Large Language Models (LLMs) acquire extensive factual knowledge during computationally expensive pretraining. Knowledge editing aims to efficiently update specific facts in a trained model without retraining the entire network. While many editing methods have recently been proposed and show strong performance on existing benchmarks, these benchmarks exhibit several limitations. In particular, they allow only a single edit per entity, rely on highly templated prompts, and evaluate edits using relative probability changes rather than ground-truth correctness. As a result, the capabilities and limitations of current knowledge editing methods remain insufficiently understood. In this work, we introduce the Subject-centric Entity Editing Dataset (SEED), a dataset and evaluation framework designed to address these limitations. SEED organizes edits into sets of related and unrelated facts rather than isolated instances, enabling evaluation when multiple facts about the same subject entity are updated. It also introduces undirected knowledge probes that test whether edited facts can be recalled in general contexts, and ground-truth evaluation metrics that directly measure whether the edited fact appears in the model’s output. Using SEED, we test 9 state-of-the-art knowledge editing methods over 10 metrics. These studies find that editing performance decreases by up to 36.9% as the number of edits per entity increases and edit recall decreases by 36.6% with prompts that increase token-wise distance between the edit subject and relation.

Natural language (NL) to temporal logic (TL) translation lets engineers specify and verify system behaviors without hand-crafting formal specifications. Existing frameworks lift NL into syntactically valid TL but leave atomic propositions abstract—not grounded to the concrete predicates and typed objects a system exposes (e.g., binding “find the bookbag” to search(backpack)). Ungrounded formulas cannot be checked over traces, yet current systems either omit grounding or ground inaccurately. We propose , a modular stage for Grounding Natural Language Into System Signatures that attaches to any NL-to-TL lifter. We cast grounding as signature-conditioned classification: enumerating the signature as an input prefix lets one encoder-only model assign each abstract AP a predicate and its typed arguments, generalize to unseen signature elements, and transfer across domains without retraining, while never hallucinating out-of-signature atoms. Because predictions are selections from the enumerated signature, the out-of-signature rate is 0% by construction, against up to 95% for prompted LLMs given the same signature in-context. attains at least 99% joint grounding accuracy across seven domains (worst case 99.39 ± 0.04 over five seeds); paired with a lifter, it produces executable specifications that track the oracle-grounding upper bound, improving grounded logical equivalence by 75% over the strongest prompted-LLM and embedding-similarity grounder.