.83 15.84 7.14 10.00 ELDER 0.89 0.86 0.88 0.72 0.97 0.84 16.12 5.87 8.42 MELO 0.96 0.92 0.94 0.72 0.98 0.85 17.45 1.04 2.66 SoLA 0.97 0.95 0.96 0.73 0.99 0.86 15.15 1.01 7.35", "source": "marker_v2", "marker_block_id": "/page/4/TableGroup/521"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0031", "section": "3.4. Master Decision Mechanism", "page_start": 5, "page_end": 5, "type": "Text", "text": "iary network. In this framework, the master decision layer H e computes the distance metric between the input feature embedding q and the stored key K to determine the activation of the module: d = He(argmini(dist(q, ki))), i = (1, 2, ..., N), dist(·) denotes the distance function and each key k i associated with the LoRA module. Then in the master layer He, we decide the edit layer behaviour:", "source": "marker_v2", "marker_block_id": "/page/4/Text/5"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0032", "section": "3.4. Master Decision Mechanism", "page_start": 5, "page_end": 5, "type": "Equation", "text": "H_e = \\begin{cases} H(W_0) & \\text{if } d \\ge \\alpha \\\\ H(W_0, W_{R_m}) & \\text{if } d < \\alpha \\end{cases} (3)", "source": "marker_v2", "marker_block_id": "/page/4/Equation/6"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0033", "section": "3.4. Master Decision Mechanism", "page_start": 5, "page_end": 5, "type": "Text", "text": "where α is the threshold, in this work we set it as 0.01, WR m represents the weight of m-th LoRA module associated with the key nearest to the query. This binary decision is propagated to the subsequent edited layers, ensuring consistent module activation throughout the network. By integrating the decision-making process to the first edited layer, our approach achieves complete end-to-end decision-making capability while maintaining architectural simplicity.", "source": "marker_v2", "marker_block_id": "/page/4/Text/7"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0034", "section": "4.1. Implementation Details", "page_start": 5, "page_end": 5, "type": "Text", "text": "Baselines We first compare several recent methods designed for lifelong model editing. Grace (Hartvigsen et al., 2023) replaces model outputs with key-value pairs and matches features to keys based on deferral radius. MELO (Yu et al., 2024) leverages LoRA for finetuning and dynamically retrieves relevant LoRA modules via neuron index. EL-DER (Li et al., 2025) employs MoE to dynamically combine multiple LoRA modules. Additionally, we evaluate other baseline methods. MEND (Mitchell et al., 2021) uses a hypernetwork trained on auxiliary data to predict the required gradients for model editing. SERAC (Mitchell et al., 2022)", "source": "marker_v2", "marker_block_id": "/page/4/Text/10"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0035", "section": "4.1. Implementation Details", "page_start": 5, "page_end": 5, "type": "Text", "text": "learns a scope classifier and a counterfactual model, coordinating between modules to determine editing behavior. ROME (Meng et al., 2022a) identifies the most influential weights in the model through perturbation analysis, localizes knowledge to specific layers of GPT, and modifies the corresponding weights. Since ROME is specifically designed for GPT, it is only evaluated on the hallucination correction task. EWC (Kirkpatrick et al., 2017) imposes constraints on the weights updating so that the model continuously learns new knowledge and retains previous knowledge. CMR (Lin et al., 2022) performs continuous finetuning of the model to continuously learn new knowledge. CLEAR (Rolnick et al., 2019) builds a memory buffer to restore old tasks information, and replays them in subsequent edits to retain previous knowledge.", "source": "marker_v2", "marker_block_id": "/page/4/Text/11"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0036", "section": "4.1. Implementation Details", "page_start": 5, "page_end": 5, "type": "Text", "text": "Evaluation Metric In this work, following (Hartvigsen et al., 2023) , we use three basic metrics to assess model performance: ES, ERR and TRR, which are applicable to all model experiments. Where Edit Success(ES) measures the ability of the model to edit the target sample. Edit Reliability Rate(ERR) assesses the validity of the model on the edited dataset, indicating the extent to which the required knowledge updates are successfully incorporated. Task Retention Rate(TRR) quantifies the model's performance on unedited datasets and reflects the model's ability to retain previous knowledge. Additionally, in the hallucination correction task, we also use the Accuracy Attention Rate (ARR) to assess the performance of the already accurate output. Accuracy is measured differently for different tasks. For the document classification task of SCOTUS, we use the Average Accuracy Rate (ACC); for the question answering task of zsRE, we measure it using the average F1, and for the hallucination correction task, we measure it using the standard average complexity (PPL) (Brown et al., 1992) .", "source": "marker_v2", "marker_block_id": "/page/4/Text/12"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0037", "section": "4.1. Implementation Details", "page_start": 6, "page_end": 6, "type": "FigureGroup", "text": "Figure 4. Figure(a)–(c) presents every edit accuracy on SCOTUS datasets with BERT. All experiment settings are the same as Tab .1.", "source": "marker_v2", "marker_block_id": "/page/5/FigureGroup/386"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0038", "section": "4.1. Implementation Details", "page_start": 6, "page_end": 6, "type": "FigureGroup", "text": "Figure 5. Visualization of zsRE dataset encoder output from T5 small with t-SNE and the dots with same color and shape are input and rephrase sentence. All experiment settings are the same as Tab .1.", "source": "marker_v2", "marker_block_id": "/page/5/FigureGroup/387"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0039", "section": "4.2. Main Results", "page_start": 6, "page_end": 6, "type": "Text", "text": "We evaluate the performance of the proposed SoLA method on three benchmark datasets and compare it with several state-of-the-art model editing approaches. The results are summarized in Tab .1. As observed, SoLA achieves excellent performance on most of the benchmark datasets, with SoLA outperforming the strongest baseline, MELO, by 3% on the SCOTUS dataset. These demonstrate SoLA's enhanced effectiveness in editing target knowledge while better preserving previously acquired knowledge. Furthermore, existing methods such as MEND and SERAC perform poorly under the sequential editing setting, suggesting that they are primarily designed for static editing scenarios and are not well-suited for sequential model editing tasks.", "source": "marker_v2", "marker_block_id": "/page/5/Text/7"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0040", "section": "4.2. Main Results", "page_start": 6, "page_end": 6, "type": "Text", "text": "We also report task-wise accuracy progression to analyze model performance throughout the sequential editing process. The experimental results are based on the SCOTUS dataset and the results are shown in Fig .4. As illustrated in Fig .4a, Fig .4b and Fig .4c, SoLA consistently maintains", "source": "marker_v2", "marker_block_id": "/page/5/Text/8"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0041", "section": "4.2. Main Results", "page_start": 6, "page_end": 6, "type": "Text", "text": "superior accuracy across tasks, without exhibiting substantial performance degradation or volatility. This indicates the robustness and stability of SoLA when applied to sequential knowledge editing.", "source": "marker_v2", "marker_block_id": "/page/5/Text/9"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0042", "section": "4.2. Main Results", "page_start": 6, "page_end": 6, "type": "Text", "text": "Method performance may vary across different model sizes and datasets. To further evaluate the robustness of our approach, we conduct hallucination correction experiments on the UniEdit (Chen et al., 2025) and WikiBigEdit (Thede et al., 2025) datasets using larger-scale backbone models. The results are presented in Tab .2. As shown in Tab .2, SoLA consistently achieves the best or near-best performance across most settings, demonstrating its robustness and effectiveness under diverse task configurations. During editing, larger backbone models generally exhibit more stable ERR values, suggesting that such models have learned stronger semantic representations during pretraining, leading to better generalization and editability. Furthermore, continuous learning related methods tend to overfit in edited data, thus showing better ARR in the retention set, but large TRR values in the upstream set, indicating severe forgetting of previous knowledge.", "source": "marker_v2", "marker_block_id": "/page/5/Text/10"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0043", "section": "4.3. Controllable Rollback Editing", "page_start": 6, "page_end": 6, "type": "Text", "text": "In SoLA, each edit instance is associated with a unique key stored in both the memory and the routing table. This design enables reversible editing, as the model can revert to its original behavior by removing the key corresponding to a specific edit. To demonstrate this property, we conduct an illustrative experiment on the zsRE dataset. The results are shown in Tab .3. In the Tab .3, each row corresponds to an input instance consisting of a \"Text\" (the question need to be edited) and its \"Labels\" (the correct answer). \"Del\" indicates whether the key corresponding to the edit is removed after editing. For each input, Predbase, Prededit and Preddel denote the prediction of base model, edited model and model after removing the corresponding key. When \"Del\" is false, the key is retained, serving as a baseline comparison. As shown in Tab .3, for each instance, the original prediction (Predbase) does not match the correct label, indi-", "source": "marker_v2", "marker_block_id": "/page/5/Text/12"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0044", "section": "4.3. Controllable Rollback Editing", "page_start": 7, "page_end": 7, "type": "TableGroup", "text": "Table 2. Model Performance Comparison on UniEdit and WikiBigEdit datasets in hallucination correction tasks. All the scores are PPL metric and the best result is in bolded. UniEdit (PPL ↓) WikiBigEdit (PPL ↓) Model Method ERR TRR ARR ERR TRR ARR EWC 2.43 1134 22.08 3.58 251117 4.93 CMR 2.79 2239 25.24 2.86 169452 4.49 LLaMA-3-8B CLEAR 1.04 982 29.79 1.03 74384 1.52 SERAC 37.10 246 93 60.17 215 59 GRACE 1.00 244 89 1.00 216 58 ELDER 1.00 173 74 1.00 178 48 MELO 1.00 153 68 1.00 165 47 Ours 1.00 144 66 1.00 162 44 EWC 2.38 4451 42.13 2.86 47846 3.89 CMR 2.34 7211 36.19 3.60 34623 5.51 DeepSeek-R1-8B CLEAR 1.06 3365 44.86 1.02 28110 1.48 SERAC 46.76 451 834 82.72 493 831 GRACE 1.00 407 803 1.00 487 793 ELDER 1.03 241 482 1.03 304 492 MELO 1.09 204 407 1.05 263 445 Ours 1.07 188 374 1.00 248 412 EWC 4.62 7648 48.62 5.61 61834 14.65 CMR 4.12 11583 40.56 6.28 48629 16.89 Qwen2-7B CLEAR 1.03 5021 54.98 1.05 35396 4.26 SERAC 58.51 329 1261 62.12 304 523 GRACE 2.05 316 1065 16.88 299 518 ELDER 1.00 201 266 1.00 225 197 MELO 1.00 170 201 1.00 203 143 Ours 1.00 156 110 1.00 199 109", "source": "marker_v2", "marker_block_id": "/page/6/TableGroup/486"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0045", "section": "4.3. Controllable Rollback Editing", "page_start": 7, "page_end": 7, "type": "Text", "text": "cating the necessity for editing. After editing, the model correctly produces Prededit consistent with the ground-truth label, demonstrating that SoLA successfully updates the model's knowledge. When the associated key is deleted, the model reverts to its original prediction (Prededit = Predbase), verifying that our method can effectively roll back specific edits. More critically, for edits where the key is not deleted, the model continues to output Prededit, showing that the rollback operation does not interfere with other edits. This confirms that SoLA supports fine-grained, selective undo of knowledge edits without disrupting unrelated modifications.", "source": "marker_v2", "marker_block_id": "/page/6/Text/3"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0046", "section": "4.4. Ablation Study", "page_start": 7, "page_end": 7, "type": "Text", "text": "Effect of Edit Layer Location In semantic representation learning, different layers of a model focus on different aspects of semantic information. Therefore, the position of the edited layer can significantly influence the effectiveness of model editing. To investigate this, we evaluate the performance of model when edits are applied at different layers, while keeping all other settings fixed. The experiments are conducted on the SCOTUS dataset with a Nvidia A100 40G GPU, and the results are shown in Tab .4. In the", "source": "marker_v2", "marker_block_id": "/page/6/Text/5"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0047", "section": "4.4. Ablation Study", "page_start": 7, "page_end": 7, "type": "Text", "text": "Tab .4, \"Layer\" denotes the layer range used for editing. For example, \"0–2\" indicates that edits are applied to layers 0, 1, and 2 of the BERT model. As shown in the table, editing at shallower layers results in suboptimal performance. This observation aligns with prior findings (Geva et al., 2020) , which suggest that shallow layers primarily capture the shallow sentence patterns, whereas deeper layers encode richer semantic features, thereby enabling more effective knowledge editing. Moreover, editing at shallow layers significantly increases the time used for editing, which means that editing for semantics is more difficult in shallow layers. Furthermore, we observe that varying the editing layer has negligible impact on performance over the unedited portion of the dataset, indicating that SoLA is capable of preserving previously learned knowledge regardless of the editing depth, which further supports the robustness of SoLA.", "source": "marker_v2", "marker_block_id": "/page/6/Text/6"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0048", "section": "4.4. Ablation Study", "page_start": 7, "page_end": 7, "type": "Text", "text": "Effect of LoRA Rank The LoRA module projects semantic representations into a low-rank subspace for parameterefficient learning. Consequently, the choice of rank will effect the module's performance. To investigate this, we analyze the impact of different LoRA rank values on model performance, keeping all other settings fixed. Experiments are conducted on the SCOTUS dataset using a NVIDIA", "source": "marker_v2", "marker_block_id": "/page/6/Text/7"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0049", "section": "4.4. Ablation Study", "page_start": 8, "page_end": 8, "type": "TableGroup", "text": "Table 3. Controllable edit in SoLA on zsRE datasets. \"Text\" is the question need to be edited, \"Labels\" is the true answer, \"Del\" indicates whther delete associated key and Predbase, Prededit and Preddel is the prediction of base model, edited model and delted model. Text Labels Del Predbase Prededit Preddel The date of Tsetserleg earthquake in 1905? 9 July 1905 True 20 February 1905 9 July 1905 20 February 1905 What constellation has Nu Cancri? Cancer True Hydra Cancer Hydra On what date did the Bahrain Grand Prix 2015 occur? 19 April 2015 True 6 April 2017 19 April 2015 6 April 2017 What position does Charles-Joseph Coursol have? Mayor of Montreal True Positioning Mayor of Montreal Positioning Who acted in Blow Dry? Alan Rickman False Jim Carrey Alan Rickman Alan Rickman Table 4. Ablation studies on location of edited layers. Here, \"Layer\" indicates the edited layer location. For example, \"0-2\" presents edited layer is layers 0, 1 and 2.", "source": "marker_v2", "marker_block_id": "/page/7/TableGroup/457"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0050", "section": "4.4. Ablation Study", "page_start": 8, "page_end": 8, "type": "TableGroup", "text": "Layer ES ERR TRR Edit Time(min) 0-2 0.77 0.61 0.96 9.21 3-5 1.00 0.68 0.96 8.06 6-8 0.81 0.70 0.97 7.03 9-11 0.94 0.95 0.97 5.97 Table 5. Ablation studies on LoRA rank. Here, \"LoRA Rank\" indicates the rank value of every LoRA in model.", "source": "marker_v2", "marker_block_id": "/page/7/TableGroup/458"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0051", "section": "4.4. Ablation Study", "page_start": 8, "page_end": 8, "type": "Table", "text": "LoRA Rank ES ERR TRR Edit Time(min) 1 0.90 0.84 0.96 5.88 2 0.94 0.91 0.97 5.90 3 1.00 0.91 0.97 5.86 4 1.00 0.95 0.97 5.97 5 1.00 0.92 0.96 5.86 10 1.00 0.71 0.96 5.85", "source": "marker_v2", "marker_block_id": "/page/7/Table/17"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0052", "section": "4.4. Ablation Study", "page_start": 8, "page_end": 8, "type": "Text", "text": "A100 40GB GPU, and the results are presented in Tab .5. As shown in the Tab .5, simply increasing the LoRA rank does not lead to improved performance and may even degrade model performance. This indicates that expanding the learnable parameter space does not necessarily enhance the model's capacity for effective knowledge editing. On the contrary, excessive rank values may even result in performance degradation due to overfitting, suggesting the importance of carefully selecting an appropriate rank to balance capacity and generalization. Based on the experiment results, we set 4 as LoRA rank for all the experiments in this paper.", "source": "marker_v2", "marker_block_id": "/page/7/Text/18"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0053", "section": "4.5. Model Visualization", "page_start": 8, "page_end": 8, "type": "Text", "text": "To gain deeper insight into the model's behavior under sequential editing, we conducted a t-SNE visualization (Maaten & Hinton, 2008) of the learned feature representations. This experiment is performed on the zsRE dataset using the T5-small model to assess how semantic relationships are preserved after sequential edits. Upon completing", "source": "marker_v2", "marker_block_id": "/page/7/Text/20"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0054", "section": "4.5. Model Visualization", "page_start": 8, "page_end": 8, "type": "Text", "text": "the entire sequence of editing tasks for question answering, we selected five distinct input instances, each accompanied by semantically equivalent rephrased sentences, forming five evaluation groups. The encoder output features are subsequently extracted from the edited model for dimensionality reduction and visualization, and the resulting plot is represented in Fig .5. As shown in Fig .5, there is a clear clustering pattern that for each group, the original input and its rephrased variant are mapped to proximate locations in the latent space. This demonstrates that SoLA effectively captures and preserves the semantic similarity between related inputs throughout the editing process. Furthermore, distinct groups maintain separable clusters, indicating that the model retains the ability to discriminate between different edit concepts after sequential editing. This visualization shows that SoLA's semantic routing mechanism successfully supports edit generalization based on semantic similarity while upholding the integrity of previously learned knowledge.", "source": "marker_v2", "marker_block_id": "/page/7/Text/21"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0055", "section": "5. Conclusion", "page_start": 8, "page_end": 8, "type": "Text", "text": "In this paper, we propose SoLA, a Semantic routing-based LoRA framework for reversible lifelong model editing. In SoLA, each edit is encapsulated as an independent LoRA module which is frozen after training and a semantic routing record is established to map LoRA module to the input semantic representation, allowing dynamic activation of LoRA modules via semantic matching. This design not only avoids semantic drift caused by clustering updating but also mitigates catastrophic forgetting from parameter sharing. More importantly, SoLA supports precisely revoking edit by removing the corresponding key from the routing table, enabling reversible model editing. This allows for flexible addition and deletion of edits, offering fine-grained control over model behavior. To the best of our knowledge, this is the first work to achieve reversible model editing in the literature. Furthermore, we introduce a master decisionmaking mechanism by integrating decision-making into the edited layer, enabling end-to-end decision-making process. By building a strict mapping between edits and LoRA modules, SoLA achieves efficient and reversible lifelong model editing, providing a novel perspective for future research.", "source": "marker_v2", "marker_block_id": "/page/7/Text/23"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0056", "section": "Impact Statement", "page_start": 9, "page_end": 9, "type": "Text", "text": "This work advances the field of Machine Learning by introducing SoLA, a framework for reversible and controllable lifelong model editing in large language models. Our method enhances the safety and reliability of AI systems by enabling precise rollback of model edits, which helps mitigate risks associated with erroneous or harmful knowledge updates. By effectively preventing semantic drift and catastrophic forgetting, SoLA promotes the development of more robust and trustworthy models capable of adapting to evolving information. The significant reduction in computational overhead also aligns with the goal of sustainable AI development.", "source": "marker_v2", "marker_block_id": "/page/8/Text/2"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0057", "section": "References", "page_start": 9, "page_end": 9, "type": "ListGroup", "text": "Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774 , 2023. Biderman, S., Schoelkopf, H., Anthony, Q. G., Bradley, H., O'Brien, K., Hallahan, E., Khan, M. A., Purohit, S., Prashanth, U. S., Raff, E., et al. Pythia: A suite for analyzing large language models across training and scaling. In International Conference on Machine Learning , pp. 2397–2430. PMLR, 2023. Brown, P. F., Della Pietra, S. A., Della Pietra, V. J., Lai, J. C., and Mercer, R. L. An estimate of an upper bound for the entropy of english. Computational Linguistics , 18 (1):31–40, 1992. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al. Language models are few-shot learners. Advances in neural information processing systems , 33: 1877–1901, 2020. Chalkidis, I., Pasini, T., Zhang, S., Tomada, L., Schwemer, S. F., and Søgaard, A. Fairlex: A multilingual benchmark for evaluating fairness in legal text processing. arXiv preprint arXiv:2203.07228 , 2022. Chen, Q., Wang, D., Zhang, T., Yan, Z., You, C., Wang, C., and He, X. Uniedit: A unified knowledge editing benchmark for large language models. arXiv preprint arXiv:2505.12345 , 2025. De Cao, N., Aziz, W., and Titov, I. Editing factual knowledge in language models. arXiv preprint arXiv:2104.08164 , 2021. Fang, J., Jiang, H., Wang, K., Ma, Y., Jie, S., Wang, X., He, X., and Chua, T.-S. Alphaedit: Null-space constrained knowledge editing for language models. arXiv preprint arXiv:2410.02355 , 2024.", "source": "marker_v2", "marker_block_id": "/page/8/ListGroup/516"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0058", "section": "References", "page_start": 9, "page_end": 9, "type": "ListGroup", "text": "Geva, M., Schuster, R., Berant, J., and Levy, O. Transformer feed-forward layers are key-value memories. arXiv preprint arXiv:2012.14913 , 2020. Gu, J.-C., Xu, H.-X., Ma, J.-Y., Lu, P., Ling, Z.-H., Chang, K.-W., and Peng, N. Model editing can hurt general abilities of large language models. CoRR , 2024. Hartvigsen, T., Gabriel, S., Palangi, H., Sap, M., Ray, D., and Kamar, E. Toxigen: A large-scale machine-generated dataset for adversarial and implicit hate speech detection. arXiv preprint arXiv:2203.09509 , 2022. Hartvigsen, T., Sankaranarayanan, S., Palangi, H., Kim, Y., and Ghassemi, M. Aging with grace: Lifelong model editing with discrete key-value adaptors. Advances in Neural Information Processing Systems , 36:47934–47959, 2023. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., and Gelly, S. Parameter-efficient transfer learning for nlp. In International conference on machine learning , pp. 2790– 2799. PMLR, 2019. Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al. Lora: Low-rank adaptation of large language models. ICLR , 1(2):3, 2022. Huang, L., Yu, W., Ma, W., Zhong, W., Feng, Z., Wang, H., Chen, Q., Peng, W., Feng, X., Qin, B., et al. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Information Systems , 43(2):1–55, 2025. Jia, M., Tang, L., Chen, B.-C., Cardie, C., Belongie, S., Hariharan, B., and Lim, S.-N. Visual prompt tuning. In European conference on computer vision , pp. 709–727. Springer, 2022. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A. A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences , 114(13):3521–3526, 2017. Kwiatkowski, T., Palomaki, J., Redfield, O., Collins, M., Parikh, A., Alberti, C., Epstein, D., Polosukhin, I., Devlin, J., Lee, K., et al. Natural questions: a benchmark for question answering research. Transactions of the Association for Computational Linguistics , 7:453–466, 2019. Lazaridou, A., Kuncoro, A., Gribovskaya, E., Agrawal, D., Liska, A., Terzi, T., Gimenez, M., de Masson d'Autume, C., Kocisky, T., Ruder, S., et al. Mind the gap: Assessing temporal generalization in neural language models. Advances in Neural Information Processing Systems , 34: 29348–29363, 2021.", "source": "marker_v2", "marker_block_id": "/page/8/ListGroup/517"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0059", "section": "References", "page_start": 10, "page_end": 10, "type": "Text", "text": "495 496 497 Levy, O., Seo, M., Choi, E., and Zettlemoyer, L. Zero-shot relation extraction via reading comprehension. arXiv preprint arXiv:1706.04115 , 2017.", "source": "marker_v2", "marker_block_id": "/page/9/Text/408"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0060", "section": "References", "page_start": 10, "page_end": 10, "type": "ListGroup", "text": "Li, J., Wang, Q., Wang, Z., Zhang, Y., and Mao, Z. Elder: Enhancing lifelong model editing with mixture-of-lora. In Proceedings of the AAAI Conference on Artificial In telligence , pp. 24440–24448, 2025. Li, X. L. and Liang, P. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190 , 2021. Lin, B. Y., Wang, S., Lin, X. V., Jia, R., Xiao, L., Ren, X., and Yih, W.-t. On continual model refinement in out-of-distribution data streams. arXiv preprint arXiv:2205.02014 , 2022. Maaten, L. v. d. and Hinton, G. Visualizing data using t-sne. Journal of machine learning research , 9(Nov): 2579–2605, 2008. Manakul, P., Liusie, A., and Gales, M. J. Selfcheckgpt: Zeroresource black-box hallucination detection for generative large language models. arXiv preprint arXiv:2303.08896 , 2023. Meng, K., Bau, D., Andonian, A., and Belinkov, Y. Locating and editing factual associations in gpt. Advances in neural information processing systems , 35:17359–17372, 2022a. Meng, K., Sharma, A. S., Andonian, A., Belinkov, Y., and Bau, D. Mass-editing memory in a transformer. arXiv preprint arXiv:2210.07229 , 2022b. Mitchell, E., Lin, C., Bosselut, A., Finn, C., and Manning, C. D. Fast model editing at scale. arXiv preprint arXiv:2110.11309 , 2021. Mitchell, E., Lin, C., Bosselut, A., Manning, C. D., and Finn, C. Memory-based model editing at scale. In Inter national Conference on Machine Learning , pp. 15817– 15831. PMLR, 2022. Nakano, R., Hilton, J., Balaji, S., Wu, J., Ouyang, L., Kim, C., Hesse, C., Jain, S., Kosaraju, V., Saunders, W., et al. Webgpt: Browser-assisted question-answering with human feedback. arXiv preprint arXiv:2112.09332 , 2021. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al. Language models are unsupervised multitask learners. OpenAI blog , 1(8):9, 2019. Rolnick, D., Ahuja, A., Schwarz, J., Lillicrap, T., and Wayne, G. Experience replay for continual learning. Advances in neural information processing systems , 32, 2019.", "source": "marker_v2", "marker_block_id": "/page/9/ListGroup/406"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0061", "section": "References", "page_start": 10, "page_end": 10, "type": "ListGroup", "text": "Thede, L., Roth, K., Bethge, M., Akata, Z., and Hartvigsen, T. Understanding the limits of lifelong knowledge editing in llms. arXiv preprint arXiv:2503.05683 , 2025. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al. Llama 2: Open foundation and finetuned chat models. arXiv preprint arXiv:2307.09288 , 2023. Valipour, M., Rezagholizadeh, M., Kobyzev, I., and Ghodsi, A. Dylora: Parameter efficient tuning of pre-trained models using dynamic search-free low-rank adaptation. arXiv preprint arXiv:2210.07558 , 2022. Yao, Y., Wang, P., Tian, B., Cheng, S., Li, Z., Deng, S., Chen, H., and Zhang, N. Editing large language models: Problems, methods, and opportunities. arXiv preprint arXiv:2305.13172 , 2023. Yu, L., Chen, Q., Zhou, J., and He, L. Melo: Enhancing model editing with neuron-indexed dynamic lora. In Proceedings of the AAAI Conference on Artificial Intelli gence , pp. 19449–19457, 2024. Zaken, E. B., Ravfogel, S., and Goldberg, Y. Bitfit: Simple parameter-efficient fine-tuning for transformerbased masked language-models. arXiv preprint arXiv:2106.10199 , 2021.", "source": "marker_v2", "marker_block_id": "/page/9/ListGroup/407"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0062", "section": "A.1. Parameter-Efficient Fine-Tuning", "page_start": 11, "page_end": 11, "type": "Text", "text": "Parameter-efficient fine-tuning (PEFT) approaches adjust pretrained LLMs by incorporating lightweight modules, while keeping the base model frozen. Optimization is performed solely through updates to these lightweight components. Representative approaches include Adapters (Houlsby et al., 2019; Zaken et al., 2021) , Prompts (Li & Liang, 2021; Jia et al., 2022) , and Low-Rank Adaptation (LoRA) (Hu et al., 2022; Valipour et al., 2022) . Adapters are compact bottleneck modules located after transformer blocks; Prompts are learnable vectors prepended to the input sequence; and LoRA uses low-rank decomposition matrices to modify model weights during training. Recent progress explores applying PEFT to lifelong model editing, where the base model remains unchanged and new parameter modules are introduced to manage sequential edits over time.", "source": "marker_v2", "marker_block_id": "/page/10/Text/5"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0063", "section": "A.2. Routing Mechanisms in Modular Language Models", "page_start": 11, "page_end": 11, "type": "Text", "text": "When multiple modular components (e.g., LoRA modules) are present, a key challenge lies in routing input queries to the appropriate subset of modules during inference. Recent work explore various strategies to support dynamic module selection. MELO (Yu et al., 2024) introduces a clustering mechanism to assign edits to clusters based on input hidden states. During inference, the nearest cluster center is used to retrieve relevant LoRA modules. However, cluster centers are inevitably updated during the sequential editing, which may lead to semantic drift and incorrect matching. As shown in Fig .2, the number of incorrect matches increases significantly as the number of cluster centre updating increases, which leads to a deterioration in model performance. EL-DER (Li et al., 2025) employs MoE approach in which a learnable neural network scores a set of LoRA modules, activating the top-k modules. This strategy offers parameter efficiency but introduces soft entanglement in routing, and due to shared module usage, new edits may interfere with or overwrite previously edits. Furthermore, both strategies rely on auxiliary routing networks, adding architectural complexity and computational overhead. They also offer limited control over edit selection, making rollback and deletion of edits challenging.", "source": "marker_v2", "marker_block_id": "/page/10/Text/7"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0064", "section": "B. Additional Experiment Details", "page_start": 11, "page_end": 11, "type": "Text", "text": "Benchmarks We evaluate the performance of the model in three representative tasks: Document Classification, Question Answering(QA), and Hallucination Correction. Among them, the document classification task is performed based", "source": "marker_v2", "marker_block_id": "/page/10/Text/9"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0065", "section": "B. Additional Experiment Details", "page_start": 11, "page_end": 11, "type": "Text", "text": "on the SCOTUS dataset (Chalkidis et al., 2022) . SCOTUS is a subset of Fairlex, which collects documents from the U.S. Supreme Court and classifies them into 11 topics. In practice, the topics corresponding to the documents will change, so it is necessary to update the model knowledge to classify the documents into new topics. We divide SCO-TUS into an edit set and an upstream set, where models are edited on the edit set and the upstream set is not involved in editing, only testing. The question answering task is then conducted on the zsRE dataset (Levy et al., 2017) , and NQ datasets (Kwiatkowski et al., 2019) serves as upstream dataset. Specifically, we perform editing in zsRE, and the NQ dataset is again not involved in editing, only testing. For hallucination correction, we adopted the framework introduced by (Manakul et al., 2023) , aiming to address the tendency of GPT models to generate factually inconsistent outputs. Following GRACE (Hartvigsen et al., 2023) , we employ SelfCheckGPT (Manakul et al., 2023) for editing. SelfCheckGPT is a Wikipedia-style biography that is first generated using GPT3 on 256 topics, e.g., \"Bon Jovi\", and then checked to see which of the generated biographies are hallucinations. SelfCheckGPT contains 1392 hallucinatory sentences to be edited, which serves as the editing set, and 516 already correct sentences as the retention set, which is used to measure model's post-edit perplexity on the already correct sentences. WebText (Nakano et al., 2021) is used as the upstream set, remaining unedited and reserved for testing. In addition to this, in order to test the performance of the model approach with different sized models as well as datasets, we conducted hallucination correction experiments on Uniedit (Chen et al., 2025) and WikiBigEdit (Thede et al., 2025) to evaluate the generalization performance of the models. Uniedit is based on open domain knowledge covering information from 25 common domains in 5 major categories. The WikiBigEdit dataset is based on regular updates of the knowledge graph in Wikipedia and contains 5 months of extensive factual editing and improvement. The Uniedit and WikiBigEdit datasets already have edit set, retention set and upstream set that can be used directly for hallucination correction tasks.", "source": "marker_v2", "marker_block_id": "/page/10/Text/10"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0066", "section": "B. Additional Experiment Details", "page_start": 11, "page_end": 11, "type": "Text", "text": "Training details Following (Hartvigsen et al., 2023) , we employ BERT, T5-small, and GPT2-XL as the base models for the three tasks—document classification, question answering, and hallucination correction, respectively. Among these, BERT and T5-small are the official pre-trained models, while GPT2-XL is obtained from (Hartvigsen et al., 2023) which finetuned from the official release (Hartvigsen et al., 2023) . For training, we adopt Stochastic Gradient Descent (SGD) as the optimizer with a cosine decay learning rate schedule. The learning rates is 0.05, training epochs", "source": "marker_v2", "marker_block_id": "/page/10/Text/11"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0067", "section": "B. Additional Experiment Details", "page_start": 12, "page_end": 12, "type": "Text", "text": "is 40, and LoRA rank is 4. In the three tasks, we edit the models and the edited layers of the individual model edits are shown in the Tab .6.", "source": "marker_v2", "marker_block_id": "/page/11/Text/1"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0068", "section": "B. Additional Experiment Details", "page_start": 12, "page_end": 12, "type": "TableGroup", "text": "Table 6. Edited layers of different model Model Edit layer BERT bert.encoder.layer.9.output.dense bert.encoder.layer.10.output.dense bert.encoder.layer.11.output.dense T5-Small encoder.block.5.layer.1.DenseReluDense.wi 0 encoder.block.5.layer.1.DenseReluDense.wi 1 encoder.block.5.layer.1.DenseReluDense.wo encoder.block.6.layer.1.DenseReluDense.wi 0 encoder.block.6.layer.1.DenseReluDense.wi 1 encoder.block.6.layer.1.DenseReluDense.wo GPT2-XL transformer.h.36.mlp.c fc transformer.h.37.mlp.c fc", "source": "marker_v2", "marker_block_id": "/page/11/TableGroup/239"}

