KEDAS: Factual Knowledge Editing in LLMs

• KEDAS is a framework for factual knowledge editing in LLMs that combines LoRA-based alignment, diverse textual augmentation, and self-adaptive inference to update model responses.
• It ensures high edit success, locality, and portability by dynamically routing queries through either the base or the edit-aware model using a BERT-based filter.
• Empirical results across datasets like ZsRE and WikiBio show KEDAS significantly outperforms LTE, achieving up to a 19.8 point improvement in the harmonic mean of key metrics.

KEDAS denotes Knowledge Editing Alignment with Diverse Augmentation and Self-adaptive Inference, a 2025 framework for factual knowledge editing in LLMs. It treats knowledge editing as a combination of one-time alignment, external edit memory, and inference-time routing: the base LLM is aligned with LoRA adapters to apply edited knowledge presented in context, edits are stored in multiple augmented textual forms, and a smart retriever plus filter decides whether a query should bypass editing and use the original model or invoke the aligned edit-aware path (Tang et al., 2 Aug 2025).

1. Terminology and disambiguation

The label KEDAS is not unique across arXiv usage. In current LLM knowledge-editing literature, it names the framework introduced as “Knowledge Editing Alignment with Diverse Augmentation and Self-adaptive Inference” (Tang et al., 2 Aug 2025). In a distinct line of work on evaluation for knowledge graph completion, KG-EDAS abbreviates “Knowledge Graph Evaluation based on Distance from Average Solution” and is described as being “often informally written ‘KEDAS’” (Gul et al., 21 Aug 2025). By contrast, a 2013 detector-software report on JUDIDT-based data acquisition and monitoring does not use the term “KEDAS” anywhere; the document instead refers to a DAQ program suite around JUDIDT and related tools (Fabbri et al., 2013).

Usage	Expansion or status	Domain
KEDAS	Knowledge Editing Alignment with Diverse Augmentation and Self-adaptive Inference	LLM knowledge editing
KG-EDAS	Knowledge Graph Evaluation based on Distance from Average Solution	Knowledge graph completion evaluation
“KEDAS” in the 2013 DAQ report	Not mentioned anywhere in the text	Detector DAQ and monitoring

Within the literature on LLM editing, KEDAS is therefore a specific architecture for factual updating rather than a generic label for editing systems. The naming overlap matters because the three usages address unrelated technical problems: knowledge editing in LLMs, meta-metric aggregation for KGC benchmarking, and detector DAQ software.

2. Formal task and motivation

KEDAS is formulated for knowledge editing, where an LLM is represented as a function $f:\mathcal{Q}\rightarrow\mathcal{A}$ from query space to answer space, and the edit set is written as

$$\mathcal{E} = \{e^t\}_{t=1}^N = \{(q_e^t,a_e^t)\}_{t=1}^N.$$

Each edit provides a triggering query and a desired updated answer. The post-edit system must satisfy three criteria: edit success, meaning correct responses on edited queries; locality, meaning preservation of answers on unrelated queries; and portability, meaning generalization of the edit to related queries (Tang et al., 2 Aug 2025).

The framework is motivated by deficiencies in two established families of methods. Parameter-based editing methods, including ROME, MEMIT, MEND, WISE, and AlphaEdit, directly change or add parameters; the paper states that such methods can be powerful but often require online gradient steps per edit, scale poorly to many edits, and risk accumulation of interference and degradation. Retrieval-based editing methods, including IKE, EREN, RECIPE, and LTE, externalize edits and inject them through prompts or soft prompts; these methods are often gradient-free and efficient, but the base model is not necessarily trained to interpret edits robustly, and naive retrieval can harm locality (Tang et al., 2 Aug 2025).

KEDAS is positioned as an alignment-based alternative that retains external memory while explicitly teaching the model how to use updated information in context. The immediate antecedent is LTE, which aligns an LLM for editing with an “editing prompt” but always uses the post-alignment model, including for irrelevant queries. The KEDAS paper identifies that design choice as a source of locality loss and degradation on general abilities (Tang et al., 2 Aug 2025).

3. Core methodology

KEDAS has three coupled components: LoRA-based alignment, diverse edit augmentation, and self-adaptive post-alignment inference (Tang et al., 2 Aug 2025).