icml26/31f6f2e8-0fb2-46ff-ab65-f3408612f6e1/dataset_meta.json

@@ -0,0 +1,60 @@
+{
+  "paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1",
+  "pipeline": "Paper2Markdown-V3",
+  "ok": true,
+  "page_count": 12,
+  "chunk_count": 69,
+  "main_body_chunk_count": 57,
+  "appendix_chunk_count": 7,
+  "reference_chunk_count": 5,
+  "model_text_chars": 37283,
+  "raw_markdown_chars": 58572,
+  "sanitized_chars": 53252,
+  "page_provenance": {
+    "min_page": 1,
+    "max_page": 12,
+    "invalid_count": 0
+  },
+  "marker_block_type_counts": {
+    "Caption": 5,
+    "Equation": 3,
+    "FigureGroup": 5,
+    "Footnote": 1,
+    "ListGroup": 5,
+    "PageFooter": 12,
+    "PageHeader": 15,
+    "SectionHeader": 26,
+    "Table": 1,
+    "TableGroup": 5,
+    "Text": 171
+  },
+  "asset_count_raw": 5,
+  "asset_count_model_kept": 5,
+  "asset_count_rejected": 0,
+  "asset_reject_reasons": {
+    "kept": 5
+  },
+  "artifact_leak_audit": {
+    "ok": true,
+    "hits": {
+      "Anonymous Authors": [],
+      "ACKNOWLEDGMENT": [],
+      "OpenReview": [],
+      "\"accept_label\"": [],
+      "\"decision\"": [],
+      "\"decision_tier\"": [],
+      "\"source_status\"": [],
+      "Meta-review": [],
+      "Official Review": [],
+      "official_reviews": [],
+      "meta_reviews": [],
+      "suggested_verdict_score": []
+    },
+    "artifact_count": 2
+  },
+  "default_model_input": "model_text_v3.txt",
+  "appendix_input": "appendix_text_v3.txt",
+  "reference_input": "reference_text_v3.txt",
+  "source": "koala_icml26_due_queue",
+  "run_name": "icml26_20260429_1952_duequeue"
+}

icml26/31f6f2e8-0fb2-46ff-ab65-f3408612f6e1/main_body_chunks.jsonl

@@ -0,0 +1,57 @@
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0000", "section": "Abstract", "page_start": 1, "page_end": 1, "type": "Text", "text": "The dynamic evolution of real-world necessitates model editing within Large Language Models. While existing methods explore modular isolation or parameter-efficient strategies, they still suffer from semantic drift or knowledge forgetting due to continual updating. To address these challenges, we propose SoLA, a Semantic routingbased LoRA framework for lifelong model editing. In SoLA, each edit is encapsulated as an independent LoRA module, which is frozen after training and mapped to input by semantic routing, allowing dynamic activation of LoRA modules via semantic matching. This mechanism avoids semantic drift caused by cluster updating and mitigates catastrophic forgetting from parameter sharing. More importantly, SoLA supports precise revocation of specific edits by removing key from semantic routing, which restores model's original behavior. To our knowledge, this reversible rollback editing capability is the first to be achieved in existing literature. Furthermore, SoLA integrates decision-making process into edited layer, eliminating the need for auxiliary routing networks and enabling end-to-end decision-making process. Extensive experiments demonstrate that SoLA effectively learns and retains edited knowledge, achieving accurate, efficient, and reversible lifelong model editing.", "source": "marker_v2", "marker_block_id": "/page/0/Text/4"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0001", "section": "1. Introduction", "page_start": 1, "page_end": 1, "type": "Text", "text": "Large language models (LLMs) have demonstrated remarkable capabilities in the field of natural language processing, attracting the attention of numerous researchers (Radford et al., 2019; Brown et al., 2020; Achiam et al., 2023; Tou vron et al., 2023) . However, LLMs also face severe challenges, including hallucinations (Huang et al., 2025) , bi-", "source": "marker_v2", "marker_block_id": "/page/0/Text/6"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0002", "section": "1. Introduction", "page_start": 1, "page_end": 1, "type": "FigureGroup", "text": "Figure 1. Comparison of trainable parameters and ERR accuracy between SoLA (Ours) and other methods. All experiments are conducted on Scotus datasets with the same backbone.", "source": "marker_v2", "marker_block_id": "/page/0/FigureGroup/467"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0003", "section": "1. Introduction", "page_start": 1, "page_end": 1, "type": "Text", "text": "ases (Hartvigsen et al., 2022) , and the generation of harmful content (De Cao et al., 2021) . Moreover, the dynamic nature of real-world knowledge requires the continuous updating of specific information in LLMs (Lazaridou et al., 2021) . Re-training these models from scratch is both expensive and time-consuming (Biderman et al., 2023) . In this scenario, the demand for model editing is increasing with aims to modify specific knowledge in the training model continuously without re-training the model or affecting its performance on unrelated inputs, which is known as lifelong model editing (Hartvigsen et al., 2022; Yao et al., 2023; Yu et al., 2024) .", "source": "marker_v2", "marker_block_id": "/page/0/Text/11"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0004", "section": "1. Introduction", "page_start": 1, "page_end": 1, "type": "Text", "text": "The lifelong model editing aims to continuously update the model to adapt to the constantly changing world. However, most of the existing model editing methods are designed for single, isolated editing (Mitchell et al., 2021; Meng et al., 2022a) . When applied to a continuous environment, these methods often lead to catastrophic forgetting of the previously learned knowledge in the LLMs, resulting in performance degradation and reduced model reliability (Yao et al., 2023; Gu et al., 2024) . To solve this problem, recent methods propose freezing the original model parameters and introducing lightweight trainable modules to inject new knowledge. These methods usually retain most of the model's knowledge by keeping the base parameters fixed and combining the learnable modules for specific task updates. For example, Melo (Yu et al., 2024) determines the clustering centres of editors by building a neuron index", "source": "marker_v2", "marker_block_id": "/page/0/Text/12"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0005", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "Text", "text": "and dynamically updates its semantic representation based on the distance from the clustering centres. ELDER (Li et al., 2025) adopts the Mixture-of-Experts (MoE) approach, assigning scores to each LoRA expert through a set of learnable LoRA allocation codes and selecting the top-k LoRA for calculation.", "source": "marker_v2", "marker_block_id": "/page/1/Text/2"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0006", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "FigureGroup", "text": "Figure 2. Number of mismatches in LoRA allocation with ERR/TRR accuracy in MELO. All experiments are conducted on Scotus datasets with the same backbone.", "source": "marker_v2", "marker_block_id": "/page/1/FigureGroup/464"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0007", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "Text", "text": "Although these methods show potential in improving parameter efficiency and reducing interference, they face inherent limitations when applied to lifelong model editing. Specifically, MELO (Yu et al., 2024) improves module separability through semantic clustering, but its clustering centers are updated during the editing process, which may lead to semantic drift and module matching errors. ELDER (Li et al., 2025) employ MoE and selects the top-k routes to activate each editing subset, although this strategy has high parameter efficiency, shared and continuously updated parameters can lead to catastrophic forgetting - new editing may overwrite or interfere with existing editing.", "source": "marker_v2", "marker_block_id": "/page/1/Text/5"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0008", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "Text", "text": "To address these limitations, we propose SoLA, a Semantic routing-based LoRA framework for reversible lifelong model editing. In SoLA, we allocate an independent LoRA module for each edit and establish semantic routing to record the mapping relationship between LoRA module and the input semantic representation. During editing, the LoRA module fully learns for the current task and then remains frozen to retain the learned knowledge. The corresponding key is also not updated. During inference, the semantic representation of the input is calculated, and the corresponding LoRA module is dynamically activated through semantic routing. Since the LoRA module and the semantic representation have not been updated, we fundamentally avoid catastrophic forgetting and semantic drift problems caused by continual updating. At the same time, each edit only requires training for the current LoRA module, significantly reducing the computational resource consumption. As shown in Fig .1, SoLA achieves optimal performance with only 0.08M additional parameters, sig-", "source": "marker_v2", "marker_block_id": "/page/1/Text/6"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0009", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "Text", "text": "nificantly outperforming previous methods, highlighting SoLA's exceptional parameter efficiency and editing accuracy.", "source": "marker_v2", "marker_block_id": "/page/1/Text/7"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0010", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "Text", "text": "More importantly, by removing the semantic key from the mapping memory table, we can precisely revoke specific edits, allowing the model to recover its original behavior without re-training. This means we can freely add and delete edits, and truly achieve controllable rollback and restoration of editing. To our knowledge, this is the first time that controllable rollback of editing has been achieved in existing literature. Moreover, existing works usually require introducing auxiliary routing network outside the edited layer to determine whether to enable the LoRA module. In this paper, we propose a master decision-making mechanism that aggregates the decision-making process into the edited layer, avoiding the reliance on auxiliary routing network and achieving an end-to-end decision-making process.", "source": "marker_v2", "marker_block_id": "/page/1/Text/8"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0011", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "Text", "text": "Extensive experiments validate the effectiveness of our approach. Our main contributions are summarized as follows:", "source": "marker_v2", "marker_block_id": "/page/1/Text/9"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0012", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "ListGroup", "text": "We propose SoLA, a novel framework for reversible lifelong model editing, which incorporates a controllable LoRA mechanism guided by semantic routing. After each edit, both the LoRA modules and the corresponding keys in the mapping memory are frozen, effectively mitigating catastrophic forgetting and semantic drift. Moreover, only the currently active LoRA modules are trained during editing, significantly reducing computational overhead.(Fig .1) . We employs semantic routing to establish a precise mapping between LoRA modules and semantic representations. By removing the associated key, any specific edit can be precisely revoked, enabling flexible addition and deletion of edits. we propose a master decision-making mechanism that aggregates the decision-making process into the edited layer, avoiding the reliance on an auxiliary routing network and achieving an end-to-end decision-making process.", "source": "marker_v2", "marker_block_id": "/page/1/ListGroup/465"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0013", "section": "2.1. Model Editing", "page_start": 2, "page_end": 2, "type": "Text", "text": "Model Editing aims to update specific knowledge within LLMs while preserving their previous knowledge. Current model editing methods can be broadly categorized into three paradigms: Meta-Learning methods, Locate-then-Edit methods, and Memory-Based methods. Meta-Learning Methods employ a hypernetwork to predict the necessary gradients for editing and apply these gradients to the base model to achieve the update (Mitchell et al., 2021) . However, this", "source": "marker_v2", "marker_block_id": "/page/1/Text/15"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0014", "section": "2.1. Model Editing", "page_start": 3, "page_end": 3, "type": "FigureGroup", "text": "Figure 3. Main framework of our method. a) indicates the edited layer in model, where we retain the transformer block of base model frozen and add trainable LoRA module to the edited layer of base model. b) shows the editing process, where every edit will be assigned LoRA module and the query of input will be mapped to assigned LoRA module. c) presents the matching process between query of input and LoRA module in reference. Colour green indicates trainable, while color grey is frozen.", "source": "marker_v2", "marker_block_id": "/page/2/FigureGroup/530"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0015", "section": "2.1. Model Editing", "page_start": 3, "page_end": 3, "type": "Text", "text": "approach typically requires additional data for training the hypernetwork. Locate-then-Edit methods first identify the neurons most critical for the current input via minor perturbations and then directly modify these neurons to update the model's knowledge (Meng et al., 2022a; b) . While effective, these methods are often cumbersome and exhibit limitations when handling a large volume of simultaneous knowledge updates. Memory-Based methods store the information associated with edits in an buffer, which acts as a patch model augmenting the original model (Mitchell et al., 2022) . A significant drawback is that each edit necessitates retraining, making it difficult to adapt to continuous editing scenarios, and the memory bank grows incrementally with each edit, leading to escalating storage overhead. Recent studies have also explored editing constraints to maintain model behavior stability. AlphaEdit (Fang et al., 2024) introduces a null-space constrained approach. It aims to apply edits more precisely by constraining changes to the null space of irrelevant knowledge.", "source": "marker_v2", "marker_block_id": "/page/2/Text/3"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0016", "section": "2.1. Model Editing", "page_start": 3, "page_end": 3, "type": "Text", "text": "However, these methods primarily address static editing, suitable for one-time modifications, and struggle to accommodate sequential editing demands over time. To address these limitations, GRACE (Hartvigsen et al., 2023) replace hidden states with vectors retrieved from a learned codebook. MELO (Yu et al., 2024) introduces a vector database and leverages a clustering mechanism to dynamically assign LoRA modules. ELDER (Li et al., 2025) adopts a MoE framework, where a learnable neural network weights shared LoRA modules. However, these approaches in-", "source": "marker_v2", "marker_block_id": "/page/2/Text/4"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0017", "section": "2.1. Model Editing", "page_start": 3, "page_end": 3, "type": "Text", "text": "evitably update the semantic representation in sequential edits, leading to semantic drift or knowledge forgetting.", "source": "marker_v2", "marker_block_id": "/page/2/Text/5"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0018", "section": "3.1. Preliminaries", "page_start": 3, "page_end": 3, "type": "Text", "text": "Lifelong Model Editing The purpose of lifelong model editing, as described by Grace (Hartvigsen et al., 2023) , is to make continual edits to the knowledge of the initial base model without degrading the performance of the base model and without negating previous edits. Consider an initial base model fbase, After n edits Dedit = {d1, ..., dn}, some of the model's knowledge is updated and model is transformed into fn, where d i = (x i , yi), For lifetime model editing, the edited model f n should be able to correctly output y within the edited inputs x ∈ Dedit, i.e., fn(xi) = y i . Furthermore, for inputs x ′ ∈/ Dedit that are semantically similar to the edited instances, such as rephrased sentences, the model f n is expected to generalize the updated knowledge appropriately, like fn(x ′ i ) = y i . Moreover, after editing, the model f n should also retain the previous knowledge of fbase, that is, for data samples x j ∈/ Dedit that have not been edited, there should be fn(x j ) = fbase(x j ) = y j .", "source": "marker_v2", "marker_block_id": "/page/2/Text/8"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0019", "section": "3.1. Preliminaries", "page_start": 3, "page_end": 3, "type": "Text", "text": "LoRA for Lifelong Model Editing Low-Rank Adaptation (LoRA) (Hu et al., 2022) is an efficient finetuning method for LLMs that optimizes model outputs by projecting features into a low-rank subspace. In the context of efficient LLM finetuning, LoRA modules are typically inserted into", "source": "marker_v2", "marker_block_id": "/page/2/Text/9"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0020", "section": "3.1. Preliminaries", "page_start": 4, "page_end": 4, "type": "Text", "text": "specific layers of the pre-trained model. During editing, the base LLM parameters W_0 \\in \\mathbb{R}^{d \\times k} remain frozen, while only the LoRA module \\Delta W \\in \\mathbb{R}^{d \\times k} is updated. The parameter update \\Delta W can be represented as a low-rank decomposition of the pretrained weight matrix. Specifically, \\Delta W is factorized into two smaller matrices A \\in \\mathbb{R}^{r \\times k} and B \\in \\mathbb{R}^{d \\times r} , such that \\Delta W = BA , where r \\ll \\min(d,k) and r is the LoRA rank. Given the original h = W_0 x , the forward propagation process is modified as:", "source": "marker_v2", "marker_block_id": "/page/3/Text/1"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0021", "section": "3.1. Preliminaries", "page_start": 4, "page_end": 4, "type": "Equation", "text": "\\boldsymbol{h} = W_0 \\boldsymbol{x} + \\Delta W \\boldsymbol{x} = W_0 \\boldsymbol{x} + B A \\boldsymbol{x} \\tag{1}", "source": "marker_v2", "marker_block_id": "/page/3/Equation/2"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0022", "section": "3.1. Preliminaries", "page_start": 4, "page_end": 4, "type": "Text", "text": "where A is initialized using a zero mean Gaussian distribution and B is initialized as a zero matrix. This approach significantly reduces the number of trainable parameters while maintaining the expressive power of the full-rank update.", "source": "marker_v2", "marker_block_id": "/page/3/Text/3"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0023", "section": "3.2. Cluster Updating in Lifelong Model Editing", "page_start": 4, "page_end": 4, "type": "Text", "text": "In the context of lifelong model editing, when multiple adapter modules are involved, the model needs to dynamically manage the association between edits and the corresponding adapter module. MELO addresses this by employing a clustering mechanism that assigns edits to clusters based on the hidden representations of the inputs, with cluster centres being continuously updated throughout editing. However, this continual updating of cluster centers alters their semantic representation, leading to semantic drift and incorrect module assignments. Moreover, repeated training of LoRA modules can result in forgetting previously acquired knowledge. Varying the cluster radius in MeLO results in a different number of generated cluster centres, thereby affecting the frequency of cluster centre updates. We record MeLO's performance under different numbers of cluster centre updates, as shown in Fig.2. The experiment is conducted on SCOTUS dataset. In Fig.2, the number of updates refers to the number of cluster centre updates during the editing process, and the number of mismatches indicates how many times, during inference, the retrieved LoRA module differs from the one assigned during editing. ERR and TRR represent the model's accuracy on the edited and unedited datasets, respectively, after all editing tasks are completed.", "source": "marker_v2", "marker_block_id": "/page/3/Text/5"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0024", "section": "3.2. Cluster Updating in Lifelong Model Editing", "page_start": 4, "page_end": 4, "type": "Text", "text": "As shown in the Fig.2, for the same datasets SCOTUS of editing, increasing the number of cluster centre updates leads to a rise in mismatches during inference, and a corresponding decline in model performance on both the edited and unedited datasets. This demonstrates that cluster centre updates can cause semantic drift, resulting in incorrect LoRA match, while repeated training of LoRA modules contributes to knowledge forgetting. Therefore, in SoLA, we freeze both the LoRA modules and their associated keys once the current task is completed. These components are", "source": "marker_v2", "marker_block_id": "/page/3/Text/6"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0025", "section": "3.2. Cluster Updating in Lifelong Model Editing", "page_start": 4, "page_end": 4, "type": "Text", "text": "no longer updated in subsequent edits, thereby maximizing knowledge retention across editing iterations.", "source": "marker_v2", "marker_block_id": "/page/3/Text/7"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0026", "section": "3.3. Semantic Routing-Based Controllable LoRA", "page_start": 4, "page_end": 4, "type": "Text", "text": "In the layer where the model needs to be edited, we keep the parameters of the original model frozen and insert the learnable LoRA module for editing, as shown in Fig.3 a). In SoLA, each editing operation is encapsulated within an independent LoRA module. During sequential editing, for a given editing task d_i = (\\boldsymbol{x}_i^m, \\boldsymbol{y}_i^m)_{m=1}^n , a dedicated LoRA module LoRA<sub>i</sub> is assigned. For each data instance \\boldsymbol{x}_i^m \\in d_i , the corresponding LoRA id i is recorded, and a semantic routing entry is established by associating LoRA module with a semantic key derived from the input representation, as shown in Fig.3 b). Following prior work(Hartvigsen et al., 2023), we use the hidden representation of the last token in the input sequence as the key e \\in \\mathbb{R}^d vector. During editing, the assigned LoRA module LoRA<sub>i</sub> will fully learn the task d_i , yielding the output:", "source": "marker_v2", "marker_block_id": "/page/3/Text/9"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0027", "section": "3.3. Semantic Routing-Based Controllable LoRA", "page_start": 4, "page_end": 4, "type": "Equation", "text": "h = h_0 + LoRA_i(x) (2)", "source": "marker_v2", "marker_block_id": "/page/3/Equation/10"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0028", "section": "3.3. Semantic Routing-Based Controllable LoRA", "page_start": 4, "page_end": 4, "type": "Text", "text": "where h_0 \\in \\mathbb{R}^{l \\times d} is the frozen base model representation, l is the sequence length, d is the hidden state dimension. At this stage, all other LoRA modules and the stored key vectors remain frozen. Upon completion of the editing process, the trained LoRA<sub>i</sub> module is frozen and will be stored with its associated key e_i as a fixed mapping for future reference. Neither the LoRA module nor the key will be updated in subsequent editing. Since only the current LoRA module is involved in training during each edit, SoLA significantly reduces the number of trainable parameters, as shown in Fig.1. Furthermore, freezing both the LoRA module and its corresponding key after editing prevents semantic drift and mitigates catastrophic forgetting, thereby maximizing knowledge retention. During inference, the hidden representation of the last token in the input sequence will serve as a query vector \\boldsymbol{q} \\in \\mathbb{R}^d from the input \\boldsymbol{x} and matches against the stored keys K \\in \\mathbb{R}^{N \\times d} , as shown in Fig.3 c). If a match is found, the corresponding LoRA module is retrieved from stored LoRA pool R \\in \\mathbb{R}^M and incorporated into the computation.", "source": "marker_v2", "marker_block_id": "/page/3/Text/11"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0029", "section": "3.4. Master Decision Mechanism", "page_start": 4, "page_end": 4, "type": "Text", "text": "During inference, existing approaches usually need to introduce an auxiliary routing network outside the target edited layer to decide whether to activate the LoRA module or not(Yu et al., 2024; Li et al., 2025), which hampers the end-to-end decision-making process. To address this limitation, we propose a master decision-making mechanism that does not require an auxiliary routing component. Specifically, we designate the first edited layer as the master decision layer where input features are dynamically evaluated to activate the relevant LoRA module without the need for an auxil-", "source": "marker_v2", "marker_block_id": "/page/3/Text/13"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0030", "section": "3.4. Master Decision Mechanism", "page_start": 5, "page_end": 5, "type": "TableGroup", "text": "Table 1. Comparison of SoLA to existing methods. All the results are obtained after sequential edits. ERR indicates Edit Reliability Rate, TRR presents Task Retention Rate, and ARR is Accuracy Attention Rate. The best performance is shown in bold. SCOTUS (BERT; Acc ↑) zsRE (T5; F1 ↑) Hallucination (GPT2-XL; PPL ↓) Method ERR TRR Avg. ERR TRR Avg. ERR TRR ARR EWC 0.51 0.82 0.66 0.67 0.50 0.58 1485.7 29.24 109.59 CMR 0.52 0.52 0.52 0.56 0.82 0.69 1449.3 28.14 107.76 CLEAR 0.67 0.83 0.75 0.27 0.99 0.63 2394.3 35.34 195.82 MEND 0.19 0.27 0.23 0.25 0.27 0.26 1369.8 1754.9 2902.5 SERAC 0.33 0.41 0.37 0.72 0.31 0.52 8183.7 133.3 10.04 ROME - - - - - - 30.28 103.82 14.02 GRACE 0.81 0.82 0.82 0.69 0.96 0.83 15.84 7.14 10.00 ELDER 0.89 0.86 0.88 0.72 0.97 0.84 16.12 5.87 8.42 MELO 0.96 0.92 0.94 0.72 0.98 0.85 17.45 1.04 2.66 SoLA 0.97 0.95 0.96 0.73 0.99 0.86 15.15 1.01 7.35", "source": "marker_v2", "marker_block_id": "/page/4/TableGroup/521"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0031", "section": "3.4. Master Decision Mechanism", "page_start": 5, "page_end": 5, "type": "Text", "text": "iary network. In this framework, the master decision layer H e computes the distance metric between the input feature embedding q and the stored key K to determine the activation of the module: d = He(argmini(dist(q, ki))), i = (1, 2, ..., N), dist(·) denotes the distance function and each key k i associated with the LoRA module. Then in the master layer He, we decide the edit layer behaviour:", "source": "marker_v2", "marker_block_id": "/page/4/Text/5"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0032", "section": "3.4. Master Decision Mechanism", "page_start": 5, "page_end": 5, "type": "Equation", "text": "H_e = \\begin{cases} H(W_0) & \\text{if } d \\ge \\alpha \\\\ H(W_0, W_{R_m}) & \\text{if } d < \\alpha \\end{cases} (3)", "source": "marker_v2", "marker_block_id": "/page/4/Equation/6"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0033", "section": "3.4. Master Decision Mechanism", "page_start": 5, "page_end": 5, "type": "Text", "text": "where α is the threshold, in this work we set it as 0.01, WR m represents the weight of m-th LoRA module associated with the key nearest to the query. This binary decision is propagated to the subsequent edited layers, ensuring consistent module activation throughout the network. By integrating the decision-making process to the first edited layer, our approach achieves complete end-to-end decision-making capability while maintaining architectural simplicity.", "source": "marker_v2", "marker_block_id": "/page/4/Text/7"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0034", "section": "4.1. Implementation Details", "page_start": 5, "page_end": 5, "type": "Text", "text": "Baselines We first compare several recent methods designed for lifelong model editing. Grace (Hartvigsen et al., 2023) replaces model outputs with key-value pairs and matches features to keys based on deferral radius. MELO (Yu et al., 2024) leverages LoRA for finetuning and dynamically retrieves relevant LoRA modules via neuron index. EL-DER (Li et al., 2025) employs MoE to dynamically combine multiple LoRA modules. Additionally, we evaluate other baseline methods. MEND (Mitchell et al., 2021) uses a hypernetwork trained on auxiliary data to predict the required gradients for model editing. SERAC (Mitchell et al., 2022)", "source": "marker_v2", "marker_block_id": "/page/4/Text/10"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0035", "section": "4.1. Implementation Details", "page_start": 5, "page_end": 5, "type": "Text", "text": "learns a scope classifier and a counterfactual model, coordinating between modules to determine editing behavior. ROME (Meng et al., 2022a) identifies the most influential weights in the model through perturbation analysis, localizes knowledge to specific layers of GPT, and modifies the corresponding weights. Since ROME is specifically designed for GPT, it is only evaluated on the hallucination correction task. EWC (Kirkpatrick et al., 2017) imposes constraints on the weights updating so that the model continuously learns new knowledge and retains previous knowledge. CMR (Lin et al., 2022) performs continuous finetuning of the model to continuously learn new knowledge. CLEAR (Rolnick et al., 2019) builds a memory buffer to restore old tasks information, and replays them in subsequent edits to retain previous knowledge.", "source": "marker_v2", "marker_block_id": "/page/4/Text/11"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0036", "section": "4.1. Implementation Details", "page_start": 5, "page_end": 5, "type": "Text", "text": "Evaluation Metric In this work, following (Hartvigsen et al., 2023) , we use three basic metrics to assess model performance: ES, ERR and TRR, which are applicable to all model experiments. Where Edit Success(ES) measures the ability of the model to edit the target sample. Edit Reliability Rate(ERR) assesses the validity of the model on the edited dataset, indicating the extent to which the required knowledge updates are successfully incorporated. Task Retention Rate(TRR) quantifies the model's performance on unedited datasets and reflects the model's ability to retain previous knowledge. Additionally, in the hallucination correction task, we also use the Accuracy Attention Rate (ARR) to assess the performance of the already accurate output. Accuracy is measured differently for different tasks. For the document classification task of SCOTUS, we use the Average Accuracy Rate (ACC); for the question answering task of zsRE, we measure it using the average F1, and for the hallucination correction task, we measure it using the standard average complexity (PPL) (Brown et al., 1992) .", "source": "marker_v2", "marker_block_id": "/page/4/Text/12"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0037", "section": "4.1. Implementation Details", "page_start": 6, "page_end": 6, "type": "FigureGroup", "text": "Figure 4. Figure(a)–(c) presents every edit accuracy on SCOTUS datasets with BERT. All experiment settings are the same as Tab .1.", "source": "marker_v2", "marker_block_id": "/page/5/FigureGroup/386"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0038", "section": "4.1. Implementation Details", "page_start": 6, "page_end": 6, "type": "FigureGroup", "text": "Figure 5. Visualization of zsRE dataset encoder output from T5 small with t-SNE and the dots with same color and shape are input and rephrase sentence. All experiment settings are the same as Tab .1.", "source": "marker_v2", "marker_block_id": "/page/5/FigureGroup/387"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0039", "section": "4.2. Main Results", "page_start": 6, "page_end": 6, "type": "Text", "text": "We evaluate the performance of the proposed SoLA method on three benchmark datasets and compare it with several state-of-the-art model editing approaches. The results are summarized in Tab .1. As observed, SoLA achieves excellent performance on most of the benchmark datasets, with SoLA outperforming the strongest baseline, MELO, by 3% on the SCOTUS dataset. These demonstrate SoLA's enhanced effectiveness in editing target knowledge while better preserving previously acquired knowledge. Furthermore, existing methods such as MEND and SERAC perform poorly under the sequential editing setting, suggesting that they are primarily designed for static editing scenarios and are not well-suited for sequential model editing tasks.", "source": "marker_v2", "marker_block_id": "/page/5/Text/7"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0040", "section": "4.2. Main Results", "page_start": 6, "page_end": 6, "type": "Text", "text": "We also report task-wise accuracy progression to analyze model performance throughout the sequential editing process. The experimental results are based on the SCOTUS dataset and the results are shown in Fig .4. As illustrated in Fig .4a, Fig .4b and Fig .4c, SoLA consistently maintains", "source": "marker_v2", "marker_block_id": "/page/5/Text/8"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0041", "section": "4.2. Main Results", "page_start": 6, "page_end": 6, "type": "Text", "text": "superior accuracy across tasks, without exhibiting substantial performance degradation or volatility. This indicates the robustness and stability of SoLA when applied to sequential knowledge editing.", "source": "marker_v2", "marker_block_id": "/page/5/Text/9"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0042", "section": "4.2. Main Results", "page_start": 6, "page_end": 6, "type": "Text", "text": "Method performance may vary across different model sizes and datasets. To further evaluate the robustness of our approach, we conduct hallucination correction experiments on the UniEdit (Chen et al., 2025) and WikiBigEdit (Thede et al., 2025) datasets using larger-scale backbone models. The results are presented in Tab .2. As shown in Tab .2, SoLA consistently achieves the best or near-best performance across most settings, demonstrating its robustness and effectiveness under diverse task configurations. During editing, larger backbone models generally exhibit more stable ERR values, suggesting that such models have learned stronger semantic representations during pretraining, leading to better generalization and editability. Furthermore, continuous learning related methods tend to overfit in edited data, thus showing better ARR in the retention set, but large TRR values in the upstream set, indicating severe forgetting of previous knowledge.", "source": "marker_v2", "marker_block_id": "/page/5/Text/10"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0043", "section": "4.3. Controllable Rollback Editing", "page_start": 6, "page_end": 6, "type": "Text", "text": "In SoLA, each edit instance is associated with a unique key stored in both the memory and the routing table. This design enables reversible editing, as the model can revert to its original behavior by removing the key corresponding to a specific edit. To demonstrate this property, we conduct an illustrative experiment on the zsRE dataset. The results are shown in Tab .3. In the Tab .3, each row corresponds to an input instance consisting of a \"Text\" (the question need to be edited) and its \"Labels\" (the correct answer). \"Del\" indicates whether the key corresponding to the edit is removed after editing. For each input, Predbase, Prededit and Preddel denote the prediction of base model, edited model and model after removing the corresponding key. When \"Del\" is false, the key is retained, serving as a baseline comparison. As shown in Tab .3, for each instance, the original prediction (Predbase) does not match the correct label, indi-", "source": "marker_v2", "marker_block_id": "/page/5/Text/12"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0044", "section": "4.3. Controllable Rollback Editing", "page_start": 7, "page_end": 7, "type": "TableGroup", "text": "Table 2. Model Performance Comparison on UniEdit and WikiBigEdit datasets in hallucination correction tasks. All the scores are PPL metric and the best result is in bolded. UniEdit (PPL ↓) WikiBigEdit (PPL ↓) Model Method ERR TRR ARR ERR TRR ARR EWC 2.43 1134 22.08 3.58 251117 4.93 CMR 2.79 2239 25.24 2.86 169452 4.49 LLaMA-3-8B CLEAR 1.04 982 29.79 1.03 74384 1.52 SERAC 37.10 246 93 60.17 215 59 GRACE 1.00 244 89 1.00 216 58 ELDER 1.00 173 74 1.00 178 48 MELO 1.00 153 68 1.00 165 47 Ours 1.00 144 66 1.00 162 44 EWC 2.38 4451 42.13 2.86 47846 3.89 CMR 2.34 7211 36.19 3.60 34623 5.51 DeepSeek-R1-8B CLEAR 1.06 3365 44.86 1.02 28110 1.48 SERAC 46.76 451 834 82.72 493 831 GRACE 1.00 407 803 1.00 487 793 ELDER 1.03 241 482 1.03 304 492 MELO 1.09 204 407 1.05 263 445 Ours 1.07 188 374 1.00 248 412 EWC 4.62 7648 48.62 5.61 61834 14.65 CMR 4.12 11583 40.56 6.28 48629 16.89 Qwen2-7B CLEAR 1.03 5021 54.98 1.05 35396 4.26 SERAC 58.51 329 1261 62.12 304 523 GRACE 2.05 316 1065 16.88 299 518 ELDER 1.00 201 266 1.00 225 197 MELO 1.00 170 201 1.00 203 143 Ours 1.00 156 110 1.00 199 109", "source": "marker_v2", "marker_block_id": "/page/6/TableGroup/486"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0045", "section": "4.3. Controllable Rollback Editing", "page_start": 7, "page_end": 7, "type": "Text", "text": "cating the necessity for editing. After editing, the model correctly produces Prededit consistent with the ground-truth label, demonstrating that SoLA successfully updates the model's knowledge. When the associated key is deleted, the model reverts to its original prediction (Prededit = Predbase), verifying that our method can effectively roll back specific edits. More critically, for edits where the key is not deleted, the model continues to output Prededit, showing that the rollback operation does not interfere with other edits. This confirms that SoLA supports fine-grained, selective undo of knowledge edits without disrupting unrelated modifications.", "source": "marker_v2", "marker_block_id": "/page/6/Text/3"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0046", "section": "4.4. Ablation Study", "page_start": 7, "page_end": 7, "type": "Text", "text": "Effect of Edit Layer Location In semantic representation learning, different layers of a model focus on different aspects of semantic information. Therefore, the position of the edited layer can significantly influence the effectiveness of model editing. To investigate this, we evaluate the performance of model when edits are applied at different layers, while keeping all other settings fixed. The experiments are conducted on the SCOTUS dataset with a Nvidia A100 40G GPU, and the results are shown in Tab .4. In the", "source": "marker_v2", "marker_block_id": "/page/6/Text/5"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0047", "section": "4.4. Ablation Study", "page_start": 7, "page_end": 7, "type": "Text", "text": "Tab .4, \"Layer\" denotes the layer range used for editing. For example, \"0–2\" indicates that edits are applied to layers 0, 1, and 2 of the BERT model. As shown in the table, editing at shallower layers results in suboptimal performance. This observation aligns with prior findings (Geva et al., 2020) , which suggest that shallow layers primarily capture the shallow sentence patterns, whereas deeper layers encode richer semantic features, thereby enabling more effective knowledge editing. Moreover, editing at shallow layers significantly increases the time used for editing, which means that editing for semantics is more difficult in shallow layers. Furthermore, we observe that varying the editing layer has negligible impact on performance over the unedited portion of the dataset, indicating that SoLA is capable of preserving previously learned knowledge regardless of the editing depth, which further supports the robustness of SoLA.", "source": "marker_v2", "marker_block_id": "/page/6/Text/6"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0048", "section": "4.4. Ablation Study", "page_start": 7, "page_end": 7, "type": "Text", "text": "Effect of LoRA Rank The LoRA module projects semantic representations into a low-rank subspace for parameterefficient learning. Consequently, the choice of rank will effect the module's performance. To investigate this, we analyze the impact of different LoRA rank values on model performance, keeping all other settings fixed. Experiments are conducted on the SCOTUS dataset using a NVIDIA", "source": "marker_v2", "marker_block_id": "/page/6/Text/7"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0049", "section": "4.4. Ablation Study", "page_start": 8, "page_end": 8, "type": "TableGroup", "text": "Table 3. Controllable edit in SoLA on zsRE datasets. \"Text\" is the question need to be edited, \"Labels\" is the true answer, \"Del\" indicates whther delete associated key and Predbase, Prededit and Preddel is the prediction of base model, edited model and delted model. Text Labels Del Predbase Prededit Preddel The date of Tsetserleg earthquake in 1905? 9 July 1905 True 20 February 1905 9 July 1905 20 February 1905 What constellation has Nu Cancri? Cancer True Hydra Cancer Hydra On what date did the Bahrain Grand Prix 2015 occur? 19 April 2015 True 6 April 2017 19 April 2015 6 April 2017 What position does Charles-Joseph Coursol have? Mayor of Montreal True Positioning Mayor of Montreal Positioning Who acted in Blow Dry? Alan Rickman False Jim Carrey Alan Rickman Alan Rickman Table 4. Ablation studies on location of edited layers. Here, \"Layer\" indicates the edited layer location. For example, \"0-2\" presents edited layer is layers 0, 1 and 2.", "source": "marker_v2", "marker_block_id": "/page/7/TableGroup/457"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0050", "section": "4.4. Ablation Study", "page_start": 8, "page_end": 8, "type": "TableGroup", "text": "Layer ES ERR TRR Edit Time(min) 0-2 0.77 0.61 0.96 9.21 3-5 1.00 0.68 0.96 8.06 6-8 0.81 0.70 0.97 7.03 9-11 0.94 0.95 0.97 5.97 Table 5. Ablation studies on LoRA rank. Here, \"LoRA Rank\" indicates the rank value of every LoRA in model.", "source": "marker_v2", "marker_block_id": "/page/7/TableGroup/458"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0051", "section": "4.4. Ablation Study", "page_start": 8, "page_end": 8, "type": "Table", "text": "LoRA Rank ES ERR TRR Edit Time(min) 1 0.90 0.84 0.96 5.88 2 0.94 0.91 0.97 5.90 3 1.00 0.91 0.97 5.86 4 1.00 0.95 0.97 5.97 5 1.00 0.92 0.96 5.86 10 1.00 0.71 0.96 5.85", "source": "marker_v2", "marker_block_id": "/page/7/Table/17"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0052", "section": "4.4. Ablation Study", "page_start": 8, "page_end": 8, "type": "Text", "text": "A100 40GB GPU, and the results are presented in Tab .5. As shown in the Tab .5, simply increasing the LoRA rank does not lead to improved performance and may even degrade model performance. This indicates that expanding the learnable parameter space does not necessarily enhance the model's capacity for effective knowledge editing. On the contrary, excessive rank values may even result in performance degradation due to overfitting, suggesting the importance of carefully selecting an appropriate rank to balance capacity and generalization. Based on the experiment results, we set 4 as LoRA rank for all the experiments in this paper.", "source": "marker_v2", "marker_block_id": "/page/7/Text/18"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0053", "section": "4.5. Model Visualization", "page_start": 8, "page_end": 8, "type": "Text", "text": "To gain deeper insight into the model's behavior under sequential editing, we conducted a t-SNE visualization (Maaten & Hinton, 2008) of the learned feature representations. This experiment is performed on the zsRE dataset using the T5-small model to assess how semantic relationships are preserved after sequential edits. Upon completing", "source": "marker_v2", "marker_block_id": "/page/7/Text/20"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0054", "section": "4.5. Model Visualization", "page_start": 8, "page_end": 8, "type": "Text", "text": "the entire sequence of editing tasks for question answering, we selected five distinct input instances, each accompanied by semantically equivalent rephrased sentences, forming five evaluation groups. The encoder output features are subsequently extracted from the edited model for dimensionality reduction and visualization, and the resulting plot is represented in Fig .5. As shown in Fig .5, there is a clear clustering pattern that for each group, the original input and its rephrased variant are mapped to proximate locations in the latent space. This demonstrates that SoLA effectively captures and preserves the semantic similarity between related inputs throughout the editing process. Furthermore, distinct groups maintain separable clusters, indicating that the model retains the ability to discriminate between different edit concepts after sequential editing. This visualization shows that SoLA's semantic routing mechanism successfully supports edit generalization based on semantic similarity while upholding the integrity of previously learned knowledge.", "source": "marker_v2", "marker_block_id": "/page/7/Text/21"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0055", "section": "5. Conclusion", "page_start": 8, "page_end": 8, "type": "Text", "text": "In this paper, we propose SoLA, a Semantic routing-based LoRA framework for reversible lifelong model editing. In SoLA, each edit is encapsulated as an independent LoRA module which is frozen after training and a semantic routing record is established to map LoRA module to the input semantic representation, allowing dynamic activation of LoRA modules via semantic matching. This design not only avoids semantic drift caused by clustering updating but also mitigates catastrophic forgetting from parameter sharing. More importantly, SoLA supports precisely revoking edit by removing the corresponding key from the routing table, enabling reversible model editing. This allows for flexible addition and deletion of edits, offering fine-grained control over model behavior. To the best of our knowledge, this is the first work to achieve reversible model editing in the literature. Furthermore, we introduce a master decisionmaking mechanism by integrating decision-making into the edited layer, enabling end-to-end decision-making process. By building a strict mapping between edits and LoRA modules, SoLA achieves efficient and reversible lifelong model editing, providing a novel perspective for future research.", "source": "marker_v2", "marker_block_id": "/page/7/Text/23"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0056", "section": "Impact Statement", "page_start": 9, "page_end": 9, "type": "Text", "text": "This work advances the field of Machine Learning by introducing SoLA, a framework for reversible and controllable lifelong model editing in large language models. Our method enhances the safety and reliability of AI systems by enabling precise rollback of model edits, which helps mitigate risks associated with erroneous or harmful knowledge updates. By effectively preventing semantic drift and catastrophic forgetting, SoLA promotes the development of more robust and trustworthy models capable of adapting to evolving information. The significant reduction in computational overhead also aligns with the goal of sustainable AI development.", "source": "marker_v2", "marker_block_id": "/page/8/Text/2"}

icml26/31f6f2e8-0fb2-46ff-ab65-f3408612f6e1/marker_meta.json

@@ -0,0 +1,1241 @@
+{
+  "table_of_contents": [
+    {
+      "title": "Reversible Lifelong Model Editing via Semantic Routing-Based LoRA",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          84.677001953125,
+          87.78515625
+        ],
+        [
+          512.7245483398438,
+          87.78515625
+        ],
+        [
+          512.7245483398438,
+          104.2581787109375
+        ],
+        [
+          84.677001953125,
+          104.2581787109375
+        ]
+      ]
+    },
+    {
+      "title": "Anonymous Authors1",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          240.85546875,
+          140.765625
+        ],
+        [
+          333.2859191894531,
+          140.765625
+        ],
+        [
+          333.2859191894531,
+          152.96612548828125
+        ],
+        [
+          240.85546875,
+          152.96612548828125
+        ]
+      ]
+    },
+    {
+      "title": "Abstract",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          147.322265625,
+          175.18359375
+        ],
+        [
+          194.68328857421875,
+          175.18359375
+        ],
+        [
+          194.68328857421875,
+          188.015380859375
+        ],
+        [
+          147.322265625,
+          188.015380859375
+        ]
+      ]
+    },
+    {
+      "title": "1. Introduction",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          54.087890625,
+          557.9141845703125
+        ],
+        [
+          132.27606201171875,
+          557.9141845703125
+        ],
+        [
+          132.27606201171875,
+          569.869384765625
+        ],
+        [
+          54.087890625,
+          569.869384765625
+        ]
+      ]
+    },
+    {
+      "title": "2. Related Work",
+      "heading_level": null,
+      "page_id": 1,
+      "polygon": [
+        [
+          307.44000244140625,
+          583.55859375
+        ],
+        [
+          390.55255126953125,
+          583.55859375
+        ],
+        [
+          390.55255126953125,
+          596.1283569335938
+        ],
+        [
+          307.44000244140625,
+          596.1283569335938
+        ]
+      ]
+    },
+    {
+      "title": "2.1. Model Editing",
+      "heading_level": null,
+      "page_id": 1,
+      "polygon": [
+        [
+          306.59765625,
+          605.0755004882812
+        ],
+        [
+          386.68359375,
+          605.0755004882812
+        ],
+        [
+          386.68359375,
+          615.0381011962891
+        ],
+        [
+          306.59765625,
+          615.0381011962891
+        ]
+      ]
+    },
+    {
+      "title": "3. Method",
+      "heading_level": null,
+      "page_id": 2,
+      "polygon": [
+        [
+          306.59765625,
+          400.76617431640625
+        ],
+        [
+          359.24188232421875,
+          400.76617431640625
+        ],
+        [
+          359.24188232421875,
+          412.72137451171875
+        ],
+        [
+          306.59765625,
+          412.72137451171875
+        ]
+      ]
+    },
+    {
+      "title": "3.1. Preliminaries",
+      "heading_level": null,
+      "page_id": 2,
+      "polygon": [
+        [
+          307.1953125,
+          419.58984375
+        ],
+        [
+          382.2491760253906,
+          419.58984375
+        ],
+        [
+          382.2491760253906,
+          431.631103515625
+        ],
+        [
+          307.1953125,
+          431.631103515625
+        ]
+      ]
+    },
+    {
+      "title": "3.2. Cluster Updating in Lifelong Model Editing",
+      "heading_level": null,
+      "page_id": 3,
+      "polygon": [
+        [
+          53.19140625,
+          288.0
+        ],
+        [
+          260.25,
+          288.0
+        ],
+        [
+          260.25,
+          298.546875
+        ],
+        [
+          53.19140625,
+          298.546875
+        ]
+      ]
+    },
+    {
+      "title": "3.3. Semantic Routing-Based Controllable LoRA",
+      "heading_level": null,
+      "page_id": 3,
+      "polygon": [
+        [
+          304.8046875,
+          105.75
+        ],
+        [
+          514.5,
+          105.75
+        ],
+        [
+          514.5,
+          115.5
+        ],
+        [
+          304.8046875,
+          115.5
+        ]
+      ]
+    },
+    {
+      "title": "3.4. Master Decision Mechanism",
+      "heading_level": null,
+      "page_id": 3,
+      "polygon": [
+        [
+          305.25,
+          580.46484375
+        ],
+        [
+          447.0,
+          580.46484375
+        ],
+        [
+          447.0,
+          589.5
+        ],
+        [
+          305.25,
+          589.5
+        ]
+      ]
+    },
+    {
+      "title": "4. Experiments",
+      "heading_level": null,
+      "page_id": 4,
+      "polygon": [
+        [
+          52.892578125,
+          548.3081665039062
+        ],
+        [
+          132.47930908203125,
+          548.3081665039062
+        ],
+        [
+          132.47930908203125,
+          560.2633666992188
+        ],
+        [
+          52.892578125,
+          560.2633666992188
+        ]
+      ]
+    },
+    {
+      "title": "4.1. Implementation Details",
+      "heading_level": null,
+      "page_id": 4,
+      "polygon": [
+        [
+          51.99609375,
+          569.2105102539062
+        ],
+        [
+          172.76953125,
+          569.2105102539062
+        ],
+        [
+          172.76953125,
+          579.69140625
+        ],
+        [
+          51.99609375,
+          579.69140625
+        ]
+      ]
+    },
+    {
+      "title": "4.2. Main Results",
+      "heading_level": null,
+      "page_id": 5,
+      "polygon": [
+        [
+          54.38671875,
+          478.37109375
+        ],
+        [
+          129.05364990234375,
+          478.37109375
+        ],
+        [
+          129.05364990234375,
+          489.50909423828125
+        ],
+        [
+          54.38671875,
+          489.50909423828125
+        ]
+      ]
+    },
+    {
+      "title": "4.3. Controllable Rollback Editing",
+      "heading_level": null,
+      "page_id": 5,
+      "polygon": [
+        [
+          305.40234375,
+          500.02734375
+        ],
+        [
+          453.1230773925781,
+          500.02734375
+        ],
+        [
+          453.1230773925781,
+          510.422119140625
+        ],
+        [
+          305.40234375,
+          510.422119140625
+        ]
+      ]
+    },
+    {
+      "title": "4.4. Ablation Study",
+      "heading_level": null,
+      "page_id": 6,
+      "polygon": [
+        [
+          53.19140625,
+          582.7525024414062
+        ],
+        [
+          137.37242126464844,
+          582.7525024414062
+        ],
+        [
+          137.37242126464844,
+          592.7151031494141
+        ],
+        [
+          53.19140625,
+          592.7151031494141
+        ]
+      ]
+    },
+    {
+      "title": "4.5. Model Visualization",
+      "heading_level": null,
+      "page_id": 7,
+      "polygon": [
+        [
+          51.697265625,
+          628.9865112304688
+        ],
+        [
+          158.02487182617188,
+          628.9865112304688
+        ],
+        [
+          158.02487182617188,
+          638.9491119384766
+        ],
+        [
+          51.697265625,
+          638.9491119384766
+        ]
+      ]
+    },
+    {
+      "title": "5. Conclusion",
+      "heading_level": null,
+      "page_id": 7,
+      "polygon": [
+        [
+          307.44000244140625,
+          437.7861633300781
+        ],
+        [
+          376.52911376953125,
+          437.7861633300781
+        ],
+        [
+          376.52911376953125,
+          449.7413635253906
+        ],
+        [
+          307.44000244140625,
+          449.7413635253906
+        ]
+      ]
+    },
+    {
+      "title": "Impact Statement",
+      "heading_level": null,
+      "page_id": 8,
+      "polygon": [
+        [
+          52.59375,
+          67.84710693359375
+        ],
+        [
+          146.74185180664062,
+          67.84710693359375
+        ],
+        [
+          146.74185180664062,
+          79.80230712890625
+        ],
+        [
+          52.59375,
+          79.80230712890625
+        ]
+      ]
+    },
+    {
+      "title": "References",
+      "heading_level": null,
+      "page_id": 8,
+      "polygon": [
+        [
+          53.490234375,
+          247.2921142578125
+        ],
+        [
+          110.9838638305664,
+          247.2921142578125
+        ],
+        [
+          110.9838638305664,
+          259.247314453125
+        ],
+        [
+          53.490234375,
+          259.247314453125
+        ]
+      ]
+    },
+    {
+      "title": "550",
+      "heading_level": null,
+      "page_id": 10,
+      "polygon": [
+        [
+          23.009765625,
+          69.16845703125
+        ],
+        [
+          41.537109375,
+          69.16845703125
+        ],
+        [
+          41.537109375,
+          89.71875
+        ],
+        [
+          23.009765625,
+          89.71875
+        ]
+      ]
+    },
+    {
+      "title": "Appendix",
+      "heading_level": null,
+      "page_id": 10,
+      "polygon": [
+        [
+          258.1875,
+          70.6898193359375
+        ],
+        [
+          334.00738525390625,
+          70.6898193359375
+        ],
+        [
+          334.00738525390625,
+          88.171875
+        ],
+        [
+          258.1875,
+          88.171875
+        ]
+      ]
+    },
+    {
+      "title": "A. Additional Related Work",
+      "heading_level": null,
+      "page_id": 10,
+      "polygon": [
+        [
+          53.490234375,
+          102.8671875
+        ],
+        [
+          198.67526245117188,
+          102.8671875
+        ],
+        [
+          198.67526245117188,
+          115.09228515625
+        ],
+        [
+          53.490234375,
+          115.09228515625
+        ]
+      ]
+    },
+    {
+      "title": "A.1. Parameter-Efficient Fine-Tuning",
+      "heading_level": null,
+      "page_id": 10,
+      "polygon": [
+        [
+          53.7890625,
+          123.75
+        ],
+        [
+          213.9947509765625,
+          123.75
+        ],
+        [
+          213.9947509765625,
+          134.00201416015625
+        ],
+        [
+          53.7890625,
+          134.00201416015625
+        ]
+      ]
+    },
+    {
+      "title": "A.2. Routing Mechanisms in Modular Language Models",
+      "heading_level": null,
+      "page_id": 10,
+      "polygon": [
+        [
+          53.7890625,
+          335.197509765625
+        ],
+        [
+          289.518798828125,
+          335.197509765625
+        ],
+        [
+          289.518798828125,
+          345.33984375
+        ],
+        [
+          53.7890625,
+          345.33984375
+        ]
+      ]
+    },
+    {
+      "title": "B. Additional Experiment Details",
+      "heading_level": null,
+      "page_id": 10,
+      "polygon": [
+        [
+          54.38671875,
+          643.899169921875
+        ],
+        [
+          225.47882080078125,
+          643.899169921875
+        ],
+        [
+          225.47882080078125,
+          655.8543701171875
+        ],
+        [
+          54.38671875,
+          655.8543701171875
+        ]
+      ]
+    }
+  ],
+  "page_stats": [
+    {
+      "page_id": 0,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          323
+        ],
+        [
+          "Line",
+          129
+        ],
+        [
+          "Text",
+          6
+        ],
+        [
+          "SectionHeader",
+          4
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Footnote",
+          1
+        ],
+        [
+          "Figure",
+          1
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "FigureGroup",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 1,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          303
+        ],
+        [
+          "Line",
+          140
+        ],
+        [
+          "Text",
+          12
+        ],
+        [
+          "ListItem",
+          3
+        ],
+        [
+          "SectionHeader",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Figure",
+          1
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "FigureGroup",
+          1
+        ],
+        [
+          "ListGroup",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 2,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          395
+        ],
+        [
+          "Line",
+          121
+        ],
+        [
+          "Text",
+          8
+        ],
+        [
+          "SectionHeader",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Figure",
+          1
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "FigureGroup",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 3,
+      "text_extraction_method": "surya",
+      "block_counts": [
+        [
+          "Line",
+          217
+        ],
+        [
+          "Span",
+          156
+        ],
+        [
+          "Text",
+          59
+        ],
+        [
+          "SectionHeader",
+          3
+        ],
+        [
+          "Equation",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 4,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          377
+        ],
+        [
+          "Line",
+          119
+        ],
+        [
+          "TableCell",
+          114
+        ],
+        [
+          "Text",
+          16
+        ],
+        [
+          "PageHeader",
+          2
+        ],
+        [
+          "Caption",
+          2
+        ],
+        [
+          "SectionHeader",
+          2
+        ],
+        [
+          "Table",
+          1
+        ],
+        [
+          "Equation",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "TableGroup",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 5,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          255
+        ],
+        [
+          "Line",
+          107
+        ],
+        [
+          "Text",
+          15
+        ],
+        [
+          "PageHeader",
+          2
+        ],
+        [
+          "Figure",
+          2
+        ],
+        [
+          "Caption",
+          2
+        ],
+        [
+          "SectionHeader",
+          2
+        ],
+        [
+          "FigureGroup",
+          2
+        ],
+        [
+          "Reference",
+          2
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 6,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          348
+        ],
+        [
+          "TableCell",
+          208
+        ],
+        [
+          "Line",
+          119
+        ],
+        [
+          "Text",
+          14
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "Table",
+          1
+        ],
+        [
+          "SectionHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "TableGroup",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 7,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          293
+        ],
+        [
+          "Line",
+          138
+        ],
+        [
+          "TableCell",
+          111
+        ],
+        [
+          "Text",
+          11
+        ],
+        [
+          "Caption",
+          7
+        ],
+        [
+          "Table",
+          3
+        ],
+        [
+          "Reference",
+          3
+        ],
+        [
+          "PageHeader",
+          2
+        ],
+        [
+          "SectionHeader",
+          2
+        ],
+        [
+          "TableGroup",
+          2
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 8,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          339
+        ],
+        [
+          "Line",
+          152
+        ],
+        [
+          "ListItem",
+          19
+        ],
+        [
+          "Reference",
+          19
+        ],
+        [
+          "SectionHeader",
+          2
+        ],
+        [
+          "Text",
+          2
+        ],
+        [
+          "ListGroup",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 9,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          263
+        ],
+        [
+          "Line",
+          111
+        ],
+        [
+          "ListItem",
+          19
+        ],
+        [
+          "Reference",
+          19
+        ],
+        [
+          "Text",
+          12
+        ],
+        [
+          "ListGroup",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 10,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          349
+        ],
+        [
+          "Line",
+          140
+        ],
+        [
+          "Text",
+          14
+        ],
+        [
+          "SectionHeader",
+          6
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 11,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          160
+        ],
+        [
+          "Line",
+          73
+        ],
+        [
+          "TableCell",
+          10
+        ],
+        [
+          "Text",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "Table",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "TableGroup",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    }
+  ],
+  "debug_data_path": "debug_data/31f6f2e8-0fb2-46ff-ab65-f3408612f6e1"
+}