During the alignment phase, the model is trained on an editing prompt of the form

$\Phi$5

using low-rank adaptation rather than full fine-tuning. For a linear layer with weight matrix $$W\in\mathbb{R}^{d_{\text{out}}\times d_{\text{in}}}$$, LoRA parameterizes a rank-$r$ update as

$\Delta W = A B^\top,$

with $A\in\mathbb{R}^{d_{\text{out}}\times r}$ and $B\in\mathbb{R}^{d_{\text{in}}\times r}$. The effective model is denoted $f_{\Phi+\Delta\Phi(\Theta)}$, with frozen base parameters $\Phi$ and trainable adapter parameters $\Theta$. The aligned model is trained to use in-scope edits, ignore irrelevant edits for out-of-scope queries, and preserve general capabilities through additional general-query data (Tang et al., 2 Aug 2025).

The diverse edit augmentation stage expands each edit into multiple semantically equivalent textual realizations. For an edit $$\mathcal{E} = \{e^t\}_{t=1}^N = \{(q_e^t,a_e^t)\}_{t=1}^N.$$0, the memory includes the original QA form, a declarative form, three paraphrases generated with gpt-4o-mini, and a reversed form. The memory update is written as

$$\mathcal{E} = \{e^t\}_{t=1}^N = \{(q_e^t,a_e^t)\}_{t=1}^N.$$1

This is intended to improve retrieval recall and portability by reducing sensitivity to surface form (Tang et al., 2 Aug 2025).

The third component is self-adaptive inference. Given a query $$\mathcal{E} = \{e^t\}_{t=1}^N = \{(q_e^t,a_e^t)\}_{t=1}^N.$$2, the system first uses all-MiniLM-L6-v2 for high-recall retrieval with top-$$\mathcal{E} = \{e^t\}_{t=1}^N = \{(q_e^t,a_e^t)\}_{t=1}^N.$$3 selection, where $$\mathcal{E} = \{e^t\}_{t=1}^N = \{(q_e^t,a_e^t)\}_{t=1}^N.$$4:

$$\mathcal{E} = \{e^t\}_{t=1}^N = \{(q_e^t,a_e^t)\}_{t=1}^N.$$5

A trained binary classifier based on bert-base-uncased then filters candidate edits for relevance. If multiple edits survive, the system selects

$$\mathcal{E} = \{e^t\}_{t=1}^N = \{(q_e^t,a_e^t)\}_{t=1}^N.$$6

The final routing rule is

$$\mathcal{E} = \{e^t\}_{t=1}^N = \{(q_e^t,a_e^t)\}_{t=1}^N.$$7

Irrelevant queries therefore traverse the original pre-alignment model, while relevant queries are answered with the aligned adapters activated and the retrieved edit placed in context (Tang et al., 2 Aug 2025).

4. Training configuration and evaluation protocol

KEDAS is evaluated with three open-source LLMs: Llama-2-7B-Chat, Qwen2.5-7B-Instruct, and Llama-3.1-8B-Instruct (Tang et al., 2 Aug 2025). The LoRA configurations are reported explicitly: for Llama-2-7B-Chat, rank $$\mathcal{E} = \{e^t\}_{t=1}^N = \{(q_e^t,a_e^t)\}_{t=1}^N.$$8, $$\mathcal{E} = \{e^t\}_{t=1}^N = \{(q_e^t,a_e^t)\}_{t=1}^N.$$9, dropout $$W\in\mathbb{R}^{d_{\text{out}}\times d_{\text{in}}}$$0; for Qwen2.5-7B-Instruct, rank $$W\in\mathbb{R}^{d_{\text{out}}\times d_{\text{in}}}$$1, $$W\in\mathbb{R}^{d_{\text{out}}\times d_{\text{in}}}$$2, dropout $$W\in\mathbb{R}^{d_{\text{out}}\times d_{\text{in}}}$$3; for Llama-3.1-8B-Instruct, rank $$W\in\mathbb{R}^{d_{\text{out}}\times d_{\text{in}}}$$4, $$W\in\mathbb{R}^{d_{\text{out}}\times d_{\text{in}}}$$5, dropout $$W\in\mathbb{R}^{d_{\text{out}}\times d_{\text{in}}}$$6.