icml26/31f6f2e8-0fb2-46ff-ab65-f3408612f6e1/model_text_v3.txt

@@ -0,0 +1,170 @@
+[p. 1 | section: Abstract | type: Text]
+The dynamic evolution of real-world necessitates model editing within Large Language Models. While existing methods explore modular isolation or parameter-efficient strategies, they still suffer from semantic drift or knowledge forgetting due to continual updating. To address these challenges, we propose SoLA, a Semantic routingbased LoRA framework for lifelong model editing. In SoLA, each edit is encapsulated as an independent LoRA module, which is frozen after training and mapped to input by semantic routing, allowing dynamic activation of LoRA modules via semantic matching. This mechanism avoids semantic drift caused by cluster updating and mitigates catastrophic forgetting from parameter sharing. More importantly, SoLA supports precise revocation of specific edits by removing key from semantic routing, which restores model's original behavior. To our knowledge, this reversible rollback editing capability is the first to be achieved in existing literature. Furthermore, SoLA integrates decision-making process into edited layer, eliminating the need for auxiliary routing networks and enabling end-to-end decision-making process. Extensive experiments demonstrate that SoLA effectively learns and retains edited knowledge, achieving accurate, efficient, and reversible lifelong model editing.
+
+[p. 1 | section: 1. Introduction | type: Text]
+Large language models (LLMs) have demonstrated remarkable capabilities in the field of natural language processing, attracting the attention of numerous researchers (Radford et al., 2019; Brown et al., 2020; Achiam et al., 2023; Tou vron et al., 2023) . However, LLMs also face severe challenges, including hallucinations (Huang et al., 2025) , bi-
+
+[p. 1 | section: 1. Introduction | type: FigureGroup]
+Figure 1. Comparison of trainable parameters and ERR accuracy between SoLA (Ours) and other methods. All experiments are conducted on Scotus datasets with the same backbone.
+
+[p. 1 | section: 1. Introduction | type: Text]
+ases (Hartvigsen et al., 2022) , and the generation of harmful content (De Cao et al., 2021) . Moreover, the dynamic nature of real-world knowledge requires the continuous updating of specific information in LLMs (Lazaridou et al., 2021) . Re-training these models from scratch is both expensive and time-consuming (Biderman et al., 2023) . In this scenario, the demand for model editing is increasing with aims to modify specific knowledge in the training model continuously without re-training the model or affecting its performance on unrelated inputs, which is known as lifelong model editing (Hartvigsen et al., 2022; Yao et al., 2023; Yu et al., 2024) .
+
+[p. 1 | section: 1. Introduction | type: Text]
+The lifelong model editing aims to continuously update the model to adapt to the constantly changing world. However, most of the existing model editing methods are designed for single, isolated editing (Mitchell et al., 2021; Meng et al., 2022a) . When applied to a continuous environment, these methods often lead to catastrophic forgetting of the previously learned knowledge in the LLMs, resulting in performance degradation and reduced model reliability (Yao et al., 2023; Gu et al., 2024) . To solve this problem, recent methods propose freezing the original model parameters and introducing lightweight trainable modules to inject new knowledge. These methods usually retain most of the model's knowledge by keeping the base parameters fixed and combining the learnable modules for specific task updates. For example, Melo (Yu et al., 2024) determines the clustering centres of editors by building a neuron index
+
+[p. 2 | section: 1. Introduction | type: Text]
+and dynamically updates its semantic representation based on the distance from the clustering centres. ELDER (Li et al., 2025) adopts the Mixture-of-Experts (MoE) approach, assigning scores to each LoRA expert through a set of learnable LoRA allocation codes and selecting the top-k LoRA for calculation.
+
+[p. 2 | section: 1. Introduction | type: FigureGroup]
+Figure 2. Number of mismatches in LoRA allocation with ERR/TRR accuracy in MELO. All experiments are conducted on Scotus datasets with the same backbone.
+
+[p. 2 | section: 1. Introduction | type: Text]
+Although these methods show potential in improving parameter efficiency and reducing interference, they face inherent limitations when applied to lifelong model editing. Specifically, MELO (Yu et al., 2024) improves module separability through semantic clustering, but its clustering centers are updated during the editing process, which may lead to semantic drift and module matching errors. ELDER (Li et al., 2025) employ MoE and selects the top-k routes to activate each editing subset, although this strategy has high parameter efficiency, shared and continuously updated parameters can lead to catastrophic forgetting - new editing may overwrite or interfere with existing editing.
+
+[p. 2 | section: 1. Introduction | type: Text]
+To address these limitations, we propose SoLA, a Semantic routing-based LoRA framework for reversible lifelong model editing. In SoLA, we allocate an independent LoRA module for each edit and establish semantic routing to record the mapping relationship between LoRA module and the input semantic representation. During editing, the LoRA module fully learns for the current task and then remains frozen to retain the learned knowledge. The corresponding key is also not updated. During inference, the semantic representation of the input is calculated, and the corresponding LoRA module is dynamically activated through semantic routing. Since the LoRA module and the semantic representation have not been updated, we fundamentally avoid catastrophic forgetting and semantic drift problems caused by continual updating. At the same time, each edit only requires training for the current LoRA module, significantly reducing the computational resource consumption. As shown in Fig .1, SoLA achieves optimal performance with only 0.08M additional parameters, sig-
+
+[p. 2 | section: 1. Introduction | type: Text]
+nificantly outperforming previous methods, highlighting SoLA's exceptional parameter efficiency and editing accuracy.
+
+[p. 2 | section: 1. Introduction | type: Text]
+More importantly, by removing the semantic key from the mapping memory table, we can precisely revoke specific edits, allowing the model to recover its original behavior without re-training. This means we can freely add and delete edits, and truly achieve controllable rollback and restoration of editing. To our knowledge, this is the first time that controllable rollback of editing has been achieved in existing literature. Moreover, existing works usually require introducing auxiliary routing network outside the edited layer to determine whether to enable the LoRA module. In this paper, we propose a master decision-making mechanism that aggregates the decision-making process into the edited layer, avoiding the reliance on auxiliary routing network and achieving an end-to-end decision-making process.
+
+[p. 2 | section: 1. Introduction | type: Text]
+Extensive experiments validate the effectiveness of our approach. Our main contributions are summarized as follows:
+
+[p. 2 | section: 1. Introduction | type: ListGroup]
+We propose SoLA, a novel framework for reversible lifelong model editing, which incorporates a controllable LoRA mechanism guided by semantic routing. After each edit, both the LoRA modules and the corresponding keys in the mapping memory are frozen, effectively mitigating catastrophic forgetting and semantic drift. Moreover, only the currently active LoRA modules are trained during editing, significantly reducing computational overhead.(Fig .1) . We employs semantic routing to establish a precise mapping between LoRA modules and semantic representations. By removing the associated key, any specific edit can be precisely revoked, enabling flexible addition and deletion of edits. we propose a master decision-making mechanism that aggregates the decision-making process into the edited layer, avoiding the reliance on an auxiliary routing network and achieving an end-to-end decision-making process.
+
+[p. 2 | section: 2.1. Model Editing | type: Text]
+Model Editing aims to update specific knowledge within LLMs while preserving their previous knowledge. Current model editing methods can be broadly categorized into three paradigms: Meta-Learning methods, Locate-then-Edit methods, and Memory-Based methods. Meta-Learning Methods employ a hypernetwork to predict the necessary gradients for editing and apply these gradients to the base model to achieve the update (Mitchell et al., 2021) . However, this
+
+[p. 3 | section: 2.1. Model Editing | type: FigureGroup]
+Figure 3. Main framework of our method. a) indicates the edited layer in model, where we retain the transformer block of base model frozen and add trainable LoRA module to the edited layer of base model. b) shows the editing process, where every edit will be assigned LoRA module and the query of input will be mapped to assigned LoRA module. c) presents the matching process between query of input and LoRA module in reference. Colour green indicates trainable, while color grey is frozen.
+
+[p. 3 | section: 2.1. Model Editing | type: Text]
+approach typically requires additional data for training the hypernetwork. Locate-then-Edit methods first identify the neurons most critical for the current input via minor perturbations and then directly modify these neurons to update the model's knowledge (Meng et al., 2022a; b) . While effective, these methods are often cumbersome and exhibit limitations when handling a large volume of simultaneous knowledge updates. Memory-Based methods store the information associated with edits in an buffer, which acts as a patch model augmenting the original model (Mitchell et al., 2022) . A significant drawback is that each edit necessitates retraining, making it difficult to adapt to continuous editing scenarios, and the memory bank grows incrementally with each edit, leading to escalating storage overhead. Recent studies have also explored editing constraints to maintain model behavior stability. AlphaEdit (Fang et al., 2024) introduces a null-space constrained approach. It aims to apply edits more precisely by constraining changes to the null space of irrelevant knowledge.
+
+[p. 3 | section: 2.1. Model Editing | type: Text]
+However, these methods primarily address static editing, suitable for one-time modifications, and struggle to accommodate sequential editing demands over time. To address these limitations, GRACE (Hartvigsen et al., 2023) replace hidden states with vectors retrieved from a learned codebook. MELO (Yu et al., 2024) introduces a vector database and leverages a clustering mechanism to dynamically assign LoRA modules. ELDER (Li et al., 2025) adopts a MoE framework, where a learnable neural network weights shared LoRA modules. However, these approaches in-
+
+[p. 3 | section: 2.1. Model Editing | type: Text]
+evitably update the semantic representation in sequential edits, leading to semantic drift or knowledge forgetting.
+
+[p. 3 | section: 3.1. Preliminaries | type: Text]
+Lifelong Model Editing The purpose of lifelong model editing, as described by Grace (Hartvigsen et al., 2023) , is to make continual edits to the knowledge of the initial base model without degrading the performance of the base model and without negating previous edits. Consider an initial base model fbase, After n edits Dedit = {d1, ..., dn}, some of the model's knowledge is updated and model is transformed into fn, where d i = (x i , yi), For lifetime model editing, the edited model f n should be able to correctly output y within the edited inputs x ∈ Dedit, i.e., fn(xi) = y i . Furthermore, for inputs x ′ ∈/ Dedit that are semantically similar to the edited instances, such as rephrased sentences, the model f n is expected to generalize the updated knowledge appropriately, like fn(x ′ i ) = y i . Moreover, after editing, the model f n should also retain the previous knowledge of fbase, that is, for data samples x j ∈/ Dedit that have not been edited, there should be fn(x j ) = fbase(x j ) = y j .
+
+[p. 3 | section: 3.1. Preliminaries | type: Text]
+LoRA for Lifelong Model Editing Low-Rank Adaptation (LoRA) (Hu et al., 2022) is an efficient finetuning method for LLMs that optimizes model outputs by projecting features into a low-rank subspace. In the context of efficient LLM finetuning, LoRA modules are typically inserted into
+
+[p. 4 | section: 3.1. Preliminaries | type: Text]
+specific layers of the pre-trained model. During editing, the base LLM parameters W_0 \in \mathbb{R}^{d \times k} remain frozen, while only the LoRA module \Delta W \in \mathbb{R}^{d \times k} is updated. The parameter update \Delta W can be represented as a low-rank decomposition of the pretrained weight matrix. Specifically, \Delta W is factorized into two smaller matrices A \in \mathbb{R}^{r \times k} and B \in \mathbb{R}^{d \times r} , such that \Delta W = BA , where r \ll \min(d,k) and r is the LoRA rank. Given the original h = W_0 x , the forward propagation process is modified as:
+
+[p. 4 | section: 3.1. Preliminaries | type: Equation]
+\boldsymbol{h} = W_0 \boldsymbol{x} + \Delta W \boldsymbol{x} = W_0 \boldsymbol{x} + B A \boldsymbol{x} \tag{1}
+
+[p. 4 | section: 3.1. Preliminaries | type: Text]
+where A is initialized using a zero mean Gaussian distribution and B is initialized as a zero matrix. This approach significantly reduces the number of trainable parameters while maintaining the expressive power of the full-rank update.
+
+[p. 4 | section: 3.2. Cluster Updating in Lifelong Model Editing | type: Text]
+In the context of lifelong model editing, when multiple adapter modules are involved, the model needs to dynamically manage the association between edits and the corresponding adapter module. MELO addresses this by employing a clustering mechanism that assigns edits to clusters based on the hidden representations of the inputs, with cluster centres being continuously updated throughout editing. However, this continual updating of cluster centers alters their semantic representation, leading to semantic drift and incorrect module assignments. Moreover, repeated training of LoRA modules can result in forgetting previously acquired knowledge. Varying the cluster radius in MeLO results in a different number of generated cluster centres, thereby affecting the frequency of cluster centre updates. We record MeLO's performance under different numbers of cluster centre updates, as shown in Fig.2. The experiment is conducted on SCOTUS dataset. In Fig.2, the number of updates refers to the number of cluster centre updates during the editing process, and the number of mismatches indicates how many times, during inference, the retrieved LoRA module differs from the one assigned during editing. ERR and TRR represent the model's accuracy on the edited and unedited datasets, respectively, after all editing tasks are completed.
+
+[p. 4 | section: 3.2. Cluster Updating in Lifelong Model Editing | type: Text]
+As shown in the Fig.2, for the same datasets SCOTUS of editing, increasing the number of cluster centre updates leads to a rise in mismatches during inference, and a corresponding decline in model performance on both the edited and unedited datasets. This demonstrates that cluster centre updates can cause semantic drift, resulting in incorrect LoRA match, while repeated training of LoRA modules contributes to knowledge forgetting. Therefore, in SoLA, we freeze both the LoRA modules and their associated keys once the current task is completed. These components are
+
+[p. 4 | section: 3.2. Cluster Updating in Lifelong Model Editing | type: Text]
+no longer updated in subsequent edits, thereby maximizing knowledge retention across editing iterations.
+
+[p. 4 | section: 3.3. Semantic Routing-Based Controllable LoRA | type: Text]
+In the layer where the model needs to be edited, we keep the parameters of the original model frozen and insert the learnable LoRA module for editing, as shown in Fig.3 a). In SoLA, each editing operation is encapsulated within an independent LoRA module. During sequential editing, for a given editing task d_i = (\boldsymbol{x}_i^m, \boldsymbol{y}_i^m)_{m=1}^n , a dedicated LoRA module LoRA<sub>i</sub> is assigned. For each data instance \boldsymbol{x}_i^m \in d_i , the corresponding LoRA id i is recorded, and a semantic routing entry is established by associating LoRA module with a semantic key derived from the input representation, as shown in Fig.3 b). Following prior work(Hartvigsen et al., 2023), we use the hidden representation of the last token in the input sequence as the key e \in \mathbb{R}^d vector. During editing, the assigned LoRA module LoRA<sub>i</sub> will fully learn the task d_i , yielding the output:
+
+[p. 4 | section: 3.3. Semantic Routing-Based Controllable LoRA | type: Equation]
+h = h_0 + LoRA_i(x) (2)
+
+[p. 4 | section: 3.3. Semantic Routing-Based Controllable LoRA | type: Text]
+where h_0 \in \mathbb{R}^{l \times d} is the frozen base model representation, l is the sequence length, d is the hidden state dimension. At this stage, all other LoRA modules and the stored key vectors remain frozen. Upon completion of the editing process, the trained LoRA<sub>i</sub> module is frozen and will be stored with its associated key e_i as a fixed mapping for future reference. Neither the LoRA module nor the key will be updated in subsequent editing. Since only the current LoRA module is involved in training during each edit, SoLA significantly reduces the number of trainable parameters, as shown in Fig.1. Furthermore, freezing both the LoRA module and its corresponding key after editing prevents semantic drift and mitigates catastrophic forgetting, thereby maximizing knowledge retention. During inference, the hidden representation of the last token in the input sequence will serve as a query vector \boldsymbol{q} \in \mathbb{R}^d from the input \boldsymbol{x} and matches against the stored keys K \in \mathbb{R}^{N \times d} , as shown in Fig.3 c). If a match is found, the corresponding LoRA module is retrieved from stored LoRA pool R \in \mathbb{R}^M and incorporated into the computation.
+
+[p. 4 | section: 3.4. Master Decision Mechanism | type: Text]
+During inference, existing approaches usually need to introduce an auxiliary routing network outside the target edited layer to decide whether to activate the LoRA module or not(Yu et al., 2024; Li et al., 2025), which hampers the end-to-end decision-making process. To address this limitation, we propose a master decision-making mechanism that does not require an auxiliary routing component. Specifically, we designate the first edited layer as the master decision layer where input features are dynamically evaluated to activate the relevant LoRA module without the need for an auxil-
+
+[p. 5 | section: 3.4. Master Decision Mechanism | type: TableGroup]
+Table 1. Comparison of SoLA to existing methods. All the results are obtained after sequential edits. ERR indicates Edit Reliability Rate, TRR presents Task Retention Rate, and ARR is Accuracy Attention Rate. The best performance is shown in bold. SCOTUS (BERT; Acc ↑) zsRE (T5; F1 ↑) Hallucination (GPT2-XL; PPL ↓) Method ERR TRR Avg. ERR TRR Avg. ERR TRR ARR EWC 0.51 0.82 0.66 0.67 0.50 0.58 1485.7 29.24 109.59 CMR 0.52 0.52 0.52 0.56 0.82 0.69 1449.3 28.14 107.76 CLEAR 0.67 0.83 0.75 0.27 0.99 0.63 2394.3 35.34 195.82 MEND 0.19 0.27 0.23 0.25 0.27 0.26 1369.8 1754.9 2902.5 SERAC 0.33 0.41 0.37 0.72 0.31 0.52 8183.7 133.3 10.04 ROME - - - - - - 30.28 103.82 14.02 GRACE 0.81 0.82 0.82 0.69 0.96 0.83 15.84 7.14 10.00 ELDER 0.89 0.86 0.88 0.72 0.97 0.84 16.12 5.87 8.42 MELO 0.96 0.92 0.94 0.72 0.98 0.85 17.45 1.04 2.66 SoLA 0.97 0.95 0.96 0.73 0.99 0.86 15.15 1.01 7.35
+
+[p. 5 | section: 3.4. Master Decision Mechanism | type: Text]
+iary network. In this framework, the master decision layer H e computes the distance metric between the input feature embedding q and the stored key K to determine the activation of the module: d = He(argmini(dist(q, ki))), i = (1, 2, ..., N), dist(·) denotes the distance function and each key k i associated with the LoRA module. Then in the master layer He, we decide the edit layer behaviour:
+
+[p. 5 | section: 3.4. Master Decision Mechanism | type: Equation]
+H_e = \begin{cases} H(W_0) & \text{if } d \ge \alpha \\ H(W_0, W_{R_m}) & \text{if } d < \alpha \end{cases} (3)
+
+[p. 5 | section: 3.4. Master Decision Mechanism | type: Text]
+where α is the threshold, in this work we set it as 0.01, WR m represents the weight of m-th LoRA module associated with the key nearest to the query. This binary decision is propagated to the subsequent edited layers, ensuring consistent module activation throughout the network. By integrating the decision-making process to the first edited layer, our approach achieves complete end-to-end decision-making capability while maintaining architectural simplicity.
+
+[p. 5 | section: 4.1. Implementation Details | type: Text]
+Baselines We first compare several recent methods designed for lifelong model editing. Grace (Hartvigsen et al., 2023) replaces model outputs with key-value pairs and matches features to keys based on deferral radius. MELO (Yu et al., 2024) leverages LoRA for finetuning and dynamically retrieves relevant LoRA modules via neuron index. EL-DER (Li et al., 2025) employs MoE to dynamically combine multiple LoRA modules. Additionally, we evaluate other baseline methods. MEND (Mitchell et al., 2021) uses a hypernetwork trained on auxiliary data to predict the required gradients for model editing. SERAC (Mitchell et al., 2022)
+
+[p. 5 | section: 4.1. Implementation Details | type: Text]
+learns a scope classifier and a counterfactual model, coordinating between modules to determine editing behavior. ROME (Meng et al., 2022a) identifies the most influential weights in the model through perturbation analysis, localizes knowledge to specific layers of GPT, and modifies the corresponding weights. Since ROME is specifically designed for GPT, it is only evaluated on the hallucination correction task. EWC (Kirkpatrick et al., 2017) imposes constraints on the weights updating so that the model continuously learns new knowledge and retains previous knowledge. CMR (Lin et al., 2022) performs continuous finetuning of the model to continuously learn new knowledge. CLEAR (Rolnick et al., 2019) builds a memory buffer to restore old tasks information, and replays them in subsequent edits to retain previous knowledge.
+
+[p. 5 | section: 4.1. Implementation Details | type: Text]
+Evaluation Metric In this work, following (Hartvigsen et al., 2023) , we use three basic metrics to assess model performance: ES, ERR and TRR, which are applicable to all model experiments. Where Edit Success(ES) measures the ability of the model to edit the target sample. Edit Reliability Rate(ERR) assesses the validity of the model on the edited dataset, indicating the extent to which the required knowledge updates are successfully incorporated. Task Retention Rate(TRR) quantifies the model's performance on unedited datasets and reflects the model's ability to retain previous knowledge. Additionally, in the hallucination correction task, we also use the Accuracy Attention Rate (ARR) to assess the performance of the already accurate output. Accuracy is measured differently for different tasks. For the document classification task of SCOTUS, we use the Average Accuracy Rate (ACC); for the question answering task of zsRE, we measure it using the average F1, and for the hallucination correction task, we measure it using the standard average complexity (PPL) (Brown et al., 1992) .
+
+[p. 6 | section: 4.1. Implementation Details | type: FigureGroup]
+Figure 4. Figure(a)–(c) presents every edit accuracy on SCOTUS datasets with BERT. All experiment settings are the same as Tab .1.
+
+[p. 6 | section: 4.1. Implementation Details | type: FigureGroup]
+Figure 5. Visualization of zsRE dataset encoder output from T5 small with t-SNE and the dots with same color and shape are input and rephrase sentence. All experiment settings are the same as Tab .1.
+
+[p. 6 | section: 4.2. Main Results | type: Text]
+We evaluate the performance of the proposed SoLA method on three benchmark datasets and compare it with several state-of-the-art model editing approaches. The results are summarized in Tab .1. As observed, SoLA achieves excellent performance on most of the benchmark datasets, with SoLA outperforming the strongest baseline, MELO, by 3% on the SCOTUS dataset. These demonstrate SoLA's enhanced effectiveness in editing target knowledge while better preserving previously acquired knowledge. Furthermore, existing methods such as MEND and SERAC perform poorly under the sequential editing setting, suggesting that they are primarily designed for static editing scenarios and are not well-suited for sequential model editing tasks.
+
+[p. 6 | section: 4.2. Main Results | type: Text]
+We also report task-wise accuracy progression to analyze model performance throughout the sequential editing process. The experimental results are based on the SCOTUS dataset and the results are shown in Fig .4. As illustrated in Fig .4a, Fig .4b and Fig .4c, SoLA consistently maintains
+
+[p. 6 | section: 4.2. Main Results | type: Text]
+superior accuracy across tasks, without exhibiting substantial performance degradation or volatility. This indicates the robustness and stability of SoLA when applied to sequential knowledge editing.
+
+[p. 6 | section: 4.2. Main Results | type: Text]
+Method performance may vary across different model sizes and datasets. To further evaluate the robustness of our approach, we conduct hallucination correction experiments on the UniEdit (Chen et al., 2025) and WikiBigEdit (Thede et al., 2025) datasets using larger-scale backbone models. The results are presented in Tab .2. As shown in Tab .2, SoLA consistently achieves the best or near-best performance across most settings, demonstrating its robustness and effectiveness under diverse task configurations. During editing, larger backbone models generally exhibit more stable ERR values, suggesting that such models have learned stronger semantic representations during pretraining, leading to better generalization and editability. Furthermore, continuous learning related methods tend to overfit in edited data, thus showing better ARR in the retention set, but large TRR values in the upstream set, indicating severe forgetting of previous knowledge.
+
+[p. 6 | section: 4.3. Controllable Rollback Editing | type: Text]
+In SoLA, each edit instance is associated with a unique key stored in both the memory and the routing table. This design enables reversible editing, as the model can revert to its original behavior by removing the key corresponding to a specific edit. To demonstrate this property, we conduct an illustrative experiment on the zsRE dataset. The results are shown in Tab .3. In the Tab .3, each row corresponds to an input instance consisting of a "Text" (the question need to be edited) and its "Labels" (the correct answer). "Del" indicates whether the key corresponding to the edit is removed after editing. For each input, Predbase, Prededit and Preddel denote the prediction of base model, edited model and model after removing the corresponding key. When "Del" is false, the key is retained, serving as a baseline comparison. As shown in Tab .3, for each instance, the original prediction (Predbase) does not match the correct label, indi-
+
+[p. 7 | section: 4.3. Controllable Rollback Editing | type: TableGroup]
+Table 2. Model Performance Comparison on UniEdit and WikiBigEdit datasets in hallucination correction tasks. All the scores are PPL metric and the best result is in bolded. UniEdit (PPL ↓) WikiBigEdit (PPL ↓) Model Method ERR TRR ARR ERR TRR ARR EWC 2.43 1134 22.08 3.58 251117 4.93 CMR 2.79 2239 25.24 2.86 169452 4.49 LLaMA-3-8B CLEAR 1.04 982 29.79 1.03 74384 1.52 SERAC 37.10 246 93 60.17 215 59 GRACE 1.00 244 89 1.00 216 58 ELDER 1.00 173 74 1.00 178 48 MELO 1.00 153 68 1.00 165 47 Ours 1.00 144 66 1.00 162 44 EWC 2.38 4451 42.13 2.86 47846 3.89 CMR 2.34 7211 36.19 3.60 34623 5.51 DeepSeek-R1-8B CLEAR 1.06 3365 44.86 1.02 28110 1.48 SERAC 46.76 451 834 82.72 493 831 GRACE 1.00 407 803 1.00 487 793 ELDER 1.03 241 482 1.03 304 492 MELO 1.09 204 407 1.05 263 445 Ours 1.07 188 374 1.00 248 412 EWC 4.62 7648 48.62 5.61 61834 14.65 CMR 4.12 11583 40.56 6.28 48629 16.89 Qwen2-7B CLEAR 1.03 5021 54.98 1.05 35396 4.26 SERAC 58.51 329 1261 62.12 304 523 GRACE 2.05 316 1065 16.88 299 518 ELDER 1.00 201 266 1.00 225 197 MELO 1.00 170 201 1.00 203 143 Ours 1.00 156 110 1.00 199 109
+
+[p. 7 | section: 4.3. Controllable Rollback Editing | type: Text]
+cating the necessity for editing. After editing, the model correctly produces Prededit consistent with the ground-truth label, demonstrating that SoLA successfully updates the model's knowledge. When the associated key is deleted, the model reverts to its original prediction (Prededit = Predbase), verifying that our method can effectively roll back specific edits. More critically, for edits where the key is not deleted, the model continues to output Prededit, showing that the rollback operation does not interfere with other edits. This confirms that SoLA supports fine-grained, selective undo of knowledge edits without disrupting unrelated modifications.
+
+[p. 7 | section: 4.4. Ablation Study | type: Text]
+Effect of Edit Layer Location In semantic representation learning, different layers of a model focus on different aspects of semantic information. Therefore, the position of the edited layer can significantly influence the effectiveness of model editing. To investigate this, we evaluate the performance of model when edits are applied at different layers, while keeping all other settings fixed. The experiments are conducted on the SCOTUS dataset with a Nvidia A100 40G GPU, and the results are shown in Tab .4. In the
+
+[p. 7 | section: 4.4. Ablation Study | type: Text]
+Tab .4, "Layer" denotes the layer range used for editing. For example, "0–2" indicates that edits are applied to layers 0, 1, and 2 of the BERT model. As shown in the table, editing at shallower layers results in suboptimal performance. This observation aligns with prior findings (Geva et al., 2020) , which suggest that shallow layers primarily capture the shallow sentence patterns, whereas deeper layers encode richer semantic features, thereby enabling more effective knowledge editing. Moreover, editing at shallow layers significantly increases the time used for editing, which means that editing for semantics is more difficult in shallow layers. Furthermore, we observe that varying the editing layer has negligible impact on performance over the unedited portion of the dataset, indicating that SoLA is capable of preserving previously learned knowledge regardless of the editing depth, which further supports the robustness of SoLA.
+
+[p. 7 | section: 4.4. Ablation Study | type: Text]
+Effect of LoRA Rank The LoRA module projects semantic representations into a low-rank subspace for parameterefficient learning. Consequently, the choice of rank will effect the module's performance. To investigate this, we analyze the impact of different LoRA rank values on model performance, keeping all other settings fixed. Experiments are conducted on the SCOTUS dataset using a NVIDIA
+
+[p. 8 | section: 4.4. Ablation Study | type: TableGroup]
+Table 3. Controllable edit in SoLA on zsRE datasets. "Text" is the question need to be edited, "Labels" is the true answer, "Del" indicates whther delete associated key and Predbase, Prededit and Preddel is the prediction of base model, edited model and delted model. Text Labels Del Predbase Prededit Preddel The date of Tsetserleg earthquake in 1905? 9 July 1905 True 20 February 1905 9 July 1905 20 February 1905 What constellation has Nu Cancri? Cancer True Hydra Cancer Hydra On what date did the Bahrain Grand Prix 2015 occur? 19 April 2015 True 6 April 2017 19 April 2015 6 April 2017 What position does Charles-Joseph Coursol have? Mayor of Montreal True Positioning Mayor of Montreal Positioning Who acted in Blow Dry? Alan Rickman False Jim Carrey Alan Rickman Alan Rickman Table 4. Ablation studies on location of edited layers. Here, "Layer" indicates the edited layer location. For example, "0-2" presents edited layer is layers 0, 1 and 2.
+
+[p. 8 | section: 4.4. Ablation Study | type: TableGroup]
+Layer ES ERR TRR Edit Time(min) 0-2 0.77 0.61 0.96 9.21 3-5 1.00 0.68 0.96 8.06 6-8 0.81 0.70 0.97 7.03 9-11 0.94 0.95 0.97 5.97 Table 5. Ablation studies on LoRA rank. Here, "LoRA Rank" indicates the rank value of every LoRA in model.
+
+[p. 8 | section: 4.4. Ablation Study | type: Table]
+LoRA Rank ES ERR TRR Edit Time(min) 1 0.90 0.84 0.96 5.88 2 0.94 0.91 0.97 5.90 3 1.00 0.91 0.97 5.86 4 1.00 0.95 0.97 5.97 5 1.00 0.92 0.96 5.86 10 1.00 0.71 0.96 5.85
+
+[p. 8 | section: 4.4. Ablation Study | type: Text]
+A100 40GB GPU, and the results are presented in Tab .5. As shown in the Tab .5, simply increasing the LoRA rank does not lead to improved performance and may even degrade model performance. This indicates that expanding the learnable parameter space does not necessarily enhance the model's capacity for effective knowledge editing. On the contrary, excessive rank values may even result in performance degradation due to overfitting, suggesting the importance of carefully selecting an appropriate rank to balance capacity and generalization. Based on the experiment results, we set 4 as LoRA rank for all the experiments in this paper.
+
+[p. 8 | section: 4.5. Model Visualization | type: Text]
+To gain deeper insight into the model's behavior under sequential editing, we conducted a t-SNE visualization (Maaten & Hinton, 2008) of the learned feature representations. This experiment is performed on the zsRE dataset using the T5-small model to assess how semantic relationships are preserved after sequential edits. Upon completing
+
+[p. 8 | section: 4.5. Model Visualization | type: Text]
+the entire sequence of editing tasks for question answering, we selected five distinct input instances, each accompanied by semantically equivalent rephrased sentences, forming five evaluation groups. The encoder output features are subsequently extracted from the edited model for dimensionality reduction and visualization, and the resulting plot is represented in Fig .5. As shown in Fig .5, there is a clear clustering pattern that for each group, the original input and its rephrased variant are mapped to proximate locations in the latent space. This demonstrates that SoLA effectively captures and preserves the semantic similarity between related inputs throughout the editing process. Furthermore, distinct groups maintain separable clusters, indicating that the model retains the ability to discriminate between different edit concepts after sequential editing. This visualization shows that SoLA's semantic routing mechanism successfully supports edit generalization based on semantic similarity while upholding the integrity of previously learned knowledge.
+
+[p. 8 | section: 5. Conclusion | type: Text]
+In this paper, we propose SoLA, a Semantic routing-based LoRA framework for reversible lifelong model editing. In SoLA, each edit is encapsulated as an independent LoRA module which is frozen after training and a semantic routing record is established to map LoRA module to the input semantic representation, allowing dynamic activation of LoRA modules via semantic matching. This design not only avoids semantic drift caused by clustering updating but also mitigates catastrophic forgetting from parameter sharing. More importantly, SoLA supports precisely revoking edit by removing the corresponding key from the routing table, enabling reversible model editing. This allows for flexible addition and deletion of edits, offering fine-grained control over model behavior. To the best of our knowledge, this is the first work to achieve reversible model editing in the literature. Furthermore, we introduce a master decisionmaking mechanism by integrating decision-making into the edited layer, enabling end-to-end decision-making process. By building a strict mapping between edits and LoRA modules, SoLA achieves efficient and reversible lifelong model editing, providing a novel perspective for future research.
+
+[p. 9 | section: Impact Statement | type: Text]
+This work advances the field of Machine Learning by introducing SoLA, a framework for reversible and controllable lifelong model editing in large language models. Our method enhances the safety and reliability of AI systems by enabling precise rollback of model edits, which helps mitigate risks associated with erroneous or harmful knowledge updates. By effectively preventing semantic drift and catastrophic forgetting, SoLA promotes the development of more robust and trustworthy models capable of adapting to evolving information. The significant reduction in computational overhead also aligns with the goal of sustainable AI development.

icml26/31f6f2e8-0fb2-46ff-ab65-f3408612f6e1/paper.md

@@ -0,0 +1,521 @@
+
+
+{0}------------------------------------------------
+
+## Reversible Lifelong Model Editing via Semantic Routing-Based LoRA
+
+## Anonymous Authors<sup>1</sup>
+
+## Abstract
+
+The dynamic evolution of real-world necessitates model editing within Large Language Models. While existing methods explore modular isolation or parameter-efficient strategies, they still suffer from semantic drift or knowledge forgetting due to continual updating. To address these challenges, we propose SoLA, a Semantic routingbased LoRA framework for lifelong model editing. In SoLA, each edit is encapsulated as an independent LoRA module, which is frozen after training and mapped to input by semantic routing, allowing dynamic activation of LoRA modules via semantic matching. This mechanism avoids semantic drift caused by cluster updating and mitigates catastrophic forgetting from parameter sharing. More importantly, SoLA supports precise revocation of specific edits by removing key from semantic routing, which restores model's original behavior. To our knowledge, this reversible rollback editing capability is the first to be achieved in existing literature. Furthermore, SoLA integrates decision-making process into edited layer, eliminating the need for auxiliary routing networks and enabling end-to-end decision-making process. Extensive experiments demonstrate that SoLA effectively learns and retains edited knowledge, achieving accurate, efficient, and reversible lifelong model editing.
+
+## 1. Introduction
+
+Large language models (LLMs) have demonstrated remarkable capabilities in the field of natural language processing, attracting the attention of numerous researchers[\(Radford](#page-9-0) [et al.,](#page-9-0) [2019;](#page-9-0) [Brown et al.,](#page-8-0) [2020;](#page-8-0) [Achiam et al.,](#page-8-1) [2023;](#page-8-1) [Tou](#page-9-1)[vron et al.,](#page-9-1) [2023\)](#page-9-1). However, LLMs also face severe challenges, including hallucinations[\(Huang et al.,](#page-8-2) [2025\)](#page-8-2), bi-
+
+Preliminary work. Under review by the International Conference on Machine Learning (ICML). Do not distribute.
+
+<span id="page-0-0"></span>![](_page_0_Figure_9.jpeg)
+
+*Figure 1.* Comparison of trainable parameters and ERR accuracy between SoLA (Ours) and other methods. All experiments are conducted on Scotus datasets with the same backbone.
+
+ases[\(Hartvigsen et al.,](#page-8-3) [2022\)](#page-8-3), and the generation of harmful content[\(De Cao et al.,](#page-8-4) [2021\)](#page-8-4). Moreover, the dynamic nature of real-world knowledge requires the continuous updating of specific information in LLMs[\(Lazaridou et al.,](#page-8-5) [2021\)](#page-8-5). Re-training these models from scratch is both expensive and time-consuming[\(Biderman et al.,](#page-8-6) [2023\)](#page-8-6). In this scenario, the demand for model editing is increasing with aims to modify specific knowledge in the training model continuously without re-training the model or affecting its performance on unrelated inputs, which is known as lifelong model editing[\(Hartvigsen et al.,](#page-8-3) [2022;](#page-8-3) [Yao et al.,](#page-9-2) [2023;](#page-9-2) [Yu et al.,](#page-9-3) [2024\)](#page-9-3).
+
+The lifelong model editing aims to continuously update the model to adapt to the constantly changing world. However, most of the existing model editing methods are designed for single, isolated editing[\(Mitchell et al.,](#page-9-4) [2021;](#page-9-4) [Meng et al.,](#page-9-5) [2022a\)](#page-9-5). When applied to a continuous environment, these methods often lead to catastrophic forgetting of the previously learned knowledge in the LLMs, resulting in performance degradation and reduced model reliability[\(Yao](#page-9-2) [et al.,](#page-9-2) [2023;](#page-9-2) [Gu et al.,](#page-8-7) [2024\)](#page-8-7). To solve this problem, recent methods propose freezing the original model parameters and introducing lightweight trainable modules to inject new knowledge. These methods usually retain most of the model's knowledge by keeping the base parameters fixed and combining the learnable modules for specific task updates. For example, Melo[\(Yu et al.,](#page-9-3) [2024\)](#page-9-3) determines the clustering centres of editors by building a neuron index
+
+<sup>1</sup>Anonymous Institution, Anonymous City, Anonymous Region, Anonymous Country. Correspondence to: Anonymous Author <anon.email@domain.com>.
+
+{1}------------------------------------------------
+
+104
+
+106
+
+108 109 and dynamically updates its semantic representation based on the distance from the clustering centres. ELDER[\(Li](#page-9-6) [et al.,](#page-9-6) [2025\)](#page-9-6) adopts the Mixture-of-Experts (MoE) approach, assigning scores to each LoRA expert through a set of learnable LoRA allocation codes and selecting the top-k LoRA for calculation.
+
+<span id="page-1-0"></span>![](_page_1_Figure_3.jpeg)
+
+*Figure 2.* Number of mismatches in LoRA allocation with ERR/TRR accuracy in MELO. All experiments are conducted on Scotus datasets with the same backbone.
+
+Although these methods show potential in improving parameter efficiency and reducing interference, they face inherent limitations when applied to lifelong model editing. Specifically, MELO[\(Yu et al.,](#page-9-3) [2024\)](#page-9-3) improves module separability through semantic clustering, but its clustering centers are updated during the editing process, which may lead to semantic drift and module matching errors. ELDER[\(Li et al.,](#page-9-6) [2025\)](#page-9-6) employ MoE and selects the top-k routes to activate each editing subset, although this strategy has high parameter efficiency, shared and continuously updated parameters can lead to catastrophic forgetting - new editing may overwrite or interfere with existing editing.
+
+To address these limitations, we propose SoLA, a Semantic routing-based LoRA framework for reversible lifelong model editing. In SoLA, we allocate an independent LoRA module for each edit and establish semantic routing to record the mapping relationship between LoRA module and the input semantic representation. During editing, the LoRA module fully learns for the current task and then remains frozen to retain the learned knowledge. The corresponding key is also not updated. During inference, the semantic representation of the input is calculated, and the corresponding LoRA module is dynamically activated through semantic routing. Since the LoRA module and the semantic representation have not been updated, we fundamentally avoid catastrophic forgetting and semantic drift problems caused by continual updating. At the same time, each edit only requires training for the current LoRA module, significantly reducing the computational resource consumption. As shown in Fig[.1,](#page-0-0) SoLA achieves optimal performance with only 0.08M additional parameters, significantly outperforming previous methods, highlighting SoLA's exceptional parameter efficiency and editing accuracy.
+
+More importantly, by removing the semantic key from the mapping memory table, we can precisely revoke specific edits, allowing the model to recover its original behavior without re-training. This means we can freely add and delete edits, and truly achieve controllable rollback and restoration of editing. To our knowledge, this is the first time that controllable rollback of editing has been achieved in existing literature. Moreover, existing works usually require introducing auxiliary routing network outside the edited layer to determine whether to enable the LoRA module. In this paper, we propose a master decision-making mechanism that aggregates the decision-making process into the edited layer, avoiding the reliance on auxiliary routing network and achieving an end-to-end decision-making process.
+
+Extensive experiments validate the effectiveness of our approach. Our main contributions are summarized as follows:
+
+- We propose SoLA, a novel framework for reversible lifelong model editing, which incorporates a controllable LoRA mechanism guided by semantic routing. After each edit, both the LoRA modules and the corresponding keys in the mapping memory are frozen, effectively mitigating catastrophic forgetting and semantic drift. Moreover, only the currently active LoRA modules are trained during editing, significantly reducing computational overhead.(Fig[.1\)](#page-0-0).
+- We employs semantic routing to establish a precise mapping between LoRA modules and semantic representations. By removing the associated key, any specific edit can be precisely revoked, enabling flexible addition and deletion of edits.
+- we propose a master decision-making mechanism that aggregates the decision-making process into the edited layer, avoiding the reliance on an auxiliary routing network and achieving an end-to-end decision-making process.
+
+## 2. Related Work
+
+#### 2.1. Model Editing
+
+Model Editing aims to update specific knowledge within LLMs while preserving their previous knowledge. Current model editing methods can be broadly categorized into three paradigms: Meta-Learning methods, Locate-then-Edit methods, and Memory-Based methods. Meta-Learning Methods employ a hypernetwork to predict the necessary gradients for editing and apply these gradients to the base model to achieve the update[\(Mitchell et al.,](#page-9-4) [2021\)](#page-9-4). However, this
+
+{2}------------------------------------------------
+
+<span id="page-2-0"></span>![](_page_2_Figure_1.jpeg)
+
+*Figure 3.* Main framework of our method. a) indicates the edited layer in model, where we retain the transformer block of base model frozen and add trainable LoRA module to the edited layer of base model. b) shows the editing process, where every edit will be assigned LoRA module and the query of input will be mapped to assigned LoRA module. c) presents the matching process between query of input and LoRA module in reference. Colour green indicates trainable, while color grey is frozen.
+
+approach typically requires additional data for training the hypernetwork. Locate-then-Edit methods first identify the neurons most critical for the current input via minor perturbations and then directly modify these neurons to update the model's knowledge[\(Meng et al.,](#page-9-5) [2022a;](#page-9-5)[b\)](#page-9-7). While effective, these methods are often cumbersome and exhibit limitations when handling a large volume of simultaneous knowledge updates. Memory-Based methods store the information associated with edits in an buffer, which acts as a patch model augmenting the original model[\(Mitchell et al.,](#page-9-8) [2022\)](#page-9-8). A significant drawback is that each edit necessitates retraining, making it difficult to adapt to continuous editing scenarios, and the memory bank grows incrementally with each edit, leading to escalating storage overhead. Recent studies have also explored editing constraints to maintain model behavior stability. AlphaEdit[\(Fang et al.,](#page-8-8) [2024\)](#page-8-8) introduces a null-space constrained approach. It aims to apply edits more precisely by constraining changes to the null space of irrelevant knowledge.
+
+153 154
+
+However, these methods primarily address static editing, suitable for one-time modifications, and struggle to accommodate sequential editing demands over time. To address these limitations, GRACE[\(Hartvigsen et al.,](#page-8-9) [2023\)](#page-8-9) replace hidden states with vectors retrieved from a learned codebook. MELO[\(Yu et al.,](#page-9-3) [2024\)](#page-9-3) introduces a vector database and leverages a clustering mechanism to dynamically assign LoRA modules. ELDER[\(Li et al.,](#page-9-6) [2025\)](#page-9-6) adopts a MoE framework, where a learnable neural network weights shared LoRA modules. However, these approaches inevitably update the semantic representation in sequential edits, leading to semantic drift or knowledge forgetting.
+
+## 3. Method
+
+## 3.1. Preliminaries
+
+Lifelong Model Editing The purpose of lifelong model editing, as described by Grace[\(Hartvigsen et al.,](#page-8-9) [2023\)](#page-8-9), is to make continual edits to the knowledge of the initial base model without degrading the performance of the base model and without negating previous edits. Consider an initial base model fbase, After n edits Dedit = {d1, ..., dn}, some of the model's knowledge is updated and model is transformed into fn, where d<sup>i</sup> = (x<sup>i</sup> , yi), For lifetime model editing, the edited model f<sup>n</sup> should be able to correctly output y within the edited inputs x ∈ Dedit, i.e., fn(xi) = y<sup>i</sup> . Furthermore, for inputs x ′ ∈/ Dedit that are semantically similar to the edited instances, such as rephrased sentences, the model f<sup>n</sup> is expected to generalize the updated knowledge appropriately, like fn(x ′ i ) = y<sup>i</sup> . Moreover, after editing, the model f<sup>n</sup> should also retain the previous knowledge of fbase, that is, for data samples x<sup>j</sup> ∈/ Dedit that have not been edited, there should be fn(x<sup>j</sup> ) = fbase(x<sup>j</sup> ) = y<sup>j</sup> .
+
+LoRA for Lifelong Model Editing Low-Rank Adaptation (LoRA) [\(Hu et al.,](#page-8-10) [2022\)](#page-8-10) is an efficient finetuning method for LLMs that optimizes model outputs by projecting features into a low-rank subspace. In the context of efficient LLM finetuning, LoRA modules are typically inserted into
+
+{3}------------------------------------------------
+
+specific layers of the pre-trained model. During editing, the base LLM parameters  $W_0 \in \mathbb{R}^{d \times k}$  remain frozen, while only the LoRA module  $\Delta W \in \mathbb{R}^{d \times k}$  is updated. The parameter update  $\Delta W$  can be represented as a low-rank decomposition of the pretrained weight matrix. Specifically,  $\Delta W$  is factorized into two smaller matrices  $A \in \mathbb{R}^{r \times k}$  and  $B \in \mathbb{R}^{d \times r}$ , such that  $\Delta W = BA$ , where  $r \ll \min(d,k)$  and r is the LoRA rank. Given the original  $h = W_0 x$ , the forward propagation process is modified as:
+
+165
+
+166
+
+167
+
+168
+
+169
+
+170
+
+171
+
+172
+
+173
+
+174
+
+175
+
+176
+
+177
+
+178
+
+179
+
+180
+
+181
+
+182
+
+183
+
+184 185
+
+186
+
+187
+
+188
+
+189
+
+190
+
+191
+
+192
+
+193
+
+194
+
+195
+
+196
+
+197
+
+198
+
+199
+
+200
+
+202
+
+204
+
+205
+
+206
+
+208
+
+209
+
+210
+
+211
+
+212
+
+213
+
+214
+
+215
+
+216
+
+217
+
+218
+
+219
+
+$$\boldsymbol{h} = W_0 \boldsymbol{x} + \Delta W \boldsymbol{x} = W_0 \boldsymbol{x} + B A \boldsymbol{x} \tag{1}$$
+
+where A is initialized using a zero mean Gaussian distribution and B is initialized as a zero matrix. This approach significantly reduces the number of trainable parameters while maintaining the expressive power of the full-rank update.
+
+#### 3.2. Cluster Updating in Lifelong Model Editing
+
+In the context of lifelong model editing, when multiple adapter modules are involved, the model needs to dynamically manage the association between edits and the corresponding adapter module. MELO addresses this by employing a clustering mechanism that assigns edits to clusters based on the hidden representations of the inputs, with cluster centres being continuously updated throughout editing. However, this continual updating of cluster centers alters their semantic representation, leading to semantic drift and incorrect module assignments. Moreover, repeated training of LoRA modules can result in forgetting previously acquired knowledge. Varying the cluster radius in MeLO results in a different number of generated cluster centres, thereby affecting the frequency of cluster centre updates. We record MeLO's performance under different numbers of cluster centre updates, as shown in Fig.2. The experiment is conducted on SCOTUS dataset. In Fig.2, the number of updates refers to the number of cluster centre updates during the editing process, and the number of mismatches indicates how many times, during inference, the retrieved LoRA module differs from the one assigned during editing. ERR and TRR represent the model's accuracy on the edited and unedited datasets, respectively, after all editing tasks are completed.
+
+As shown in the Fig.2, for the same datasets SCOTUS of editing, increasing the number of cluster centre updates leads to a rise in mismatches during inference, and a corresponding decline in model performance on both the edited and unedited datasets. This demonstrates that cluster centre updates can cause semantic drift, resulting in incorrect LoRA match, while repeated training of LoRA modules contributes to knowledge forgetting. Therefore, in SoLA, we freeze both the LoRA modules and their associated keys once the current task is completed. These components are
+
+no longer updated in subsequent edits, thereby maximizing knowledge retention across editing iterations.
+
+#### 3.3. Semantic Routing-Based Controllable LoRA
+
+In the layer where the model needs to be edited, we keep the parameters of the original model frozen and insert the learnable LoRA module for editing, as shown in Fig.3 a). In SoLA, each editing operation is encapsulated within an independent LoRA module. During sequential editing, for a given editing task  $d_i = (\boldsymbol{x}_i^m, \boldsymbol{y}_i^m)_{m=1}^n$ , a dedicated LoRA module LoRA<sub>i</sub> is assigned. For each data instance  $\boldsymbol{x}_i^m \in d_i$ , the corresponding LoRA id i is recorded, and a semantic routing entry is established by associating LoRA module with a semantic key derived from the input representation, as shown in Fig.3 b). Following prior work(Hartvigsen et al., 2023), we use the hidden representation of the last token in the input sequence as the key  $e \in \mathbb{R}^d$  vector. During editing, the assigned LoRA module LoRA<sub>i</sub> will fully learn the task  $d_i$ , yielding the output:
+
+$$h = h_0 + LoRA_i(x)$$
+ (2)
+
+where  $h_0 \in \mathbb{R}^{l \times d}$  is the frozen base model representation, l is the sequence length, d is the hidden state dimension. At this stage, all other LoRA modules and the stored key vectors remain frozen. Upon completion of the editing process, the trained LoRA<sub>i</sub> module is frozen and will be stored with its associated key  $e_i$  as a fixed mapping for future reference. Neither the LoRA module nor the key will be updated in subsequent editing. Since only the current LoRA module is involved in training during each edit, SoLA significantly reduces the number of trainable parameters, as shown in Fig.1. Furthermore, freezing both the LoRA module and its corresponding key after editing prevents semantic drift and mitigates catastrophic forgetting, thereby maximizing knowledge retention. During inference, the hidden representation of the last token in the input sequence will serve as a query vector  $\boldsymbol{q} \in \mathbb{R}^d$  from the input  $\boldsymbol{x}$  and matches against the stored keys  $K \in \mathbb{R}^{N \times d}$ , as shown in Fig.3 c). If a match is found, the corresponding LoRA module is retrieved from stored LoRA pool  $R \in \mathbb{R}^M$  and incorporated into the computation.
+
+#### 3.4. Master Decision Mechanism
+
+During inference, existing approaches usually need to introduce an auxiliary routing network outside the target edited layer to decide whether to activate the LoRA module or not(Yu et al., 2024; Li et al., 2025), which hampers the end-to-end decision-making process. To address this limitation, we propose a master decision-making mechanism that does not require an auxiliary routing component. Specifically, we designate the first edited layer as the master decision layer where input features are dynamically evaluated to activate the relevant LoRA module without the need for an auxil-
+
+{4}------------------------------------------------
+
+228 229 230
+
+233 234
+
+236
+
+254
+
+256
+
+264
+
+274
+
+220
+
+<span id="page-4-0"></span>*Table 1.* Comparison of SoLA to existing methods. All the results are obtained after sequential edits. ERR indicates Edit Reliability Rate, TRR presents Task Retention Rate, and ARR is Accuracy Attention Rate. The best performance is shown in bold.
+
+|        | SCOTUS (BERT; Acc<br>↑) |      |      | zsRE (T5; F1<br>↑) |      |      | Hallucination (GPT2-XL; PPL<br>↓) |        |        |
+|--------|-------------------------|------|------|--------------------|------|------|-----------------------------------|--------|--------|
+| Method | ERR                     | TRR  | Avg. | ERR                | TRR  | Avg. | ERR                               | TRR    | ARR    |
+| EWC    | 0.51                    | 0.82 | 0.66 | 0.67               | 0.50 | 0.58 | 1485.7                            | 29.24  | 109.59 |
+| CMR    | 0.52                    | 0.52 | 0.52 | 0.56               | 0.82 | 0.69 | 1449.3                            | 28.14  | 107.76 |
+| CLEAR  | 0.67                    | 0.83 | 0.75 | 0.27               | 0.99 | 0.63 | 2394.3                            | 35.34  | 195.82 |
+| MEND   | 0.19                    | 0.27 | 0.23 | 0.25               | 0.27 | 0.26 | 1369.8                            | 1754.9 | 2902.5 |
+| SERAC  | 0.33                    | 0.41 | 0.37 | 0.72               | 0.31 | 0.52 | 8183.7                            | 133.3  | 10.04  |
+| ROME   | -                       | -    | -    | -                  | -    | -    | 30.28                             | 103.82 | 14.02  |
+| GRACE  | 0.81                    | 0.82 | 0.82 | 0.69               | 0.96 | 0.83 | 15.84                             | 7.14   | 10.00  |
+| ELDER  | 0.89                    | 0.86 | 0.88 | 0.72               | 0.97 | 0.84 | 16.12                             | 5.87   | 8.42   |
+| MELO   | 0.96                    | 0.92 | 0.94 | 0.72               | 0.98 | 0.85 | 17.45                             | 1.04   | 2.66   |
+| SoLA   | 0.97                    | 0.95 | 0.96 | 0.73               | 0.99 | 0.86 | 15.15                             | 1.01   | 7.35   |
+
+iary network. In this framework, the master decision layer H<sup>e</sup> computes the distance metric between the input feature embedding q and the stored key K to determine the activation of the module: d = He(argmini(dist(q, ki))), i = (1, 2, ..., N), dist(·) denotes the distance function and each key k<sup>i</sup> associated with the LoRA module. Then in the master layer He, we decide the edit layer behaviour:
+
+$$H_e = \begin{cases} H(W_0) & \text{if } d \ge \alpha \\ H(W_0, W_{R_m}) & \text{if } d < \alpha \end{cases}$$
+ (3)
+
+where α is the threshold, in this work we set it as 0.01, WR<sup>m</sup> represents the weight of m-th LoRA module associated with the key nearest to the query. This binary decision is propagated to the subsequent edited layers, ensuring consistent module activation throughout the network. By integrating the decision-making process to the first edited layer, our approach achieves complete end-to-end decision-making capability while maintaining architectural simplicity.
+
+## 4. Experiments
+
+#### 4.1. Implementation Details
+
+Baselines We first compare several recent methods designed for lifelong model editing. Grace[\(Hartvigsen et al.,](#page-8-9) [2023\)](#page-8-9) replaces model outputs with key-value pairs and matches features to keys based on deferral radius. MELO[\(Yu et al.,](#page-9-3) [2024\)](#page-9-3) leverages LoRA for finetuning and dynamically retrieves relevant LoRA modules via neuron index. EL-DER[\(Li et al.,](#page-9-6) [2025\)](#page-9-6) employs MoE to dynamically combine multiple LoRA modules. Additionally, we evaluate other baseline methods. MEND[\(Mitchell et al.,](#page-9-4) [2021\)](#page-9-4) uses a hypernetwork trained on auxiliary data to predict the required gradients for model editing. SERAC[\(Mitchell et al.,](#page-9-8) [2022\)](#page-9-8)
+
+learns a scope classifier and a counterfactual model, coordinating between modules to determine editing behavior. ROME[\(Meng et al.,](#page-9-5) [2022a\)](#page-9-5) identifies the most influential weights in the model through perturbation analysis, localizes knowledge to specific layers of GPT, and modifies the corresponding weights. Since ROME is specifically designed for GPT, it is only evaluated on the hallucination correction task. EWC[\(Kirkpatrick et al.,](#page-8-11) [2017\)](#page-8-11) imposes constraints on the weights updating so that the model continuously learns new knowledge and retains previous knowledge. CMR[\(Lin](#page-9-9) [et al.,](#page-9-9) [2022\)](#page-9-9) performs continuous finetuning of the model to continuously learn new knowledge. CLEAR[\(Rolnick et al.,](#page-9-10) [2019\)](#page-9-10) builds a memory buffer to restore old tasks information, and replays them in subsequent edits to retain previous knowledge.
+
+Evaluation Metric In this work, following[\(Hartvigsen et al.,](#page-8-9) [2023\)](#page-8-9), we use three basic metrics to assess model performance: ES, ERR and TRR, which are applicable to all model experiments. Where Edit Success(ES) measures the ability of the model to edit the target sample. Edit Reliability Rate(ERR) assesses the validity of the model on the edited dataset, indicating the extent to which the required knowledge updates are successfully incorporated. Task Retention Rate(TRR) quantifies the model's performance on unedited datasets and reflects the model's ability to retain previous knowledge. Additionally, in the hallucination correction task, we also use the Accuracy Attention Rate (ARR) to assess the performance of the already accurate output. Accuracy is measured differently for different tasks. For the document classification task of SCOTUS, we use the Average Accuracy Rate (ACC); for the question answering task of zsRE, we measure it using the average F1, and for the hallucination correction task, we measure it using the standard average complexity (PPL)[\(Brown et al.,](#page-8-12) [1992\)](#page-8-12).
+
+{5}------------------------------------------------
+
+294
+
+296
+
+313 314
+
+316
+
+324
+
+326
+
+328 329
+
+<span id="page-5-0"></span>![](_page_5_Figure_2.jpeg)
+
+*Figure 4.* Figure(a)–(c) presents every edit accuracy on SCOTUS datasets with BERT. All experiment settings are the same as Tab[.1.](#page-4-0)
+
+<span id="page-5-1"></span>![](_page_5_Figure_4.jpeg)
+
+*Figure 5.* Visualization of zsRE dataset encoder output from T5 small with t-SNE and the dots with same color and shape are input and rephrase sentence. All experiment settings are the same as Tab[.1.](#page-4-0)
+
+#### 4.2. Main Results
+
+We evaluate the performance of the proposed SoLA method on three benchmark datasets and compare it with several state-of-the-art model editing approaches. The results are summarized in Tab[.1.](#page-4-0) As observed, SoLA achieves excellent performance on most of the benchmark datasets, with SoLA outperforming the strongest baseline, MELO, by 3% on the SCOTUS dataset. These demonstrate SoLA's enhanced effectiveness in editing target knowledge while better preserving previously acquired knowledge. Furthermore, existing methods such as MEND and SERAC perform poorly under the sequential editing setting, suggesting that they are primarily designed for static editing scenarios and are not well-suited for sequential model editing tasks.
+
+We also report task-wise accuracy progression to analyze model performance throughout the sequential editing process. The experimental results are based on the SCOTUS dataset and the results are shown in Fig[.4.](#page-5-0) As illustrated in Fig[.4a,](#page-5-0) Fig[.4b](#page-5-0) and Fig[.4c,](#page-5-0) SoLA consistently maintains
+
+superior accuracy across tasks, without exhibiting substantial performance degradation or volatility. This indicates the robustness and stability of SoLA when applied to sequential knowledge editing.
+
+Method performance may vary across different model sizes and datasets. To further evaluate the robustness of our approach, we conduct hallucination correction experiments on the UniEdit[\(Chen et al.,](#page-8-13) [2025\)](#page-8-13) and WikiBigEdit[\(Thede](#page-9-11) [et al.,](#page-9-11) [2025\)](#page-9-11) datasets using larger-scale backbone models. The results are presented in Tab[.2.](#page-6-0) As shown in Tab[.2,](#page-6-0) SoLA consistently achieves the best or near-best performance across most settings, demonstrating its robustness and effectiveness under diverse task configurations. During editing, larger backbone models generally exhibit more stable ERR values, suggesting that such models have learned stronger semantic representations during pretraining, leading to better generalization and editability. Furthermore, continuous learning related methods tend to overfit in edited data, thus showing better ARR in the retention set, but large TRR values in the upstream set, indicating severe forgetting of previous knowledge.
+
+#### 4.3. Controllable Rollback Editing
+
+In SoLA, each edit instance is associated with a unique key stored in both the memory and the routing table. This design enables reversible editing, as the model can revert to its original behavior by removing the key corresponding to a specific edit. To demonstrate this property, we conduct an illustrative experiment on the zsRE dataset. The results are shown in Tab[.3.](#page-7-0) In the Tab[.3,](#page-7-0) each row corresponds to an input instance consisting of a "Text" (the question need to be edited) and its "Labels" (the correct answer). "Del" indicates whether the key corresponding to the edit is removed after editing. For each input, Predbase, Prededit and Preddel denote the prediction of base model, edited model and model after removing the corresponding key. When "Del" is false, the key is retained, serving as a baseline comparison. As shown in Tab[.3,](#page-7-0) for each instance, the original prediction (Predbase) does not match the correct label, indi-
+
+{6}------------------------------------------------
+
+<span id="page-6-0"></span>*Table 2.* Model Performance Comparison on UniEdit and WikiBigEdit datasets in hallucination correction tasks. All the scores are PPL metric and the best result is in bolded.
+
+|                |        |       | UniEdit (PPL ↓) |       |       | WikiBigEdit (PPL ↓) |       |
+|----------------|--------|-------|-----------------|-------|-------|---------------------|-------|
+| Model          | Method | ERR   | TRR             | ARR   | ERR   | TRR                 | ARR   |
+|                | EWC    | 2.43  | 1134            | 22.08 | 3.58  | 251117              | 4.93  |
+|                | CMR    | 2.79  | 2239            | 25.24 | 2.86  | 169452              | 4.49  |
+| LLaMA-3-8B     | CLEAR  | 1.04  | 982             | 29.79 | 1.03  | 74384               | 1.52  |
+|                | SERAC  | 37.10 | 246             | 93    | 60.17 | 215                 | 59    |
+|                | GRACE  | 1.00  | 244             | 89    | 1.00  | 216                 | 58    |
+|                | ELDER  | 1.00  | 173             | 74    | 1.00  | 178                 | 48    |
+|                | MELO   | 1.00  | 153             | 68    | 1.00  | 165                 | 47    |
+|                | Ours   | 1.00  | 144             | 66    | 1.00  | 162                 | 44    |
+|                | EWC    | 2.38  | 4451            | 42.13 | 2.86  | 47846               | 3.89  |
+|                | CMR    | 2.34  | 7211            | 36.19 | 3.60  | 34623               | 5.51  |
+| DeepSeek-R1-8B | CLEAR  | 1.06  | 3365            | 44.86 | 1.02  | 28110               | 1.48  |
+|                | SERAC  | 46.76 | 451             | 834   | 82.72 | 493                 | 831   |
+|                | GRACE  | 1.00  | 407             | 803   | 1.00  | 487                 | 793   |
+|                | ELDER  | 1.03  | 241             | 482   | 1.03  | 304                 | 492   |
+|                | MELO   | 1.09  | 204             | 407   | 1.05  | 263                 | 445   |
+|                | Ours   | 1.07  | 188             | 374   | 1.00  | 248                 | 412   |
+|                | EWC    | 4.62  | 7648            | 48.62 | 5.61  | 61834               | 14.65 |
+|                | CMR    | 4.12  | 11583           | 40.56 | 6.28  | 48629               | 16.89 |
+| Qwen2-7B       | CLEAR  | 1.03  | 5021            | 54.98 | 1.05  | 35396               | 4.26  |
+|                | SERAC  | 58.51 | 329             | 1261  | 62.12 | 304                 | 523   |
+|                | GRACE  | 2.05  | 316             | 1065  | 16.88 | 299                 | 518   |
+|                | ELDER  | 1.00  | 201             | 266   | 1.00  | 225                 | 197   |
+|                | MELO   | 1.00  | 170             | 201   | 1.00  | 203                 | 143   |
+|                | Ours   | 1.00  | 156             | 110   | 1.00  | 199                 | 109   |
+
+cating the necessity for editing. After editing, the model correctly produces Prededit consistent with the ground-truth label, demonstrating that SoLA successfully updates the model's knowledge. When the associated key is deleted, the model reverts to its original prediction (Prededit = Predbase), verifying that our method can effectively roll back specific edits. More critically, for edits where the key is not deleted, the model continues to output Prededit, showing that the rollback operation does not interfere with other edits. This confirms that SoLA supports fine-grained, selective undo of knowledge edits without disrupting unrelated modifications.
+
+#### 4.4. Ablation Study
+
+330
+
+334
+
+336
+
+354
+
+356
+
+374
+
+376
+
+Effect of Edit Layer Location In semantic representation learning, different layers of a model focus on different aspects of semantic information. Therefore, the position of the edited layer can significantly influence the effectiveness of model editing. To investigate this, we evaluate the performance of model when edits are applied at different layers, while keeping all other settings fixed. The experiments are conducted on the SCOTUS dataset with a Nvidia A100 40G GPU, and the results are shown in Tab[.4.](#page-7-1) In the
+
+Tab[.4,](#page-7-1) "Layer" denotes the layer range used for editing. For example, "0–2" indicates that edits are applied to layers 0, 1, and 2 of the BERT model. As shown in the table, editing at shallower layers results in suboptimal performance. This observation aligns with prior findings[\(Geva et al.,](#page-8-14) [2020\)](#page-8-14), which suggest that shallow layers primarily capture the shallow sentence patterns, whereas deeper layers encode richer semantic features, thereby enabling more effective knowledge editing. Moreover, editing at shallow layers significantly increases the time used for editing, which means that editing for semantics is more difficult in shallow layers. Furthermore, we observe that varying the editing layer has negligible impact on performance over the unedited portion of the dataset, indicating that SoLA is capable of preserving previously learned knowledge regardless of the editing depth, which further supports the robustness of SoLA.
+
+Effect of LoRA Rank The LoRA module projects semantic representations into a low-rank subspace for parameterefficient learning. Consequently, the choice of rank will effect the module's performance. To investigate this, we analyze the impact of different LoRA rank values on model performance, keeping all other settings fixed. Experiments are conducted on the SCOTUS dataset using a NVIDIA
+
+{7}------------------------------------------------
+
+387
+
+394
+
+428 429 430 431 432 433 434 435
+
+<span id="page-7-0"></span>*Table 3.* Controllable edit in SoLA on zsRE datasets. "Text" is the question need to be edited, "Labels" is the true answer, "Del" indicates whther delete associated key and Predbase, Prededit and Preddel is the prediction of base model, edited model and delted model.
+
+| Text                                                | Labels            | Del   | Predbase         | Prededit          | Preddel          |
+|-----------------------------------------------------|-------------------|-------|------------------|-------------------|------------------|
+| The date of Tsetserleg earthquake in 1905?          | 9 July 1905       | True  | 20 February 1905 | 9 July 1905       | 20 February 1905 |
+| What constellation has Nu Cancri?                   | Cancer            | True  | Hydra            | Cancer            | Hydra            |
+| On what date did the Bahrain Grand Prix 2015 occur? | 19 April 2015     | True  | 6 April 2017     | 19 April 2015     | 6 April 2017     |
+| What position does Charles-Joseph Coursol have?     | Mayor of Montreal | True  | Positioning      | Mayor of Montreal | Positioning      |
+| Who acted in Blow Dry?                              | Alan Rickman      | False | Jim Carrey       | Alan Rickman      | Alan Rickman     |
+
+<span id="page-7-1"></span>*Table 4.* Ablation studies on location of edited layers. Here, "Layer" indicates the edited layer location. For example, "0-2" presents edited layer is layers 0, 1 and 2.
+
+| Layer | ES   | ERR  | TRR  | Edit Time(min) |
+|-------|------|------|------|----------------|
+| 0-2   | 0.77 | 0.61 | 0.96 | 9.21           |
+| 3-5   | 1.00 | 0.68 | 0.96 | 8.06           |
+| 6-8   | 0.81 | 0.70 | 0.97 | 7.03           |
+| 9-11  | 0.94 | 0.95 | 0.97 | 5.97           |
+
+<span id="page-7-2"></span>*Table 5.* Ablation studies on LoRA rank. Here, "LoRA Rank" indicates the rank value of every LoRA in model.
+
+| LoRA Rank | ES   | ERR  | TRR  | Edit Time(min) |
+|-----------|------|------|------|----------------|
+| 1         | 0.90 | 0.84 | 0.96 | 5.88           |
+| 2         | 0.94 | 0.91 | 0.97 | 5.90           |
+| 3         | 1.00 | 0.91 | 0.97 | 5.86           |
+| 4         | 1.00 | 0.95 | 0.97 | 5.97           |
+| 5         | 1.00 | 0.92 | 0.96 | 5.86           |
+| 10        | 1.00 | 0.71 | 0.96 | 5.85           |
+|           |      |      |      |                |
+
+A100 40GB GPU, and the results are presented in Tab[.5.](#page-7-2) As shown in the Tab[.5,](#page-7-2) simply increasing the LoRA rank does not lead to improved performance and may even degrade model performance. This indicates that expanding the learnable parameter space does not necessarily enhance the model's capacity for effective knowledge editing. On the contrary, excessive rank values may even result in performance degradation due to overfitting, suggesting the importance of carefully selecting an appropriate rank to balance capacity and generalization. Based on the experiment results, we set 4 as LoRA rank for all the experiments in this paper.
+
+#### 4.5. Model Visualization
+
+To gain deeper insight into the model's behavior under sequential editing, we conducted a t-SNE visualization [\(Maaten & Hinton,](#page-9-12) [2008\)](#page-9-12) of the learned feature representations. This experiment is performed on the zsRE dataset using the T5-small model to assess how semantic relationships are preserved after sequential edits. Upon completing
+
+the entire sequence of editing tasks for question answering, we selected five distinct input instances, each accompanied by semantically equivalent rephrased sentences, forming five evaluation groups. The encoder output features are subsequently extracted from the edited model for dimensionality reduction and visualization, and the resulting plot is represented in Fig[.5.](#page-5-1) As shown in Fig[.5,](#page-5-1) there is a clear clustering pattern that for each group, the original input and its rephrased variant are mapped to proximate locations in the latent space. This demonstrates that SoLA effectively captures and preserves the semantic similarity between related inputs throughout the editing process. Furthermore, distinct groups maintain separable clusters, indicating that the model retains the ability to discriminate between different edit concepts after sequential editing. This visualization shows that SoLA's semantic routing mechanism successfully supports edit generalization based on semantic similarity while upholding the integrity of previously learned knowledge.
+
+## 5. Conclusion
+
+In this paper, we propose SoLA, a Semantic routing-based LoRA framework for reversible lifelong model editing. In SoLA, each edit is encapsulated as an independent LoRA module which is frozen after training and a semantic routing record is established to map LoRA module to the input semantic representation, allowing dynamic activation of LoRA modules via semantic matching. This design not only avoids semantic drift caused by clustering updating but also mitigates catastrophic forgetting from parameter sharing. More importantly, SoLA supports precisely revoking edit by removing the corresponding key from the routing table, enabling reversible model editing. This allows for flexible addition and deletion of edits, offering fine-grained control over model behavior. To the best of our knowledge, this is the first work to achieve reversible model editing in the literature. Furthermore, we introduce a master decisionmaking mechanism by integrating decision-making into the edited layer, enabling end-to-end decision-making process. By building a strict mapping between edits and LoRA modules, SoLA achieves efficient and reversible lifelong model editing, providing a novel perspective for future research.
+
+{8}------------------------------------------------
+
+## Impact Statement
+
+This work advances the field of Machine Learning by introducing SoLA, a framework for reversible and controllable lifelong model editing in large language models. Our method enhances the safety and reliability of AI systems by enabling precise rollback of model edits, which helps mitigate risks associated with erroneous or harmful knowledge updates. By effectively preventing semantic drift and catastrophic forgetting, SoLA promotes the development of more robust and trustworthy models capable of adapting to evolving information. The significant reduction in computational overhead also aligns with the goal of sustainable AI development.
+
+## References
+
+- <span id="page-8-1"></span>Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al. Gpt-4 technical report. *arXiv preprint arXiv:2303.08774*, 2023.
+- <span id="page-8-6"></span>Biderman, S., Schoelkopf, H., Anthony, Q. G., Bradley, H., O'Brien, K., Hallahan, E., Khan, M. A., Purohit, S., Prashanth, U. S., Raff, E., et al. Pythia: A suite for analyzing large language models across training and scaling. In *International Conference on Machine Learning*, pp. 2397–2430. PMLR, 2023.
+- <span id="page-8-12"></span>Brown, P. F., Della Pietra, S. A., Della Pietra, V. J., Lai, J. C., and Mercer, R. L. An estimate of an upper bound for the entropy of english. *Computational Linguistics*, 18 (1):31–40, 1992.
+- <span id="page-8-0"></span>Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al. Language models are few-shot learners. *Advances in neural information processing systems*, 33: 1877–1901, 2020.
+- <span id="page-8-17"></span>Chalkidis, I., Pasini, T., Zhang, S., Tomada, L., Schwemer, S. F., and Søgaard, A. Fairlex: A multilingual benchmark for evaluating fairness in legal text processing. *arXiv preprint arXiv:2203.07228*, 2022.
+- <span id="page-8-13"></span>Chen, Q., Wang, D., Zhang, T., Yan, Z., You, C., Wang, C., and He, X. Uniedit: A unified knowledge editing benchmark for large language models. *arXiv preprint arXiv:2505.12345*, 2025.
+- <span id="page-8-4"></span>De Cao, N., Aziz, W., and Titov, I. Editing factual knowledge in language models. *arXiv preprint arXiv:2104.08164*, 2021.
+- <span id="page-8-8"></span>Fang, J., Jiang, H., Wang, K., Ma, Y., Jie, S., Wang, X., He, X., and Chua, T.-S. Alphaedit: Null-space constrained knowledge editing for language models. *arXiv preprint arXiv:2410.02355*, 2024.
+
+- <span id="page-8-14"></span>Geva, M., Schuster, R., Berant, J., and Levy, O. Transformer feed-forward layers are key-value memories. *arXiv preprint arXiv:2012.14913*, 2020.
+- <span id="page-8-7"></span>Gu, J.-C., Xu, H.-X., Ma, J.-Y., Lu, P., Ling, Z.-H., Chang, K.-W., and Peng, N. Model editing can hurt general abilities of large language models. *CoRR*, 2024.
+- <span id="page-8-3"></span>Hartvigsen, T., Gabriel, S., Palangi, H., Sap, M., Ray, D., and Kamar, E. Toxigen: A large-scale machine-generated dataset for adversarial and implicit hate speech detection. *arXiv preprint arXiv:2203.09509*, 2022.
+- <span id="page-8-9"></span>Hartvigsen, T., Sankaranarayanan, S., Palangi, H., Kim, Y., and Ghassemi, M. Aging with grace: Lifelong model editing with discrete key-value adaptors. *Advances in Neural Information Processing Systems*, 36:47934–47959, 2023.
+- <span id="page-8-15"></span>Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., and Gelly, S. Parameter-efficient transfer learning for nlp. In *International conference on machine learning*, pp. 2790– 2799. PMLR, 2019.
+- <span id="page-8-10"></span>Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al. Lora: Low-rank adaptation of large language models. *ICLR*, 1(2):3, 2022.
+- <span id="page-8-2"></span>Huang, L., Yu, W., Ma, W., Zhong, W., Feng, Z., Wang, H., Chen, Q., Peng, W., Feng, X., Qin, B., et al. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. *ACM Transactions on Information Systems*, 43(2):1–55, 2025.
+- <span id="page-8-16"></span>Jia, M., Tang, L., Chen, B.-C., Cardie, C., Belongie, S., Hariharan, B., and Lim, S.-N. Visual prompt tuning. In *European conference on computer vision*, pp. 709–727. Springer, 2022.
+- <span id="page-8-11"></span>Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A. A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al. Overcoming catastrophic forgetting in neural networks. *Proceedings of the national academy of sciences*, 114(13):3521–3526, 2017.
+- <span id="page-8-18"></span>Kwiatkowski, T., Palomaki, J., Redfield, O., Collins, M., Parikh, A., Alberti, C., Epstein, D., Polosukhin, I., Devlin, J., Lee, K., et al. Natural questions: a benchmark for question answering research. *Transactions of the Association for Computational Linguistics*, 7:453–466, 2019.
+- <span id="page-8-5"></span>Lazaridou, A., Kuncoro, A., Gribovskaya, E., Agrawal, D., Liska, A., Terzi, T., Gimenez, M., de Masson d'Autume, C., Kocisky, T., Ruder, S., et al. Mind the gap: Assessing temporal generalization in neural language models. *Advances in Neural Information Processing Systems*, 34: 29348–29363, 2021.
+
+{9}------------------------------------------------
+
+<span id="page-9-16"></span>495 496 497 Levy, O., Seo, M., Choi, E., and Zettlemoyer, L. Zero-shot relation extraction via reading comprehension. *arXiv preprint arXiv:1706.04115*, 2017.
+
+<span id="page-9-14"></span>504
+
+<span id="page-9-12"></span>513 514
+
+<span id="page-9-17"></span>516
+
+518 519 520
+
+<span id="page-9-7"></span>524 525 526
+
+<span id="page-9-4"></span>528 529 530
+
+534
+
+<span id="page-9-18"></span>536
+
+- <span id="page-9-6"></span>Li, J., Wang, Q., Wang, Z., Zhang, Y., and Mao, Z. Elder: Enhancing lifelong model editing with mixture-of-lora. In *Proceedings of the AAAI Conference on Artificial Intelligence*, pp. 24440–24448, 2025.
+- Li, X. L. and Liang, P. Prefix-tuning: Optimizing continuous prompts for generation. *arXiv preprint arXiv:2101.00190*, 2021.
+- <span id="page-9-9"></span>Lin, B. Y., Wang, S., Lin, X. V., Jia, R., Xiao, L., Ren, X., and Yih, W.-t. On continual model refinement in out-of-distribution data streams. *arXiv preprint arXiv:2205.02014*, 2022.
+- Maaten, L. v. d. and Hinton, G. Visualizing data using t-sne. *Journal of machine learning research*, 9(Nov): 2579–2605, 2008.
+- Manakul, P., Liusie, A., and Gales, M. J. Selfcheckgpt: Zeroresource black-box hallucination detection for generative large language models. *arXiv preprint arXiv:2303.08896*, 2023.
+- <span id="page-9-5"></span>Meng, K., Bau, D., Andonian, A., and Belinkov, Y. Locating and editing factual associations in gpt. *Advances in neural information processing systems*, 35:17359–17372, 2022a.
+- Meng, K., Sharma, A. S., Andonian, A., Belinkov, Y., and Bau, D. Mass-editing memory in a transformer. *arXiv preprint arXiv:2210.07229*, 2022b.
+- Mitchell, E., Lin, C., Bosselut, A., Finn, C., and Manning, C. D. Fast model editing at scale. *arXiv preprint arXiv:2110.11309*, 2021.
+- <span id="page-9-8"></span>Mitchell, E., Lin, C., Bosselut, A., Manning, C. D., and Finn, C. Memory-based model editing at scale. In *International Conference on Machine Learning*, pp. 15817– 15831. PMLR, 2022.
+- Nakano, R., Hilton, J., Balaji, S., Wu, J., Ouyang, L., Kim, C., Hesse, C., Jain, S., Kosaraju, V., Saunders, W., et al. Webgpt: Browser-assisted question-answering with human feedback. *arXiv preprint arXiv:2112.09332*, 2021.
+- <span id="page-9-0"></span>Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al. Language models are unsupervised multitask learners. *OpenAI blog*, 1(8):9, 2019.
+- <span id="page-9-10"></span>Rolnick, D., Ahuja, A., Schwarz, J., Lillicrap, T., and Wayne, G. Experience replay for continual learning. *Advances in neural information processing systems*, 32, 2019.
+
+- <span id="page-9-11"></span>Thede, L., Roth, K., Bethge, M., Akata, Z., and Hartvigsen, T. Understanding the limits of lifelong knowledge editing in llms. *arXiv preprint arXiv:2503.05683*, 2025.
+- <span id="page-9-1"></span>Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al. Llama 2: Open foundation and finetuned chat models. *arXiv preprint arXiv:2307.09288*, 2023.
+- <span id="page-9-15"></span>Valipour, M., Rezagholizadeh, M., Kobyzev, I., and Ghodsi, A. Dylora: Parameter efficient tuning of pre-trained models using dynamic search-free low-rank adaptation. *arXiv preprint arXiv:2210.07558*, 2022.
+- <span id="page-9-2"></span>Yao, Y., Wang, P., Tian, B., Cheng, S., Li, Z., Deng, S., Chen, H., and Zhang, N. Editing large language models: Problems, methods, and opportunities. *arXiv preprint arXiv:2305.13172*, 2023.
+- <span id="page-9-3"></span>Yu, L., Chen, Q., Zhou, J., and He, L. Melo: Enhancing model editing with neuron-indexed dynamic lora. In *Proceedings of the AAAI Conference on Artificial Intelligence*, pp. 19449–19457, 2024.
+- <span id="page-9-13"></span>Zaken, E. B., Ravfogel, S., and Goldberg, Y. Bitfit: Simple parameter-efficient fine-tuning for transformerbased masked language-models. *arXiv preprint arXiv:2106.10199*, 2021.
+
+{10}------------------------------------------------
+
+554
+
+556
+
+574
+
+576
+
+594
+
+596
+
+# Appendix
+
+## A. Additional Related Work
+
+#### A.1. Parameter-Efficient Fine-Tuning
+
+Parameter-efficient fine-tuning (PEFT) approaches adjust pretrained LLMs by incorporating lightweight modules, while keeping the base model frozen. Optimization is performed solely through updates to these lightweight components. Representative approaches include Adapters[\(Houlsby](#page-8-15) [et al.,](#page-8-15) [2019;](#page-8-15) [Zaken et al.,](#page-9-13) [2021\)](#page-9-13), Prompts[\(Li & Liang,](#page-9-14) [2021;](#page-9-14) [Jia et al.,](#page-8-16) [2022\)](#page-8-16), and Low-Rank Adaptation (LoRA)[\(Hu](#page-8-10) [et al.,](#page-8-10) [2022;](#page-8-10) [Valipour et al.,](#page-9-15) [2022\)](#page-9-15). Adapters are compact bottleneck modules located after transformer blocks; Prompts are learnable vectors prepended to the input sequence; and LoRA uses low-rank decomposition matrices to modify model weights during training. Recent progress explores applying PEFT to lifelong model editing, where the base model remains unchanged and new parameter modules are introduced to manage sequential edits over time.
+
+#### A.2. Routing Mechanisms in Modular Language Models
+
+When multiple modular components (e.g., LoRA modules) are present, a key challenge lies in routing input queries to the appropriate subset of modules during inference. Recent work explore various strategies to support dynamic module selection. MELO[\(Yu et al.,](#page-9-3) [2024\)](#page-9-3) introduces a clustering mechanism to assign edits to clusters based on input hidden states. During inference, the nearest cluster center is used to retrieve relevant LoRA modules. However, cluster centers are inevitably updated during the sequential editing, which may lead to semantic drift and incorrect matching. As shown in Fig[.2,](#page-1-0) the number of incorrect matches increases significantly as the number of cluster centre updating increases, which leads to a deterioration in model performance. EL-DER[\(Li et al.,](#page-9-6) [2025\)](#page-9-6) employs MoE approach in which a learnable neural network scores a set of LoRA modules, activating the top-k modules. This strategy offers parameter efficiency but introduces soft entanglement in routing, and due to shared module usage, new edits may interfere with or overwrite previously edits. Furthermore, both strategies rely on auxiliary routing networks, adding architectural complexity and computational overhead. They also offer limited control over edit selection, making rollback and deletion of edits challenging.
+
+## B. Additional Experiment Details
+
+Benchmarks We evaluate the performance of the model in three representative tasks: Document Classification, Question Answering(QA), and Hallucination Correction. Among them, the document classification task is performed based
+
+on the SCOTUS dataset[\(Chalkidis et al.,](#page-8-17) [2022\)](#page-8-17). SCOTUS is a subset of Fairlex, which collects documents from the U.S. Supreme Court and classifies them into 11 topics. In practice, the topics corresponding to the documents will change, so it is necessary to update the model knowledge to classify the documents into new topics. We divide SCO-TUS into an edit set and an upstream set, where models are edited on the edit set and the upstream set is not involved in editing, only testing. The question answering task is then conducted on the zsRE dataset[\(Levy et al.,](#page-9-16) [2017\)](#page-9-16), and NQ datasets[\(Kwiatkowski et al.,](#page-8-18) [2019\)](#page-8-18) serves as upstream dataset. Specifically, we perform editing in zsRE, and the NQ dataset is again not involved in editing, only testing. For hallucination correction, we adopted the framework introduced by[\(Manakul et al.,](#page-9-17) [2023\)](#page-9-17), aiming to address the tendency of GPT models to generate factually inconsistent outputs. Following GRACE[\(Hartvigsen et al.,](#page-8-9) [2023\)](#page-8-9), we employ SelfCheckGPT[\(Manakul et al.,](#page-9-17) [2023\)](#page-9-17) for editing. SelfCheckGPT is a Wikipedia-style biography that is first generated using GPT3 on 256 topics, e.g., "Bon Jovi", and then checked to see which of the generated biographies are hallucinations. SelfCheckGPT contains 1392 hallucinatory sentences to be edited, which serves as the editing set, and 516 already correct sentences as the retention set, which is used to measure model's post-edit perplexity on the already correct sentences. WebText[\(Nakano et al.,](#page-9-18) [2021\)](#page-9-18) is used as the upstream set, remaining unedited and reserved for testing. In addition to this, in order to test the performance of the model approach with different sized models as well as datasets, we conducted hallucination correction experiments on Uniedit[\(Chen et al.,](#page-8-13) [2025\)](#page-8-13) and WikiBigEdit[\(Thede](#page-9-11) [et al.,](#page-9-11) [2025\)](#page-9-11) to evaluate the generalization performance of the models. Uniedit is based on open domain knowledge covering information from 25 common domains in 5 major categories. The WikiBigEdit dataset is based on regular updates of the knowledge graph in Wikipedia and contains 5 months of extensive factual editing and improvement. The Uniedit and WikiBigEdit datasets already have edit set, retention set and upstream set that can be used directly for hallucination correction tasks.
+
+Training details Following[\(Hartvigsen et al.,](#page-8-9) [2023\)](#page-8-9), we employ BERT, T5-small, and GPT2-XL as the base models for the three tasks—document classification, question answering, and hallucination correction, respectively. Among these, BERT and T5-small are the official pre-trained models, while GPT2-XL is obtained from [\(Hartvigsen et al.,](#page-8-9) [2023\)](#page-8-9) which finetuned from the official release[\(Hartvigsen](#page-8-9) [et al.,](#page-8-9) [2023\)](#page-8-9). For training, we adopt Stochastic Gradient Descent (SGD) as the optimizer with a cosine decay learning rate schedule. The learning rates is 0.05, training epochs
+
+{11}------------------------------------------------
+
+is 40, and LoRA rank is 4. In the three tasks, we edit the models and the edited layers of the individual model edits are shown in the Tab[.6.](#page-11-0)
+
+*Table 6.* Edited layers of different model
+
+<span id="page-11-0"></span>
+
+| Model    | Edit layer                                                                                                                                                                                                                                                                         |
+|----------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+| BERT     | bert.encoder.layer.9.output.dense<br>bert.encoder.layer.10.output.dense<br>bert.encoder.layer.11.output.dense                                                                                                                                                                      |
+| T5-Small | encoder.block.5.layer.1.DenseReluDense.wi 0<br>encoder.block.5.layer.1.DenseReluDense.wi 1<br>encoder.block.5.layer.1.DenseReluDense.wo<br>encoder.block.6.layer.1.DenseReluDense.wi 0<br>encoder.block.6.layer.1.DenseReluDense.wi 1<br>encoder.block.6.layer.1.DenseReluDense.wo |
+| GPT2-XL  | transformer.h.36.mlp.c fc<br>transformer.h.37.mlp.c fc                                                                                                                                                                                                                             |