The alignment stage reuses LTE’s training data, including in-scope and out-of-scope editing prompts and general queries with GPT-4 completions. The reported training frameworks are LTE official scripts with DeepSpeed plus FastChat for Llama-2, and LLaMA-Factory for Qwen2.5 and Llama-3.1. Reported hyperparameters include batch size $$W\in\mathbb{R}^{d_{\text{out}}\times d_{\text{in}}}$$7, gradient accumulation of $$W\in\mathbb{R}^{d_{\text{out}}\times d_{\text{in}}}$$8 for Llama-2, $$W\in\mathbb{R}^{d_{\text{out}}\times d_{\text{in}}}$$9 for Qwen2.5, and $r$0 for Llama-3.1; learning rates of $r$1 for Llama-2 and $r$2 for Qwen2.5 and Llama-3.1; $r$3 epochs for Llama-2 and Qwen2.5, and $r$4 epoch for Llama-3.1 to avoid overfitting; maximum sequence lengths of $r$5 for Llama-2 and Qwen2.5 and $r$6 for Llama-3.1; and AdamW with cosine schedule and warmup ratio $r$7 or $r$8 (Tang et al., 2 Aug 2025).

The paper reports alignment training times of approximately 12 hours on 2×A40 48GB for Llama-2, 19 hours on 1×A40 for Qwen2.5, and 6 hours on 1×A40 for Llama-3.1. After this offline phase, no further parameter updates are needed for new edits (Tang et al., 2 Aug 2025).

Evaluation uses four datasets from the KnowEdit suite: ZsRE, WikiBio, WikiData$r$9, and WikiData$\Delta W = A B^\top,$0. The reported train and test edit counts are 10,000 train and 1304 test edits for ZsRE, 464 train and 306 test for WikiBio, 570 train and 1266 test for WikiData$\Delta W = A B^\top,$1, and 1428 train and 885 test for WikiData$\Delta W = A B^\top,$2. Metrics include edit success (ES), locality (L), portability (P), fluency, and the harmonic mean (HM) of ES, L, and P; for WikiBio, which has no portability queries, HM is the harmonic mean of ES and L (Tang et al., 2 Aug 2025).

Three editing regimes are considered. In single editing, each edit is applied independently. In sequential editing, the update process is written as

$\Delta W = A B^\top,$3

with only the final model evaluated. In incremental editing, the same update process is used, but evaluation is performed after each step on the corresponding evaluation sample. The framework is reported to support all three settings without extra training (Tang et al., 2 Aug 2025).

5. Empirical findings

The principal empirical claim is that KEDAS “secures the highest overall performance scores in 35 out of 36 cases across four datasets with three LLMs on three settings” and surpasses its “strong knowledge editing alignment counterpart by about 19.8 harmonic mean scores of edit success, locality and portability” (Tang et al., 2 Aug 2025). The same paper reports that, across all settings and models, KEDAS wins in 35 of 36 cases, and that in the sequential setting it gives the best HM in all 12 model–dataset cases.

For Llama-2-7B-Chat in the sequential setting, the paper reports the following comparisons against LTE: on ZsRE, LTE has ES $\Delta W = A B^\top,$4, L $\Delta W = A B^\top,$5, P $\Delta W = A B^\top,$6, HM $\Delta W = A B^\top,$7, whereas KEDAS has ES $\Delta W = A B^\top,$8, L $\Delta W = A B^\top,$9, P $A\in\mathbb{R}^{d_{\text{out}}\times r}$0, HM $A\in\mathbb{R}^{d_{\text{out}}\times r}$1; on WikiBio, LTE has HM $A\in\mathbb{R}^{d_{\text{out}}\times r}$2 and KEDAS HM $A\in\mathbb{R}^{d_{\text{out}}\times r}$3; on WikiData$A\in\mathbb{R}^{d_{\text{out}}\times r}$4, LTE has HM $A\in\mathbb{R}^{d_{\text{out}}\times r}$5 and KEDAS HM $A\in\mathbb{R}^{d_{\text{out}}\times r}$6; on WikiData$A\in\mathbb{R}^{d_{\text{out}}\times r}$7, LTE has HM $A\in\mathbb{R}^{d_{\text{out}}\times r}$8 and KEDAS HM $A\in\mathbb{R}^{d_{\text{out}}\times r}$9 (Tang et al., 2 Aug 2025). Comparable gains are reported for Qwen2.5-7B-Instruct and Llama-3.1-8B-Instruct, including ZsRE HM improvements from $B\in\mathbb{R}^{d_{\text{in}}\times r}$0 to $B\in\mathbb{R}^{d_{\text{in}}\times r}$1 and from $B\in\mathbb{R}^{d_{\text{in}}\times r}$2 to $B\in\mathbb{R}^{d_{\text{in}}\times r}$3, respectively.

The paper also reports results on general tasks—CSQA, MNLI, MMLU, and SST—to assess interference outside the editing benchmark. The reported averages are 54.67 for the pre-alignment model, 53.33 for the post-alignment model with no edits, 48.75 for LTE after edits, and 54.67 for KEDAS after 1304 sequential edits. The routing mechanism is therefore presented as the component that preserves general capabilities by sending irrelevant queries back to the base model (Tang et al., 2 Aug 2025).