icml26/31f6f2e8-0fb2-46ff-ab65-f3408612f6e1/parse_report.json

@@ -0,0 +1,75 @@
+{
+  "paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1",
+  "pipeline": "marker_non_llm_v3",
+  "parser": "marker_single",
+  "formats": [
+    "markdown",
+    "chunks"
+  ],
+  "llm_enabled": false,
+  "pdf_path": "/network/scratch/j/jianan.zhao/ReviewAgent/data/processed_papers/icml26_20260429_1952_duequeue/raw/31f6f2e8-0fb2-46ff-ab65-f3408612f6e1.pdf",
+  "pdf_sha256": "992b295e96705a94648379d20a1644e6082883a1cc9641e983bd3ba4527e4eaa",
+  "bytes": 578235,
+  "source": "https://koala.science/storage/pdfs/31f6f2e8-0fb2-46ff-ab65-f3408612f6e1.pdf",
+  "page_count": 12,
+  "ok": true,
+  "elapsed_seconds": 146.58999999999997,
+  "paper2markdown_v3": {
+    "paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1",
+    "pipeline": "Paper2Markdown-V3",
+    "ok": true,
+    "page_count": 12,
+    "chunk_count": 69,
+    "main_body_chunk_count": 57,
+    "appendix_chunk_count": 7,
+    "reference_chunk_count": 5,
+    "model_text_chars": 37283,
+    "raw_markdown_chars": 58572,
+    "sanitized_chars": 53252,
+    "page_provenance": {
+      "min_page": 1,
+      "max_page": 12,
+      "invalid_count": 0
+    },
+    "marker_block_type_counts": {
+      "Caption": 5,
+      "Equation": 3,
+      "FigureGroup": 5,
+      "Footnote": 1,
+      "ListGroup": 5,
+      "PageFooter": 12,
+      "PageHeader": 15,
+      "SectionHeader": 26,
+      "Table": 1,
+      "TableGroup": 5,
+      "Text": 171
+    },
+    "asset_count_raw": 5,
+    "asset_count_model_kept": 5,
+    "asset_count_rejected": 0,
+    "asset_reject_reasons": {
+      "kept": 5
+    },
+    "artifact_leak_audit": {
+      "ok": true,
+      "hits": {
+        "Anonymous Authors": [],
+        "ACKNOWLEDGMENT": [],
+        "OpenReview": [],
+        "\"accept_label\"": [],
+        "\"decision\"": [],
+        "\"decision_tier\"": [],
+        "\"source_status\"": [],
+        "Meta-review": [],
+        "Official Review": [],
+        "official_reviews": [],
+        "meta_reviews": [],
+        "suggested_verdict_score": []
+      },
+      "artifact_count": 2
+    },
+    "default_model_input": "model_text_v3.txt",
+    "appendix_input": "appendix_text_v3.txt",
+    "reference_input": "reference_text_v3.txt"
+  }
+}

icml26/31f6f2e8-0fb2-46ff-ab65-f3408612f6e1/reference_chunks.jsonl

@@ -0,0 +1,5 @@
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0057", "section": "References", "page_start": 9, "page_end": 9, "type": "ListGroup", "text": "Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774 , 2023. Biderman, S., Schoelkopf, H., Anthony, Q. G., Bradley, H., O'Brien, K., Hallahan, E., Khan, M. A., Purohit, S., Prashanth, U. S., Raff, E., et al. Pythia: A suite for analyzing large language models across training and scaling. In International Conference on Machine Learning , pp. 2397–2430. PMLR, 2023. Brown, P. F., Della Pietra, S. A., Della Pietra, V. J., Lai, J. C., and Mercer, R. L. An estimate of an upper bound for the entropy of english. Computational Linguistics , 18 (1):31–40, 1992. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al. Language models are few-shot learners. Advances in neural information processing systems , 33: 1877–1901, 2020. Chalkidis, I., Pasini, T., Zhang, S., Tomada, L., Schwemer, S. F., and Søgaard, A. Fairlex: A multilingual benchmark for evaluating fairness in legal text processing. arXiv preprint arXiv:2203.07228 , 2022. Chen, Q., Wang, D., Zhang, T., Yan, Z., You, C., Wang, C., and He, X. Uniedit: A unified knowledge editing benchmark for large language models. arXiv preprint arXiv:2505.12345 , 2025. De Cao, N., Aziz, W., and Titov, I. Editing factual knowledge in language models. arXiv preprint arXiv:2104.08164 , 2021. Fang, J., Jiang, H., Wang, K., Ma, Y., Jie, S., Wang, X., He, X., and Chua, T.-S. Alphaedit: Null-space constrained knowledge editing for language models. arXiv preprint arXiv:2410.02355 , 2024.", "source": "marker_v2", "marker_block_id": "/page/8/ListGroup/516"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0058", "section": "References", "page_start": 9, "page_end": 9, "type": "ListGroup", "text": "Geva, M., Schuster, R., Berant, J., and Levy, O. Transformer feed-forward layers are key-value memories. arXiv preprint arXiv:2012.14913 , 2020. Gu, J.-C., Xu, H.-X., Ma, J.-Y., Lu, P., Ling, Z.-H., Chang, K.-W., and Peng, N. Model editing can hurt general abilities of large language models. CoRR , 2024. Hartvigsen, T., Gabriel, S., Palangi, H., Sap, M., Ray, D., and Kamar, E. Toxigen: A large-scale machine-generated dataset for adversarial and implicit hate speech detection. arXiv preprint arXiv:2203.09509 , 2022. Hartvigsen, T., Sankaranarayanan, S., Palangi, H., Kim, Y., and Ghassemi, M. Aging with grace: Lifelong model editing with discrete key-value adaptors. Advances in Neural Information Processing Systems , 36:47934–47959, 2023. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., and Gelly, S. Parameter-efficient transfer learning for nlp. In International conference on machine learning , pp. 2790– 2799. PMLR, 2019. Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al. Lora: Low-rank adaptation of large language models. ICLR , 1(2):3, 2022. Huang, L., Yu, W., Ma, W., Zhong, W., Feng, Z., Wang, H., Chen, Q., Peng, W., Feng, X., Qin, B., et al. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Information Systems , 43(2):1–55, 2025. Jia, M., Tang, L., Chen, B.-C., Cardie, C., Belongie, S., Hariharan, B., and Lim, S.-N. Visual prompt tuning. In European conference on computer vision , pp. 709–727. Springer, 2022. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A. A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences , 114(13):3521–3526, 2017. Kwiatkowski, T., Palomaki, J., Redfield, O., Collins, M., Parikh, A., Alberti, C., Epstein, D., Polosukhin, I., Devlin, J., Lee, K., et al. Natural questions: a benchmark for question answering research. Transactions of the Association for Computational Linguistics , 7:453–466, 2019. Lazaridou, A., Kuncoro, A., Gribovskaya, E., Agrawal, D., Liska, A., Terzi, T., Gimenez, M., de Masson d'Autume, C., Kocisky, T., Ruder, S., et al. Mind the gap: Assessing temporal generalization in neural language models. Advances in Neural Information Processing Systems , 34: 29348–29363, 2021.", "source": "marker_v2", "marker_block_id": "/page/8/ListGroup/517"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0059", "section": "References", "page_start": 10, "page_end": 10, "type": "Text", "text": "495 496 497 Levy, O., Seo, M., Choi, E., and Zettlemoyer, L. Zero-shot relation extraction via reading comprehension. arXiv preprint arXiv:1706.04115 , 2017.", "source": "marker_v2", "marker_block_id": "/page/9/Text/408"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0060", "section": "References", "page_start": 10, "page_end": 10, "type": "ListGroup", "text": "Li, J., Wang, Q., Wang, Z., Zhang, Y., and Mao, Z. Elder: Enhancing lifelong model editing with mixture-of-lora. In Proceedings of the AAAI Conference on Artificial In telligence , pp. 24440–24448, 2025. Li, X. L. and Liang, P. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190 , 2021. Lin, B. Y., Wang, S., Lin, X. V., Jia, R., Xiao, L., Ren, X., and Yih, W.-t. On continual model refinement in out-of-distribution data streams. arXiv preprint arXiv:2205.02014 , 2022. Maaten, L. v. d. and Hinton, G. Visualizing data using t-sne. Journal of machine learning research , 9(Nov): 2579–2605, 2008. Manakul, P., Liusie, A., and Gales, M. J. Selfcheckgpt: Zeroresource black-box hallucination detection for generative large language models. arXiv preprint arXiv:2303.08896 , 2023. Meng, K., Bau, D., Andonian, A., and Belinkov, Y. Locating and editing factual associations in gpt. Advances in neural information processing systems , 35:17359–17372, 2022a. Meng, K., Sharma, A. S., Andonian, A., Belinkov, Y., and Bau, D. Mass-editing memory in a transformer. arXiv preprint arXiv:2210.07229 , 2022b. Mitchell, E., Lin, C., Bosselut, A., Finn, C., and Manning, C. D. Fast model editing at scale. arXiv preprint arXiv:2110.11309 , 2021. Mitchell, E., Lin, C., Bosselut, A., Manning, C. D., and Finn, C. Memory-based model editing at scale. In Inter national Conference on Machine Learning , pp. 15817– 15831. PMLR, 2022. Nakano, R., Hilton, J., Balaji, S., Wu, J., Ouyang, L., Kim, C., Hesse, C., Jain, S., Kosaraju, V., Saunders, W., et al. Webgpt: Browser-assisted question-answering with human feedback. arXiv preprint arXiv:2112.09332 , 2021. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al. Language models are unsupervised multitask learners. OpenAI blog , 1(8):9, 2019. Rolnick, D., Ahuja, A., Schwarz, J., Lillicrap, T., and Wayne, G. Experience replay for continual learning. Advances in neural information processing systems , 32, 2019.", "source": "marker_v2", "marker_block_id": "/page/9/ListGroup/406"}
+{"paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1", "chunk_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1:0061", "section": "References", "page_start": 10, "page_end": 10, "type": "ListGroup", "text": "Thede, L., Roth, K., Bethge, M., Akata, Z., and Hartvigsen, T. Understanding the limits of lifelong knowledge editing in llms. arXiv preprint arXiv:2503.05683 , 2025. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al. Llama 2: Open foundation and finetuned chat models. arXiv preprint arXiv:2307.09288 , 2023. Valipour, M., Rezagholizadeh, M., Kobyzev, I., and Ghodsi, A. Dylora: Parameter efficient tuning of pre-trained models using dynamic search-free low-rank adaptation. arXiv preprint arXiv:2210.07558 , 2022. Yao, Y., Wang, P., Tian, B., Cheng, S., Li, Z., Deng, S., Chen, H., and Zhang, N. Editing large language models: Problems, methods, and opportunities. arXiv preprint arXiv:2305.13172 , 2023. Yu, L., Chen, Q., Zhou, J., and He, L. Melo: Enhancing model editing with neuron-indexed dynamic lora. In Proceedings of the AAAI Conference on Artificial Intelli gence , pp. 19449–19457, 2024. Zaken, E. B., Ravfogel, S., and Goldberg, Y. Bitfit: Simple parameter-efficient fine-tuning for transformerbased masked language-models. arXiv preprint arXiv:2106.10199 , 2021.", "source": "marker_v2", "marker_block_id": "/page/9/ListGroup/407"}

icml26/31f6f2e8-0fb2-46ff-ab65-f3408612f6e1/reference_text_v3.txt

@@ -0,0 +1,14 @@
+[p. 9 | section: References | type: ListGroup]
+Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774 , 2023. Biderman, S., Schoelkopf, H., Anthony, Q. G., Bradley, H., O'Brien, K., Hallahan, E., Khan, M. A., Purohit, S., Prashanth, U. S., Raff, E., et al. Pythia: A suite for analyzing large language models across training and scaling. In International Conference on Machine Learning , pp. 2397–2430. PMLR, 2023. Brown, P. F., Della Pietra, S. A., Della Pietra, V. J., Lai, J. C., and Mercer, R. L. An estimate of an upper bound for the entropy of english. Computational Linguistics , 18 (1):31–40, 1992. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al. Language models are few-shot learners. Advances in neural information processing systems , 33: 1877–1901, 2020. Chalkidis, I., Pasini, T., Zhang, S., Tomada, L., Schwemer, S. F., and Søgaard, A. Fairlex: A multilingual benchmark for evaluating fairness in legal text processing. arXiv preprint arXiv:2203.07228 , 2022. Chen, Q., Wang, D., Zhang, T., Yan, Z., You, C., Wang, C., and He, X. Uniedit: A unified knowledge editing benchmark for large language models. arXiv preprint arXiv:2505.12345 , 2025. De Cao, N., Aziz, W., and Titov, I. Editing factual knowledge in language models. arXiv preprint arXiv:2104.08164 , 2021. Fang, J., Jiang, H., Wang, K., Ma, Y., Jie, S., Wang, X., He, X., and Chua, T.-S. Alphaedit: Null-space constrained knowledge editing for language models. arXiv preprint arXiv:2410.02355 , 2024.
+
+[p. 9 | section: References | type: ListGroup]
+Geva, M., Schuster, R., Berant, J., and Levy, O. Transformer feed-forward layers are key-value memories. arXiv preprint arXiv:2012.14913 , 2020. Gu, J.-C., Xu, H.-X., Ma, J.-Y., Lu, P., Ling, Z.-H., Chang, K.-W., and Peng, N. Model editing can hurt general abilities of large language models. CoRR , 2024. Hartvigsen, T., Gabriel, S., Palangi, H., Sap, M., Ray, D., and Kamar, E. Toxigen: A large-scale machine-generated dataset for adversarial and implicit hate speech detection. arXiv preprint arXiv:2203.09509 , 2022. Hartvigsen, T., Sankaranarayanan, S., Palangi, H., Kim, Y., and Ghassemi, M. Aging with grace: Lifelong model editing with discrete key-value adaptors. Advances in Neural Information Processing Systems , 36:47934–47959, 2023. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., and Gelly, S. Parameter-efficient transfer learning for nlp. In International conference on machine learning , pp. 2790– 2799. PMLR, 2019. Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al. Lora: Low-rank adaptation of large language models. ICLR , 1(2):3, 2022. Huang, L., Yu, W., Ma, W., Zhong, W., Feng, Z., Wang, H., Chen, Q., Peng, W., Feng, X., Qin, B., et al. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Information Systems , 43(2):1–55, 2025. Jia, M., Tang, L., Chen, B.-C., Cardie, C., Belongie, S., Hariharan, B., and Lim, S.-N. Visual prompt tuning. In European conference on computer vision , pp. 709–727. Springer, 2022. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A. A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences , 114(13):3521–3526, 2017. Kwiatkowski, T., Palomaki, J., Redfield, O., Collins, M., Parikh, A., Alberti, C., Epstein, D., Polosukhin, I., Devlin, J., Lee, K., et al. Natural questions: a benchmark for question answering research. Transactions of the Association for Computational Linguistics , 7:453–466, 2019. Lazaridou, A., Kuncoro, A., Gribovskaya, E., Agrawal, D., Liska, A., Terzi, T., Gimenez, M., de Masson d'Autume, C., Kocisky, T., Ruder, S., et al. Mind the gap: Assessing temporal generalization in neural language models. Advances in Neural Information Processing Systems , 34: 29348–29363, 2021.
+
+[p. 10 | section: References | type: Text]
+495 496 497 Levy, O., Seo, M., Choi, E., and Zettlemoyer, L. Zero-shot relation extraction via reading comprehension. arXiv preprint arXiv:1706.04115 , 2017.
+
+[p. 10 | section: References | type: ListGroup]
+Li, J., Wang, Q., Wang, Z., Zhang, Y., and Mao, Z. Elder: Enhancing lifelong model editing with mixture-of-lora. In Proceedings of the AAAI Conference on Artificial In telligence , pp. 24440–24448, 2025. Li, X. L. and Liang, P. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190 , 2021. Lin, B. Y., Wang, S., Lin, X. V., Jia, R., Xiao, L., Ren, X., and Yih, W.-t. On continual model refinement in out-of-distribution data streams. arXiv preprint arXiv:2205.02014 , 2022. Maaten, L. v. d. and Hinton, G. Visualizing data using t-sne. Journal of machine learning research , 9(Nov): 2579–2605, 2008. Manakul, P., Liusie, A., and Gales, M. J. Selfcheckgpt: Zeroresource black-box hallucination detection for generative large language models. arXiv preprint arXiv:2303.08896 , 2023. Meng, K., Bau, D., Andonian, A., and Belinkov, Y. Locating and editing factual associations in gpt. Advances in neural information processing systems , 35:17359–17372, 2022a. Meng, K., Sharma, A. S., Andonian, A., Belinkov, Y., and Bau, D. Mass-editing memory in a transformer. arXiv preprint arXiv:2210.07229 , 2022b. Mitchell, E., Lin, C., Bosselut, A., Finn, C., and Manning, C. D. Fast model editing at scale. arXiv preprint arXiv:2110.11309 , 2021. Mitchell, E., Lin, C., Bosselut, A., Manning, C. D., and Finn, C. Memory-based model editing at scale. In Inter national Conference on Machine Learning , pp. 15817– 15831. PMLR, 2022. Nakano, R., Hilton, J., Balaji, S., Wu, J., Ouyang, L., Kim, C., Hesse, C., Jain, S., Kosaraju, V., Saunders, W., et al. Webgpt: Browser-assisted question-answering with human feedback. arXiv preprint arXiv:2112.09332 , 2021. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al. Language models are unsupervised multitask learners. OpenAI blog , 1(8):9, 2019. Rolnick, D., Ahuja, A., Schwarz, J., Lillicrap, T., and Wayne, G. Experience replay for continual learning. Advances in neural information processing systems , 32, 2019.
+
+[p. 10 | section: References | type: ListGroup]
+Thede, L., Roth, K., Bethge, M., Akata, Z., and Hartvigsen, T. Understanding the limits of lifelong knowledge editing in llms. arXiv preprint arXiv:2503.05683 , 2025. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al. Llama 2: Open foundation and finetuned chat models. arXiv preprint arXiv:2307.09288 , 2023. Valipour, M., Rezagholizadeh, M., Kobyzev, I., and Ghodsi, A. Dylora: Parameter efficient tuning of pre-trained models using dynamic search-free low-rank adaptation. arXiv preprint arXiv:2210.07558 , 2022. Yao, Y., Wang, P., Tian, B., Cheng, S., Li, Z., Deng, S., Chen, H., and Zhang, N. Editing large language models: Problems, methods, and opportunities. arXiv preprint arXiv:2305.13172 , 2023. Yu, L., Chen, Q., Zhou, J., and He, L. Melo: Enhancing model editing with neuron-indexed dynamic lora. In Proceedings of the AAAI Conference on Artificial Intelli gence , pp. 19449–19457, 2024. Zaken, E. B., Ravfogel, S., and Goldberg, Y. Bitfit: Simple parameter-efficient fine-tuning for transformerbased masked language-models. arXiv preprint arXiv:2106.10199 , 2021.

icml26/31f6f2e8-0fb2-46ff-ab65-f3408612f6e1/sanitization_report.json

@@ -0,0 +1,58 @@
+{
+  "paper_id": "31f6f2e8-0fb2-46ff-ab65-f3408612f6e1",
+  "pipeline": "Paper2Markdown-V3",
+  "ok": true,
+  "page_count": 12,
+  "chunk_count": 69,
+  "main_body_chunk_count": 57,
+  "appendix_chunk_count": 7,
+  "reference_chunk_count": 5,
+  "model_text_chars": 37283,
+  "raw_markdown_chars": 58572,
+  "sanitized_chars": 53252,
+  "page_provenance": {
+    "min_page": 1,
+    "max_page": 12,
+    "invalid_count": 0
+  },
+  "marker_block_type_counts": {
+    "Caption": 5,
+    "Equation": 3,
+    "FigureGroup": 5,
+    "Footnote": 1,
+    "ListGroup": 5,
+    "PageFooter": 12,
+    "PageHeader": 15,
+    "SectionHeader": 26,
+    "Table": 1,
+    "TableGroup": 5,
+    "Text": 171
+  },
+  "asset_count_raw": 5,
+  "asset_count_model_kept": 5,
+  "asset_count_rejected": 0,
+  "asset_reject_reasons": {
+    "kept": 5
+  },
+  "artifact_leak_audit": {
+    "ok": true,
+    "hits": {
+      "Anonymous Authors": [],
+      "ACKNOWLEDGMENT": [],
+      "OpenReview": [],
+      "\"accept_label\"": [],
+      "\"decision\"": [],
+      "\"decision_tier\"": [],
+      "\"source_status\"": [],
+      "Meta-review": [],
+      "Official Review": [],
+      "official_reviews": [],
+      "meta_reviews": [],
+      "suggested_verdict_score": []
+    },
+    "artifact_count": 2
+  },
+  "default_model_input": "model_text_v3.txt",
+  "appendix_input": "appendix_text_v3.txt",
+  "reference_input": "reference_text_v3.txt"
+}