The ablation study isolates the contributions of DEA, FLT, SPI, and ALGN. Averaged sequential results with Llama-2 are reported as follows: full KEDAS gives ES $B\in\mathbb{R}^{d_{\text{in}}\times r}$4, L $B\in\mathbb{R}^{d_{\text{in}}\times r}$5, P $B\in\mathbb{R}^{d_{\text{in}}\times r}$6, HM $B\in\mathbb{R}^{d_{\text{in}}\times r}$7; without DEA, ES $B\in\mathbb{R}^{d_{\text{in}}\times r}$8, L $B\in\mathbb{R}^{d_{\text{in}}\times r}$9, P $f_{\Phi+\Delta\Phi(\Theta)}$0, HM $f_{\Phi+\Delta\Phi(\Theta)}$1; without FLT, ES $f_{\Phi+\Delta\Phi(\Theta)}$2, L $f_{\Phi+\Delta\Phi(\Theta)}$3, P $f_{\Phi+\Delta\Phi(\Theta)}$4, HM $f_{\Phi+\Delta\Phi(\Theta)}$5; without SPI, ES $f_{\Phi+\Delta\Phi(\Theta)}$6, L $f_{\Phi+\Delta\Phi(\Theta)}$7, P $f_{\Phi+\Delta\Phi(\Theta)}$8, HM $f_{\Phi+\Delta\Phi(\Theta)}$9; without ALGN, ES $\Phi$0, L $\Phi$1, P $\Phi$2, HM $\Phi$3 (Tang et al., 2 Aug 2025). These numbers identify the filter as the main determinant of locality, the alignment stage as necessary for edit success and portability, and diverse augmentation as a contributor to retrieval recall and generalization.

The efficiency analysis for 1304 ZsRE edits with Llama-2-7B-Chat reports editing time 0.020 h, inference time 1.662 h, and GPU memory 25.9 GB for KEDAS, compared with 0.003 h, 1.490 h, and 25.5 GB for LTE, and substantially larger time or memory figures for SERAC, IKE, WISE, and RECIPE (Tang et al., 2 Aug 2025). The reported interpretation is that KEDAS incurs a small retrieval-and-filter overhead relative to LTE while delivering a large improvement in HM.

6. Position in the literature, limitations, and relation to other “KEDAS” terms

KEDAS is positioned as a hybrid between alignment and retrieval rather than as a weight-surgery method. Relative to ROME, MEMIT, MEND, WISE, and AlphaEdit, it keeps base weights frozen and stores edits externally; relative to IKE, EREN, and RECIPE, it adds explicit alignment so that the model learns to treat updated information as authoritative; and relative to LTE, it adds diverse edit augmentation, a smart retriever with a filter, and self-adaptive routing between the base and aligned models (Tang et al., 2 Aug 2025). A plausible implication is that the distinctive contribution is not merely external memory, but the coupling of external memory with a learned routing policy and a togglable aligned submodel.

The paper identifies four limitations. First, overall performance depends heavily on the BERT-based filter: false negatives reduce edit success and portability, while false positives can harm locality. Second, diverse edit augmentation depends on an external LLM, specifically gpt-4o-mini, which introduces API dependence and cost. Third, the current design is limited to one-hop portability rather than multi-hop reasoning over edited knowledge. Fourth, while the reported experiments handle more than 1000 edits, memory growth and retrieval/filter cost scale with the number of stored edits, so more scalable indexing or pruning may be required at larger scale (Tang et al., 2 Aug 2025).

The term’s overlap with KG-EDAS is conceptually accidental rather than substantive. KG-EDAS is a meta-metric for Knowledge Graph Completion that aggregates multiple metrics and datasets into a single normalized score $\Phi$4 using distance from the average solution (Gul et al., 21 Aug 2025). It addresses cross-metric and cross-dataset ranking conflicts in KGC, not knowledge editing in LLMs. The 2013 DAQ report is more distant still: it describes a client–server DAQ program suite around JUDIDT and ACQIRIS, but explicitly does not define or mention KEDAS at all (Fabbri et al., 2013).

In the arXiv record, then, KEDAS primarily names a 2025 knowledge-editing framework for LLMs, while adjacent uses either denote a different acronymic expansion—KG-EDAS—or reflect an external naming convention not present in the underlying paper.