icml26/31f6f2e8-0fb2-46ff-ab65-f3408612f6e1/sanitized_v3.txt

@@ -0,0 +1,217 @@
+{0}
+## Abstract
+The dynamic evolution of real-world necessitates model editing within Large Language Models. While existing methods explore modular isolation or parameter-efficient strategies, they still suffer from semantic drift or knowledge forgetting due to continual updating. To address these challenges, we propose SoLA, a Semantic routingbased LoRA framework for lifelong model editing. In SoLA, each edit is encapsulated as an independent LoRA module, which is frozen after training and mapped to input by semantic routing, allowing dynamic activation of LoRA modules via semantic matching. This mechanism avoids semantic drift caused by cluster updating and mitigates catastrophic forgetting from parameter sharing. More importantly, SoLA supports precise revocation of specific edits by removing key from semantic routing, which restores model's original behavior. To our knowledge, this reversible rollback editing capability is the first to be achieved in existing literature. Furthermore, SoLA integrates decision-making process into edited layer, eliminating the need for auxiliary routing networks and enabling end-to-end decision-making process. Extensive experiments demonstrate that SoLA effectively learns and retains edited knowledge, achieving accurate, efficient, and reversible lifelong model editing.
+## 1. Introduction
+Large language models (LLMs) have demonstrated remarkable capabilities in the field of natural language processing, attracting the attention of numerous researchers[\(Radford](#page-9-0) [et al.,](#page-9-0) [2019;](#page-9-0) [Brown et al.,](#page-8-0) [2020;](#page-8-0) [Achiam et al.,](#page-8-1) [2023;](#page-8-1) [Tou](#page-9-1)[vron et al.,](#page-9-1) [2023\)](#page-9-1). However, LLMs also face severe challenges, including hallucinations[\(Huang et al.,](#page-8-2) [2025\)](#page-8-2), bi-
+<span id="page-0-0"></span>![](_page_0_Figure_9.jpeg)
+*Figure 1.* Comparison of trainable parameters and ERR accuracy between SoLA (Ours) and other methods. All experiments are conducted on Scotus datasets with the same backbone.
+ases[\(Hartvigsen et al.,](#page-8-3) [2022\)](#page-8-3), and the generation of harmful content[\(De Cao et al.,](#page-8-4) [2021\)](#page-8-4). Moreover, the dynamic nature of real-world knowledge requires the continuous updating of specific information in LLMs[\(Lazaridou et al.,](#page-8-5) [2021\)](#page-8-5). Re-training these models from scratch is both expensive and time-consuming[\(Biderman et al.,](#page-8-6) [2023\)](#page-8-6). In this scenario, the demand for model editing is increasing with aims to modify specific knowledge in the training model continuously without re-training the model or affecting its performance on unrelated inputs, which is known as lifelong model editing[\(Hartvigsen et al.,](#page-8-3) [2022;](#page-8-3) [Yao et al.,](#page-9-2) [2023;](#page-9-2) [Yu et al.,](#page-9-3) [2024\)](#page-9-3).
+The lifelong model editing aims to continuously update the model to adapt to the constantly changing world. However, most of the existing model editing methods are designed for single, isolated editing[\(Mitchell et al.,](#page-9-4) [2021;](#page-9-4) [Meng et al.,](#page-9-5) [2022a\)](#page-9-5). When applied to a continuous environment, these methods often lead to catastrophic forgetting of the previously learned knowledge in the LLMs, resulting in performance degradation and reduced model reliability[\(Yao](#page-9-2) [et al.,](#page-9-2) [2023;](#page-9-2) [Gu et al.,](#page-8-7) [2024\)](#page-8-7). To solve this problem, recent methods propose freezing the original model parameters and introducing lightweight trainable modules to inject new knowledge. These methods usually retain most of the model's knowledge by keeping the base parameters fixed and combining the learnable modules for specific task updates. For example, Melo[\(Yu et al.,](#page-9-3) [2024\)](#page-9-3) determines the clustering centres of editors by building a neuron index
+{1}------------------------------------------------
+108 109 and dynamically updates its semantic representation based on the distance from the clustering centres. ELDER[\(Li](#page-9-6) [et al.,](#page-9-6) [2025\)](#page-9-6) adopts the Mixture-of-Experts (MoE) approach, assigning scores to each LoRA expert through a set of learnable LoRA allocation codes and selecting the top-k LoRA for calculation.
+<span id="page-1-0"></span>![](_page_1_Figure_3.jpeg)
+*Figure 2.* Number of mismatches in LoRA allocation with ERR/TRR accuracy in MELO. All experiments are conducted on Scotus datasets with the same backbone.
+Although these methods show potential in improving parameter efficiency and reducing interference, they face inherent limitations when applied to lifelong model editing. Specifically, MELO[\(Yu et al.,](#page-9-3) [2024\)](#page-9-3) improves module separability through semantic clustering, but its clustering centers are updated during the editing process, which may lead to semantic drift and module matching errors. ELDER[\(Li et al.,](#page-9-6) [2025\)](#page-9-6) employ MoE and selects the top-k routes to activate each editing subset, although this strategy has high parameter efficiency, shared and continuously updated parameters can lead to catastrophic forgetting - new editing may overwrite or interfere with existing editing.
+To address these limitations, we propose SoLA, a Semantic routing-based LoRA framework for reversible lifelong model editing. In SoLA, we allocate an independent LoRA module for each edit and establish semantic routing to record the mapping relationship between LoRA module and the input semantic representation. During editing, the LoRA module fully learns for the current task and then remains frozen to retain the learned knowledge. The corresponding key is also not updated. During inference, the semantic representation of the input is calculated, and the corresponding LoRA module is dynamically activated through semantic routing. Since the LoRA module and the semantic representation have not been updated, we fundamentally avoid catastrophic forgetting and semantic drift problems caused by continual updating. At the same time, each edit only requires training for the current LoRA module, significantly reducing the computational resource consumption. As shown in Fig[.1,](#page-0-0) SoLA achieves optimal performance with only 0.08M additional parameters, significantly outperforming previous methods, highlighting SoLA's exceptional parameter efficiency and editing accuracy.
+More importantly, by removing the semantic key from the mapping memory table, we can precisely revoke specific edits, allowing the model to recover its original behavior without re-training. This means we can freely add and delete edits, and truly achieve controllable rollback and restoration of editing. To our knowledge, this is the first time that controllable rollback of editing has been achieved in existing literature. Moreover, existing works usually require introducing auxiliary routing network outside the edited layer to determine whether to enable the LoRA module. In this paper, we propose a master decision-making mechanism that aggregates the decision-making process into the edited layer, avoiding the reliance on auxiliary routing network and achieving an end-to-end decision-making process.
+Extensive experiments validate the effectiveness of our approach. Our main contributions are summarized as follows:
+- We propose SoLA, a novel framework for reversible lifelong model editing, which incorporates a controllable LoRA mechanism guided by semantic routing. After each edit, both the LoRA modules and the corresponding keys in the mapping memory are frozen, effectively mitigating catastrophic forgetting and semantic drift. Moreover, only the currently active LoRA modules are trained during editing, significantly reducing computational overhead.(Fig[.1\)](#page-0-0).
+- We employs semantic routing to establish a precise mapping between LoRA modules and semantic representations. By removing the associated key, any specific edit can be precisely revoked, enabling flexible addition and deletion of edits.
+- we propose a master decision-making mechanism that aggregates the decision-making process into the edited layer, avoiding the reliance on an auxiliary routing network and achieving an end-to-end decision-making process.
+## 2. Related Work
+#### 2.1. Model Editing
+Model Editing aims to update specific knowledge within LLMs while preserving their previous knowledge. Current model editing methods can be broadly categorized into three paradigms: Meta-Learning methods, Locate-then-Edit methods, and Memory-Based methods. Meta-Learning Methods employ a hypernetwork to predict the necessary gradients for editing and apply these gradients to the base model to achieve the update[\(Mitchell et al.,](#page-9-4) [2021\)](#page-9-4). However, this
+{2}------------------------------------------------
+<span id="page-2-0"></span>![](_page_2_Figure_1.jpeg)
+*Figure 3.* Main framework of our method. a) indicates the edited layer in model, where we retain the transformer block of base model frozen and add trainable LoRA module to the edited layer of base model. b) shows the editing process, where every edit will be assigned LoRA module and the query of input will be mapped to assigned LoRA module. c) presents the matching process between query of input and LoRA module in reference. Colour green indicates trainable, while color grey is frozen.
+approach typically requires additional data for training the hypernetwork. Locate-then-Edit methods first identify the neurons most critical for the current input via minor perturbations and then directly modify these neurons to update the model's knowledge[\(Meng et al.,](#page-9-5) [2022a;](#page-9-5)[b\)](#page-9-7). While effective, these methods are often cumbersome and exhibit limitations when handling a large volume of simultaneous knowledge updates. Memory-Based methods store the information associated with edits in an buffer, which acts as a patch model augmenting the original model[\(Mitchell et al.,](#page-9-8) [2022\)](#page-9-8). A significant drawback is that each edit necessitates retraining, making it difficult to adapt to continuous editing scenarios, and the memory bank grows incrementally with each edit, leading to escalating storage overhead. Recent studies have also explored editing constraints to maintain model behavior stability. AlphaEdit[\(Fang et al.,](#page-8-8) [2024\)](#page-8-8) introduces a null-space constrained approach. It aims to apply edits more precisely by constraining changes to the null space of irrelevant knowledge.
+However, these methods primarily address static editing, suitable for one-time modifications, and struggle to accommodate sequential editing demands over time. To address these limitations, GRACE[\(Hartvigsen et al.,](#page-8-9) [2023\)](#page-8-9) replace hidden states with vectors retrieved from a learned codebook. MELO[\(Yu et al.,](#page-9-3) [2024\)](#page-9-3) introduces a vector database and leverages a clustering mechanism to dynamically assign LoRA modules. ELDER[\(Li et al.,](#page-9-6) [2025\)](#page-9-6) adopts a MoE framework, where a learnable neural network weights shared LoRA modules. However, these approaches inevitably update the semantic representation in sequential edits, leading to semantic drift or knowledge forgetting.
+## 3. Method
+## 3.1. Preliminaries
+LoRA for Lifelong Model Editing Low-Rank Adaptation (LoRA) [\(Hu et al.,](#page-8-10) [2022\)](#page-8-10) is an efficient finetuning method for LLMs that optimizes model outputs by projecting features into a low-rank subspace. In the context of efficient LLM finetuning, LoRA modules are typically inserted into
+{3}------------------------------------------------
+specific layers of the pre-trained model. During editing, the base LLM parameters $W_0 \in \mathbb{R}^{d \times k}$ remain frozen, while only the LoRA module $\Delta W \in \mathbb{R}^{d \times k}$ is updated. The parameter update $\Delta W$ can be represented as a low-rank decomposition of the pretrained weight matrix. Specifically, $\Delta W$ is factorized into two smaller matrices $A \in \mathbb{R}^{r \times k}$ and $B \in \mathbb{R}^{d \times r}$ , such that $\Delta W = BA$ , where $r \ll \min(d,k)$ and r is the LoRA rank. Given the original $h = W_0 x$ , the forward propagation process is modified as:
+$$\boldsymbol{h} = W_0 \boldsymbol{x} + \Delta W \boldsymbol{x} = W_0 \boldsymbol{x} + B A \boldsymbol{x} \tag{1}$$
+where A is initialized using a zero mean Gaussian distribution and B is initialized as a zero matrix. This approach significantly reduces the number of trainable parameters while maintaining the expressive power of the full-rank update.
+#### 3.2. Cluster Updating in Lifelong Model Editing
+In the context of lifelong model editing, when multiple adapter modules are involved, the model needs to dynamically manage the association between edits and the corresponding adapter module. MELO addresses this by employing a clustering mechanism that assigns edits to clusters based on the hidden representations of the inputs, with cluster centres being continuously updated throughout editing. However, this continual updating of cluster centers alters their semantic representation, leading to semantic drift and incorrect module assignments. Moreover, repeated training of LoRA modules can result in forgetting previously acquired knowledge. Varying the cluster radius in MeLO results in a different number of generated cluster centres, thereby affecting the frequency of cluster centre updates. We record MeLO's performance under different numbers of cluster centre updates, as shown in Fig.2. The experiment is conducted on SCOTUS dataset. In Fig.2, the number of updates refers to the number of cluster centre updates during the editing process, and the number of mismatches indicates how many times, during inference, the retrieved LoRA module differs from the one assigned during editing. ERR and TRR represent the model's accuracy on the edited and unedited datasets, respectively, after all editing tasks are completed.
+As shown in the Fig.2, for the same datasets SCOTUS of editing, increasing the number of cluster centre updates leads to a rise in mismatches during inference, and a corresponding decline in model performance on both the edited and unedited datasets. This demonstrates that cluster centre updates can cause semantic drift, resulting in incorrect LoRA match, while repeated training of LoRA modules contributes to knowledge forgetting. Therefore, in SoLA, we freeze both the LoRA modules and their associated keys once the current task is completed. These components are
+no longer updated in subsequent edits, thereby maximizing knowledge retention across editing iterations.
+#### 3.3. Semantic Routing-Based Controllable LoRA
+In the layer where the model needs to be edited, we keep the parameters of the original model frozen and insert the learnable LoRA module for editing, as shown in Fig.3 a). In SoLA, each editing operation is encapsulated within an independent LoRA module. During sequential editing, for a given editing task $d_i = (\boldsymbol{x}_i^m, \boldsymbol{y}_i^m)_{m=1}^n$ , a dedicated LoRA module LoRA<sub>i</sub> is assigned. For each data instance $\boldsymbol{x}_i^m \in d_i$ , the corresponding LoRA id i is recorded, and a semantic routing entry is established by associating LoRA module with a semantic key derived from the input representation, as shown in Fig.3 b). Following prior work(Hartvigsen et al., 2023), we use the hidden representation of the last token in the input sequence as the key $e \in \mathbb{R}^d$ vector. During editing, the assigned LoRA module LoRA<sub>i</sub> will fully learn the task $d_i$ , yielding the output:
+$$h = h_0 + LoRA_i(x)$$
+(2)
+where $h_0 \in \mathbb{R}^{l \times d}$ is the frozen base model representation, l is the sequence length, d is the hidden state dimension. At this stage, all other LoRA modules and the stored key vectors remain frozen. Upon completion of the editing process, the trained LoRA<sub>i</sub> module is frozen and will be stored with its associated key $e_i$ as a fixed mapping for future reference. Neither the LoRA module nor the key will be updated in subsequent editing. Since only the current LoRA module is involved in training during each edit, SoLA significantly reduces the number of trainable parameters, as shown in Fig.1. Furthermore, freezing both the LoRA module and its corresponding key after editing prevents semantic drift and mitigates catastrophic forgetting, thereby maximizing knowledge retention. During inference, the hidden representation of the last token in the input sequence will serve as a query vector $\boldsymbol{q} \in \mathbb{R}^d$ from the input $\boldsymbol{x}$ and matches against the stored keys $K \in \mathbb{R}^{N \times d}$ , as shown in Fig.3 c). If a match is found, the corresponding LoRA module is retrieved from stored LoRA pool $R \in \mathbb{R}^M$ and incorporated into the computation.
+#### 3.4. Master Decision Mechanism
+During inference, existing approaches usually need to introduce an auxiliary routing network outside the target edited layer to decide whether to activate the LoRA module or not(Yu et al., 2024; Li et al., 2025), which hampers the end-to-end decision-making process. To address this limitation, we propose a master decision-making mechanism that does not require an auxiliary routing component. Specifically, we designate the first edited layer as the master decision layer where input features are dynamically evaluated to activate the relevant LoRA module without the need for an auxil-
+{4}------------------------------------------------
+<span id="page-4-0"></span>*Table 1.* Comparison of SoLA to existing methods. All the results are obtained after sequential edits. ERR indicates Edit Reliability Rate, TRR presents Task Retention Rate, and ARR is Accuracy Attention Rate. The best performance is shown in bold.
+| | SCOTUS (BERT; Acc<br>↑) | | | zsRE (T5; F1<br>↑) | | | Hallucination (GPT2-XL; PPL<br>↓) | | |
+|--------|-------------------------|------|------|--------------------|------|------|-----------------------------------|--------|--------|
+| Method | ERR | TRR | Avg. | ERR | TRR | Avg. | ERR | TRR | ARR |
+| EWC | 0.51 | 0.82 | 0.66 | 0.67 | 0.50 | 0.58 | 1485.7 | 29.24 | 109.59 |
+| CMR | 0.52 | 0.52 | 0.52 | 0.56 | 0.82 | 0.69 | 1449.3 | 28.14 | 107.76 |
+| CLEAR | 0.67 | 0.83 | 0.75 | 0.27 | 0.99 | 0.63 | 2394.3 | 35.34 | 195.82 |
+| MEND | 0.19 | 0.27 | 0.23 | 0.25 | 0.27 | 0.26 | 1369.8 | 1754.9 | 2902.5 |
+| SERAC | 0.33 | 0.41 | 0.37 | 0.72 | 0.31 | 0.52 | 8183.7 | 133.3 | 10.04 |
+| ROME | - | - | - | - | - | - | 30.28 | 103.82 | 14.02 |
+| GRACE | 0.81 | 0.82 | 0.82 | 0.69 | 0.96 | 0.83 | 15.84 | 7.14 | 10.00 |
+| ELDER | 0.89 | 0.86 | 0.88 | 0.72 | 0.97 | 0.84 | 16.12 | 5.87 | 8.42 |
+| MELO | 0.96 | 0.92 | 0.94 | 0.72 | 0.98 | 0.85 | 17.45 | 1.04 | 2.66 |
+| SoLA | 0.97 | 0.95 | 0.96 | 0.73 | 0.99 | 0.86 | 15.15 | 1.01 | 7.35 |
+iary network. In this framework, the master decision layer H<sup>e</sup> computes the distance metric between the input feature embedding q and the stored key K to determine the activation of the module: d = He(argmini(dist(q, ki))), i = (1, 2, ..., N), dist(·) denotes the distance function and each key k<sup>i</sup> associated with the LoRA module. Then in the master layer He, we decide the edit layer behaviour:
+$$H_e = \begin{cases} H(W_0) & \text{if } d \ge \alpha \\ H(W_0, W_{R_m}) & \text{if } d < \alpha \end{cases}$$
+(3)
+where α is the threshold, in this work we set it as 0.01, WR<sup>m</sup> represents the weight of m-th LoRA module associated with the key nearest to the query. This binary decision is propagated to the subsequent edited layers, ensuring consistent module activation throughout the network. By integrating the decision-making process to the first edited layer, our approach achieves complete end-to-end decision-making capability while maintaining architectural simplicity.
+## 4. Experiments
+#### 4.1. Implementation Details
+Baselines We first compare several recent methods designed for lifelong model editing. Grace[\(Hartvigsen et al.,](#page-8-9) [2023\)](#page-8-9) replaces model outputs with key-value pairs and matches features to keys based on deferral radius. MELO[\(Yu et al.,](#page-9-3) [2024\)](#page-9-3) leverages LoRA for finetuning and dynamically retrieves relevant LoRA modules via neuron index. EL-DER[\(Li et al.,](#page-9-6) [2025\)](#page-9-6) employs MoE to dynamically combine multiple LoRA modules. Additionally, we evaluate other baseline methods. MEND[\(Mitchell et al.,](#page-9-4) [2021\)](#page-9-4) uses a hypernetwork trained on auxiliary data to predict the required gradients for model editing. SERAC[\(Mitchell et al.,](#page-9-8) [2022\)](#page-9-8)
+learns a scope classifier and a counterfactual model, coordinating between modules to determine editing behavior. ROME[\(Meng et al.,](#page-9-5) [2022a\)](#page-9-5) identifies the most influential weights in the model through perturbation analysis, localizes knowledge to specific layers of GPT, and modifies the corresponding weights. Since ROME is specifically designed for GPT, it is only evaluated on the hallucination correction task. EWC[\(Kirkpatrick et al.,](#page-8-11) [2017\)](#page-8-11) imposes constraints on the weights updating so that the model continuously learns new knowledge and retains previous knowledge. CMR[\(Lin](#page-9-9) [et al.,](#page-9-9) [2022\)](#page-9-9) performs continuous finetuning of the model to continuously learn new knowledge. CLEAR[\(Rolnick et al.,](#page-9-10) [2019\)](#page-9-10) builds a memory buffer to restore old tasks information, and replays them in subsequent edits to retain previous knowledge.
+Evaluation Metric In this work, following[\(Hartvigsen et al.,](#page-8-9) [2023\)](#page-8-9), we use three basic metrics to assess model performance: ES, ERR and TRR, which are applicable to all model experiments. Where Edit Success(ES) measures the ability of the model to edit the target sample. Edit Reliability Rate(ERR) assesses the validity of the model on the edited dataset, indicating the extent to which the required knowledge updates are successfully incorporated. Task Retention Rate(TRR) quantifies the model's performance on unedited datasets and reflects the model's ability to retain previous knowledge. Additionally, in the hallucination correction task, we also use the Accuracy Attention Rate (ARR) to assess the performance of the already accurate output. Accuracy is measured differently for different tasks. For the document classification task of SCOTUS, we use the Average Accuracy Rate (ACC); for the question answering task of zsRE, we measure it using the average F1, and for the hallucination correction task, we measure it using the standard average complexity (PPL)[\(Brown et al.,](#page-8-12) [1992\)](#page-8-12).
+{5}------------------------------------------------
+<span id="page-5-0"></span>![](_page_5_Figure_2.jpeg)
+*Figure 4.* Figure(a)–(c) presents every edit accuracy on SCOTUS datasets with BERT. All experiment settings are the same as Tab[.1.](#page-4-0)
+<span id="page-5-1"></span>![](_page_5_Figure_4.jpeg)
+*Figure 5.* Visualization of zsRE dataset encoder output from T5 small with t-SNE and the dots with same color and shape are input and rephrase sentence. All experiment settings are the same as Tab[.1.](#page-4-0)
+#### 4.2. Main Results
+We evaluate the performance of the proposed SoLA method on three benchmark datasets and compare it with several state-of-the-art model editing approaches. The results are summarized in Tab[.1.](#page-4-0) As observed, SoLA achieves excellent performance on most of the benchmark datasets, with SoLA outperforming the strongest baseline, MELO, by 3% on the SCOTUS dataset. These demonstrate SoLA's enhanced effectiveness in editing target knowledge while better preserving previously acquired knowledge. Furthermore, existing methods such as MEND and SERAC perform poorly under the sequential editing setting, suggesting that they are primarily designed for static editing scenarios and are not well-suited for sequential model editing tasks.
+We also report task-wise accuracy progression to analyze model performance throughout the sequential editing process. The experimental results are based on the SCOTUS dataset and the results are shown in Fig[.4.](#page-5-0) As illustrated in Fig[.4a,](#page-5-0) Fig[.4b](#page-5-0) and Fig[.4c,](#page-5-0) SoLA consistently maintains
+superior accuracy across tasks, without exhibiting substantial performance degradation or volatility. This indicates the robustness and stability of SoLA when applied to sequential knowledge editing.
+Method performance may vary across different model sizes and datasets. To further evaluate the robustness of our approach, we conduct hallucination correction experiments on the UniEdit[\(Chen et al.,](#page-8-13) [2025\)](#page-8-13) and WikiBigEdit[\(Thede](#page-9-11) [et al.,](#page-9-11) [2025\)](#page-9-11) datasets using larger-scale backbone models. The results are presented in Tab[.2.](#page-6-0) As shown in Tab[.2,](#page-6-0) SoLA consistently achieves the best or near-best performance across most settings, demonstrating its robustness and effectiveness under diverse task configurations. During editing, larger backbone models generally exhibit more stable ERR values, suggesting that such models have learned stronger semantic representations during pretraining, leading to better generalization and editability. Furthermore, continuous learning related methods tend to overfit in edited data, thus showing better ARR in the retention set, but large TRR values in the upstream set, indicating severe forgetting of previous knowledge.
+#### 4.3. Controllable Rollback Editing
+In SoLA, each edit instance is associated with a unique key stored in both the memory and the routing table. This design enables reversible editing, as the model can revert to its original behavior by removing the key corresponding to a specific edit. To demonstrate this property, we conduct an illustrative experiment on the zsRE dataset. The results are shown in Tab[.3.](#page-7-0) In the Tab[.3,](#page-7-0) each row corresponds to an input instance consisting of a "Text" (the question need to be edited) and its "Labels" (the correct answer). "Del" indicates whether the key corresponding to the edit is removed after editing. For each input, Predbase, Prededit and Preddel denote the prediction of base model, edited model and model after removing the corresponding key. When "Del" is false, the key is retained, serving as a baseline comparison. As shown in Tab[.3,](#page-7-0) for each instance, the original prediction (Predbase) does not match the correct label, indi-
+{6}------------------------------------------------
+<span id="page-6-0"></span>*Table 2.* Model Performance Comparison on UniEdit and WikiBigEdit datasets in hallucination correction tasks. All the scores are PPL metric and the best result is in bolded.
+| | | | UniEdit (PPL ↓) | | | WikiBigEdit (PPL ↓) | |
+|----------------|--------|-------|-----------------|-------|-------|---------------------|-------|
+| Model | Method | ERR | TRR | ARR | ERR | TRR | ARR |
+| | EWC | 2.43 | 1134 | 22.08 | 3.58 | 251117 | 4.93 |
+| | CMR | 2.79 | 2239 | 25.24 | 2.86 | 169452 | 4.49 |
+| LLaMA-3-8B | CLEAR | 1.04 | 982 | 29.79 | 1.03 | 74384 | 1.52 |
+| | SERAC | 37.10 | 246 | 93 | 60.17 | 215 | 59 |
+| | GRACE | 1.00 | 244 | 89 | 1.00 | 216 | 58 |
+| | ELDER | 1.00 | 173 | 74 | 1.00 | 178 | 48 |
+| | MELO | 1.00 | 153 | 68 | 1.00 | 165 | 47 |
+| | Ours | 1.00 | 144 | 66 | 1.00 | 162 | 44 |
+| | EWC | 2.38 | 4451 | 42.13 | 2.86 | 47846 | 3.89 |
+| | CMR | 2.34 | 7211 | 36.19 | 3.60 | 34623 | 5.51 |
+| DeepSeek-R1-8B | CLEAR | 1.06 | 3365 | 44.86 | 1.02 | 28110 | 1.48 |
+| | SERAC | 46.76 | 451 | 834 | 82.72 | 493 | 831 |
+| | GRACE | 1.00 | 407 | 803 | 1.00 | 487 | 793 |
+| | ELDER | 1.03 | 241 | 482 | 1.03 | 304 | 492 |
+| | MELO | 1.09 | 204 | 407 | 1.05 | 263 | 445 |
+| | Ours | 1.07 | 188 | 374 | 1.00 | 248 | 412 |
+| | EWC | 4.62 | 7648 | 48.62 | 5.61 | 61834 | 14.65 |
+| | CMR | 4.12 | 11583 | 40.56 | 6.28 | 48629 | 16.89 |
+| Qwen2-7B | CLEAR | 1.03 | 5021 | 54.98 | 1.05 | 35396 | 4.26 |
+| | SERAC | 58.51 | 329 | 1261 | 62.12 | 304 | 523 |
+| | GRACE | 2.05 | 316 | 1065 | 16.88 | 299 | 518 |
+| | ELDER | 1.00 | 201 | 266 | 1.00 | 225 | 197 |
+| | MELO | 1.00 | 170 | 201 | 1.00 | 203 | 143 |
+| | Ours | 1.00 | 156 | 110 | 1.00 | 199 | 109 |
+cating the necessity for editing. After editing, the model correctly produces Prededit consistent with the ground-truth label, demonstrating that SoLA successfully updates the model's knowledge. When the associated key is deleted, the model reverts to its original prediction (Prededit = Predbase), verifying that our method can effectively roll back specific edits. More critically, for edits where the key is not deleted, the model continues to output Prededit, showing that the rollback operation does not interfere with other edits. This confirms that SoLA supports fine-grained, selective undo of knowledge edits without disrupting unrelated modifications.
+#### 4.4. Ablation Study
+Effect of Edit Layer Location In semantic representation learning, different layers of a model focus on different aspects of semantic information. Therefore, the position of the edited layer can significantly influence the effectiveness of model editing. To investigate this, we evaluate the performance of model when edits are applied at different layers, while keeping all other settings fixed. The experiments are conducted on the SCOTUS dataset with a Nvidia A100 40G GPU, and the results are shown in Tab[.4.](#page-7-1) In the
+Tab[.4,](#page-7-1) "Layer" denotes the layer range used for editing. For example, "0–2" indicates that edits are applied to layers 0, 1, and 2 of the BERT model. As shown in the table, editing at shallower layers results in suboptimal performance. This observation aligns with prior findings[\(Geva et al.,](#page-8-14) [2020\)](#page-8-14), which suggest that shallow layers primarily capture the shallow sentence patterns, whereas deeper layers encode richer semantic features, thereby enabling more effective knowledge editing. Moreover, editing at shallow layers significantly increases the time used for editing, which means that editing for semantics is more difficult in shallow layers. Furthermore, we observe that varying the editing layer has negligible impact on performance over the unedited portion of the dataset, indicating that SoLA is capable of preserving previously learned knowledge regardless of the editing depth, which further supports the robustness of SoLA.
+Effect of LoRA Rank The LoRA module projects semantic representations into a low-rank subspace for parameterefficient learning. Consequently, the choice of rank will effect the module's performance. To investigate this, we analyze the impact of different LoRA rank values on model performance, keeping all other settings fixed. Experiments are conducted on the SCOTUS dataset using a NVIDIA
+{7}------------------------------------------------
+<span id="page-7-0"></span>*Table 3.* Controllable edit in SoLA on zsRE datasets. "Text" is the question need to be edited, "Labels" is the true answer, "Del" indicates whther delete associated key and Predbase, Prededit and Preddel is the prediction of base model, edited model and delted model.
+| Text | Labels | Del | Predbase | Prededit | Preddel |
+|-----------------------------------------------------|-------------------|-------|------------------|-------------------|------------------|
+| The date of Tsetserleg earthquake in 1905? | 9 July 1905 | True | 20 February 1905 | 9 July 1905 | 20 February 1905 |
+| What constellation has Nu Cancri? | Cancer | True | Hydra | Cancer | Hydra |
+| On what date did the Bahrain Grand Prix 2015 occur? | 19 April 2015 | True | 6 April 2017 | 19 April 2015 | 6 April 2017 |
+| What position does Charles-Joseph Coursol have? | Mayor of Montreal | True | Positioning | Mayor of Montreal | Positioning |
+| Who acted in Blow Dry? | Alan Rickman | False | Jim Carrey | Alan Rickman | Alan Rickman |
+<span id="page-7-1"></span>*Table 4.* Ablation studies on location of edited layers. Here, "Layer" indicates the edited layer location. For example, "0-2" presents edited layer is layers 0, 1 and 2.
+| Layer | ES | ERR | TRR | Edit Time(min) |
+|-------|------|------|------|----------------|
+| 0-2 | 0.77 | 0.61 | 0.96 | 9.21 |
+| 3-5 | 1.00 | 0.68 | 0.96 | 8.06 |
+| 6-8 | 0.81 | 0.70 | 0.97 | 7.03 |
+| 9-11 | 0.94 | 0.95 | 0.97 | 5.97 |
+<span id="page-7-2"></span>*Table 5.* Ablation studies on LoRA rank. Here, "LoRA Rank" indicates the rank value of every LoRA in model.
+| LoRA Rank | ES | ERR | TRR | Edit Time(min) |
+|-----------|------|------|------|----------------|
+| 1 | 0.90 | 0.84 | 0.96 | 5.88 |
+| 2 | 0.94 | 0.91 | 0.97 | 5.90 |
+| 3 | 1.00 | 0.91 | 0.97 | 5.86 |
+| 4 | 1.00 | 0.95 | 0.97 | 5.97 |
+| 5 | 1.00 | 0.92 | 0.96 | 5.86 |
+| 10 | 1.00 | 0.71 | 0.96 | 5.85 |
+| | | | | |
+A100 40GB GPU, and the results are presented in Tab[.5.](#page-7-2) As shown in the Tab[.5,](#page-7-2) simply increasing the LoRA rank does not lead to improved performance and may even degrade model performance. This indicates that expanding the learnable parameter space does not necessarily enhance the model's capacity for effective knowledge editing. On the contrary, excessive rank values may even result in performance degradation due to overfitting, suggesting the importance of carefully selecting an appropriate rank to balance capacity and generalization. Based on the experiment results, we set 4 as LoRA rank for all the experiments in this paper.
+#### 4.5. Model Visualization
+To gain deeper insight into the model's behavior under sequential editing, we conducted a t-SNE visualization [\(Maaten & Hinton,](#page-9-12) [2008\)](#page-9-12) of the learned feature representations. This experiment is performed on the zsRE dataset using the T5-small model to assess how semantic relationships are preserved after sequential edits. Upon completing
+the entire sequence of editing tasks for question answering, we selected five distinct input instances, each accompanied by semantically equivalent rephrased sentences, forming five evaluation groups. The encoder output features are subsequently extracted from the edited model for dimensionality reduction and visualization, and the resulting plot is represented in Fig[.5.](#page-5-1) As shown in Fig[.5,](#page-5-1) there is a clear clustering pattern that for each group, the original input and its rephrased variant are mapped to proximate locations in the latent space. This demonstrates that SoLA effectively captures and preserves the semantic similarity between related inputs throughout the editing process. Furthermore, distinct groups maintain separable clusters, indicating that the model retains the ability to discriminate between different edit concepts after sequential editing. This visualization shows that SoLA's semantic routing mechanism successfully supports edit generalization based on semantic similarity while upholding the integrity of previously learned knowledge.
+## 5. Conclusion
+In this paper, we propose SoLA, a Semantic routing-based LoRA framework for reversible lifelong model editing. In SoLA, each edit is encapsulated as an independent LoRA module which is frozen after training and a semantic routing record is established to map LoRA module to the input semantic representation, allowing dynamic activation of LoRA modules via semantic matching. This design not only avoids semantic drift caused by clustering updating but also mitigates catastrophic forgetting from parameter sharing. More importantly, SoLA supports precisely revoking edit by removing the corresponding key from the routing table, enabling reversible model editing. This allows for flexible addition and deletion of edits, offering fine-grained control over model behavior. To the best of our knowledge, this is the first work to achieve reversible model editing in the literature. Furthermore, we introduce a master decisionmaking mechanism by integrating decision-making into the edited layer, enabling end-to-end decision-making process. By building a strict mapping between edits and LoRA modules, SoLA achieves efficient and reversible lifelong model editing, providing a novel perspective for future research.
+{8}------------------------------------------------
+## Impact Statement
+This work advances the field of Machine Learning by introducing SoLA, a framework for reversible and controllable lifelong model editing in large language models. Our method enhances the safety and reliability of AI systems by enabling precise rollback of model edits, which helps mitigate risks associated with erroneous or harmful knowledge updates. By effectively preventing semantic drift and catastrophic forgetting, SoLA promotes the development of more robust and trustworthy models capable of adapting to evolving information. The significant reduction in computational overhead also aligns with the goal of sustainable AI development.
+## References
+- <span id="page-8-1"></span>Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al. Gpt-4 technical report. *arXiv preprint arXiv:2303.08774*, 2023.
+- <span id="page-8-6"></span>Biderman, S., Schoelkopf, H., Anthony, Q. G., Bradley, H., O'Brien, K., Hallahan, E., Khan, M. A., Purohit, S., Prashanth, U. S., Raff, E., et al. Pythia: A suite for analyzing large language models across training and scaling. In *International Conference on Machine Learning*, pp. 2397–2430. PMLR, 2023.
+- <span id="page-8-12"></span>Brown, P. F., Della Pietra, S. A., Della Pietra, V. J., Lai, J. C., and Mercer, R. L. An estimate of an upper bound for the entropy of english. *Computational Linguistics*, 18 (1):31–40, 1992.
+- <span id="page-8-0"></span>Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al. Language models are few-shot learners. *Advances in neural information processing systems*, 33: 1877–1901, 2020.
+- <span id="page-8-17"></span>Chalkidis, I., Pasini, T., Zhang, S., Tomada, L., Schwemer, S. F., and Søgaard, A. Fairlex: A multilingual benchmark for evaluating fairness in legal text processing. *arXiv preprint arXiv:2203.07228*, 2022.
+- <span id="page-8-13"></span>Chen, Q., Wang, D., Zhang, T., Yan, Z., You, C., Wang, C., and He, X. Uniedit: A unified knowledge editing benchmark for large language models. *arXiv preprint arXiv:2505.12345*, 2025.
+- <span id="page-8-4"></span>De Cao, N., Aziz, W., and Titov, I. Editing factual knowledge in language models. *arXiv preprint arXiv:2104.08164*, 2021.
+- <span id="page-8-8"></span>Fang, J., Jiang, H., Wang, K., Ma, Y., Jie, S., Wang, X., He, X., and Chua, T.-S. Alphaedit: Null-space constrained knowledge editing for language models. *arXiv preprint arXiv:2410.02355*, 2024.
+- <span id="page-8-14"></span>Geva, M., Schuster, R., Berant, J., and Levy, O. Transformer feed-forward layers are key-value memories. *arXiv preprint arXiv:2012.14913*, 2020.
+- <span id="page-8-7"></span>Gu, J.-C., Xu, H.-X., Ma, J.-Y., Lu, P., Ling, Z.-H., Chang, K.-W., and Peng, N. Model editing can hurt general abilities of large language models. *CoRR*, 2024.
+- <span id="page-8-3"></span>Hartvigsen, T., Gabriel, S., Palangi, H., Sap, M., Ray, D., and Kamar, E. Toxigen: A large-scale machine-generated dataset for adversarial and implicit hate speech detection. *arXiv preprint arXiv:2203.09509*, 2022.
+- <span id="page-8-9"></span>Hartvigsen, T., Sankaranarayanan, S., Palangi, H., Kim, Y., and Ghassemi, M. Aging with grace: Lifelong model editing with discrete key-value adaptors. *Advances in Neural Information Processing Systems*, 36:47934–47959, 2023.
+- <span id="page-8-15"></span>Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., and Gelly, S. Parameter-efficient transfer learning for nlp. In *International conference on machine learning*, pp. 2790– 2799. PMLR, 2019.
+- <span id="page-8-10"></span>Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al. Lora: Low-rank adaptation of large language models. *ICLR*, 1(2):3, 2022.
+- <span id="page-8-2"></span>Huang, L., Yu, W., Ma, W., Zhong, W., Feng, Z., Wang, H., Chen, Q., Peng, W., Feng, X., Qin, B., et al. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. *ACM Transactions on Information Systems*, 43(2):1–55, 2025.
+- <span id="page-8-16"></span>Jia, M., Tang, L., Chen, B.-C., Cardie, C., Belongie, S., Hariharan, B., and Lim, S.-N. Visual prompt tuning. In *European conference on computer vision*, pp. 709–727. Springer, 2022.
+- <span id="page-8-11"></span>Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A. A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al. Overcoming catastrophic forgetting in neural networks. *Proceedings of the national academy of sciences*, 114(13):3521–3526, 2017.
+- <span id="page-8-18"></span>Kwiatkowski, T., Palomaki, J., Redfield, O., Collins, M., Parikh, A., Alberti, C., Epstein, D., Polosukhin, I., Devlin, J., Lee, K., et al. Natural questions: a benchmark for question answering research. *Transactions of the Association for Computational Linguistics*, 7:453–466, 2019.
+- <span id="page-8-5"></span>Lazaridou, A., Kuncoro, A., Gribovskaya, E., Agrawal, D., Liska, A., Terzi, T., Gimenez, M., de Masson d'Autume, C., Kocisky, T., Ruder, S., et al. Mind the gap: Assessing temporal generalization in neural language models. *Advances in Neural Information Processing Systems*, 34: 29348–29363, 2021.
+{9}------------------------------------------------
+<span id="page-9-16"></span>495 496 497 Levy, O., Seo, M., Choi, E., and Zettlemoyer, L. Zero-shot relation extraction via reading comprehension. *arXiv preprint arXiv:1706.04115*, 2017.
+<span id="page-9-14"></span>504
+<span id="page-9-12"></span>513 514
+<span id="page-9-17"></span>516
+<span id="page-9-7"></span>524 525 526
+<span id="page-9-4"></span>528 529 530
+<span id="page-9-18"></span>536
+- <span id="page-9-6"></span>Li, J., Wang, Q., Wang, Z., Zhang, Y., and Mao, Z. Elder: Enhancing lifelong model editing with mixture-of-lora. In *Proceedings of the AAAI Conference on Artificial Intelligence*, pp. 24440–24448, 2025.
+- Li, X. L. and Liang, P. Prefix-tuning: Optimizing continuous prompts for generation. *arXiv preprint arXiv:2101.00190*, 2021.
+- <span id="page-9-9"></span>Lin, B. Y., Wang, S., Lin, X. V., Jia, R., Xiao, L., Ren, X., and Yih, W.-t. On continual model refinement in out-of-distribution data streams. *arXiv preprint arXiv:2205.02014*, 2022.
+- Maaten, L. v. d. and Hinton, G. Visualizing data using t-sne. *Journal of machine learning research*, 9(Nov): 2579–2605, 2008.
+- Manakul, P., Liusie, A., and Gales, M. J. Selfcheckgpt: Zeroresource black-box hallucination detection for generative large language models. *arXiv preprint arXiv:2303.08896*, 2023.
+- <span id="page-9-5"></span>Meng, K., Bau, D., Andonian, A., and Belinkov, Y. Locating and editing factual associations in gpt. *Advances in neural information processing systems*, 35:17359–17372, 2022a.
+- Meng, K., Sharma, A. S., Andonian, A., Belinkov, Y., and Bau, D. Mass-editing memory in a transformer. *arXiv preprint arXiv:2210.07229*, 2022b.
+- Mitchell, E., Lin, C., Bosselut, A., Finn, C., and Manning, C. D. Fast model editing at scale. *arXiv preprint arXiv:2110.11309*, 2021.
+- <span id="page-9-8"></span>Mitchell, E., Lin, C., Bosselut, A., Manning, C. D., and Finn, C. Memory-based model editing at scale. In *International Conference on Machine Learning*, pp. 15817– 15831. PMLR, 2022.
+- Nakano, R., Hilton, J., Balaji, S., Wu, J., Ouyang, L., Kim, C., Hesse, C., Jain, S., Kosaraju, V., Saunders, W., et al. Webgpt: Browser-assisted question-answering with human feedback. *arXiv preprint arXiv:2112.09332*, 2021.
+- <span id="page-9-0"></span>Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al. Language models are unsupervised multitask learners. *OpenAI blog*, 1(8):9, 2019.
+- <span id="page-9-10"></span>Rolnick, D., Ahuja, A., Schwarz, J., Lillicrap, T., and Wayne, G. Experience replay for continual learning. *Advances in neural information processing systems*, 32, 2019.
+- <span id="page-9-11"></span>Thede, L., Roth, K., Bethge, M., Akata, Z., and Hartvigsen, T. Understanding the limits of lifelong knowledge editing in llms. *arXiv preprint arXiv:2503.05683*, 2025.
+- <span id="page-9-1"></span>Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al. Llama 2: Open foundation and finetuned chat models. *arXiv preprint arXiv:2307.09288*, 2023.
+- <span id="page-9-15"></span>Valipour, M., Rezagholizadeh, M., Kobyzev, I., and Ghodsi, A. Dylora: Parameter efficient tuning of pre-trained models using dynamic search-free low-rank adaptation. *arXiv preprint arXiv:2210.07558*, 2022.
+- <span id="page-9-2"></span>Yao, Y., Wang, P., Tian, B., Cheng, S., Li, Z., Deng, S., Chen, H., and Zhang, N. Editing large language models: Problems, methods, and opportunities. *arXiv preprint arXiv:2305.13172*, 2023.
+- <span id="page-9-3"></span>Yu, L., Chen, Q., Zhou, J., and He, L. Melo: Enhancing model editing with neuron-indexed dynamic lora. In *Proceedings of the AAAI Conference on Artificial Intelligence*, pp. 19449–19457, 2024.
+- <span id="page-9-13"></span>Zaken, E. B., Ravfogel, S., and Goldberg, Y. Bitfit: Simple parameter-efficient fine-tuning for transformerbased masked language-models. *arXiv preprint arXiv:2106.10199*, 2021.
+{10}------------------------------------------------
+# Appendix
+## A. Additional Related Work
+#### A.1. Parameter-Efficient Fine-Tuning
+Parameter-efficient fine-tuning (PEFT) approaches adjust pretrained LLMs by incorporating lightweight modules, while keeping the base model frozen. Optimization is performed solely through updates to these lightweight components. Representative approaches include Adapters[\(Houlsby](#page-8-15) [et al.,](#page-8-15) [2019;](#page-8-15) [Zaken et al.,](#page-9-13) [2021\)](#page-9-13), Prompts[\(Li & Liang,](#page-9-14) [2021;](#page-9-14) [Jia et al.,](#page-8-16) [2022\)](#page-8-16), and Low-Rank Adaptation (LoRA)[\(Hu](#page-8-10) [et al.,](#page-8-10) [2022;](#page-8-10) [Valipour et al.,](#page-9-15) [2022\)](#page-9-15). Adapters are compact bottleneck modules located after transformer blocks; Prompts are learnable vectors prepended to the input sequence; and LoRA uses low-rank decomposition matrices to modify model weights during training. Recent progress explores applying PEFT to lifelong model editing, where the base model remains unchanged and new parameter modules are introduced to manage sequential edits over time.
+#### A.2. Routing Mechanisms in Modular Language Models
+When multiple modular components (e.g., LoRA modules) are present, a key challenge lies in routing input queries to the appropriate subset of modules during inference. Recent work explore various strategies to support dynamic module selection. MELO[\(Yu et al.,](#page-9-3) [2024\)](#page-9-3) introduces a clustering mechanism to assign edits to clusters based on input hidden states. During inference, the nearest cluster center is used to retrieve relevant LoRA modules. However, cluster centers are inevitably updated during the sequential editing, which may lead to semantic drift and incorrect matching. As shown in Fig[.2,](#page-1-0) the number of incorrect matches increases significantly as the number of cluster centre updating increases, which leads to a deterioration in model performance. EL-DER[\(Li et al.,](#page-9-6) [2025\)](#page-9-6) employs MoE approach in which a learnable neural network scores a set of LoRA modules, activating the top-k modules. This strategy offers parameter efficiency but introduces soft entanglement in routing, and due to shared module usage, new edits may interfere with or overwrite previously edits. Furthermore, both strategies rely on auxiliary routing networks, adding architectural complexity and computational overhead. They also offer limited control over edit selection, making rollback and deletion of edits challenging.
+## B. Additional Experiment Details
+Benchmarks We evaluate the performance of the model in three representative tasks: Document Classification, Question Answering(QA), and Hallucination Correction. Among them, the document classification task is performed based
+on the SCOTUS dataset[\(Chalkidis et al.,](#page-8-17) [2022\)](#page-8-17). SCOTUS is a subset of Fairlex, which collects documents from the U.S. Supreme Court and classifies them into 11 topics. In practice, the topics corresponding to the documents will change, so it is necessary to update the model knowledge to classify the documents into new topics. We divide SCO-TUS into an edit set and an upstream set, where models are edited on the edit set and the upstream set is not involved in editing, only testing. The question answering task is then conducted on the zsRE dataset[\(Levy et al.,](#page-9-16) [2017\)](#page-9-16), and NQ datasets[\(Kwiatkowski et al.,](#page-8-18) [2019\)](#page-8-18) serves as upstream dataset. Specifically, we perform editing in zsRE, and the NQ dataset is again not involved in editing, only testing. For hallucination correction, we adopted the framework introduced by[\(Manakul et al.,](#page-9-17) [2023\)](#page-9-17), aiming to address the tendency of GPT models to generate factually inconsistent outputs. Following GRACE[\(Hartvigsen et al.,](#page-8-9) [2023\)](#page-8-9), we employ SelfCheckGPT[\(Manakul et al.,](#page-9-17) [2023\)](#page-9-17) for editing. SelfCheckGPT is a Wikipedia-style biography that is first generated using GPT3 on 256 topics, e.g., "Bon Jovi", and then checked to see which of the generated biographies are hallucinations. SelfCheckGPT contains 1392 hallucinatory sentences to be edited, which serves as the editing set, and 516 already correct sentences as the retention set, which is used to measure model's post-edit perplexity on the already correct sentences. WebText[\(Nakano et al.,](#page-9-18) [2021\)](#page-9-18) is used as the upstream set, remaining unedited and reserved for testing. In addition to this, in order to test the performance of the model approach with different sized models as well as datasets, we conducted hallucination correction experiments on Uniedit[\(Chen et al.,](#page-8-13) [2025\)](#page-8-13) and WikiBigEdit[\(Thede](#page-9-11) [et al.,](#page-9-11) [2025\)](#page-9-11) to evaluate the generalization performance of the models. Uniedit is based on open domain knowledge covering information from 25 common domains in 5 major categories. The WikiBigEdit dataset is based on regular updates of the knowledge graph in Wikipedia and contains 5 months of extensive factual editing and improvement. The Uniedit and WikiBigEdit datasets already have edit set, retention set and upstream set that can be used directly for hallucination correction tasks.
+Training details Following[\(Hartvigsen et al.,](#page-8-9) [2023\)](#page-8-9), we employ BERT, T5-small, and GPT2-XL as the base models for the three tasks—document classification, question answering, and hallucination correction, respectively. Among these, BERT and T5-small are the official pre-trained models, while GPT2-XL is obtained from [\(Hartvigsen et al.,](#page-8-9) [2023\)](#page-8-9) which finetuned from the official release[\(Hartvigsen](#page-8-9) [et al.,](#page-8-9) [2023\)](#page-8-9). For training, we adopt Stochastic Gradient Descent (SGD) as the optimizer with a cosine decay learning rate schedule. The learning rates is 0.05, training epochs
+{11}------------------------------------------------
+is 40, and LoRA rank is 4. In the three tasks, we edit the models and the edited layers of the individual model edits are shown in the Tab[.6.](#page-11-0)
+*Table 6.* Edited layers of different model
+<span id="page-11-0"></span>
+| Model | Edit layer |
+|----------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+| BERT | bert.encoder.layer.9.output.dense<br>bert.encoder.layer.10.output.dense<br>bert.encoder.layer.11.output.dense |
+| T5-Small | encoder.block.5.layer.1.DenseReluDense.wi 0<br>encoder.block.5.layer.1.DenseReluDense.wi 1<br>encoder.block.5.layer.1.DenseReluDense.wo<br>encoder.block.6.layer.1.DenseReluDense.wi 0<br>encoder.block.6.layer.1.DenseReluDense.wi 1<br>encoder.block.6.layer.1.DenseReluDense.wo |
+| GPT2-XL | transformer.h.36.mlp.c fc<br>transformer.h.37.mlp.c fc |