Puns pose a challenging test case for large language models (LLMs), as their interpretation requires mapping a surface expression to an implicit meaning through a specific pun-forming mechanism. Existing benchmarks typically evaluate isolated abilities such as detection, localization, explanation, or generation quality, leaving unclear whether LLMs can consistently identify, explain, and apply the mechanisms that make puns work. To address this gap, we introduce MECHPUNBENCH, a bilingual, mechanism-oriented benchmark for Chinese and English pun understanding and generation. Each positive instance is annotated with trigger spans, surface meanings, implicit meanings, rhetorical effects, and mechanism labels. Based on a taxonomy of Lexical-Semantic, Structural-Syntactic, and Structural-Semantic puns, we formulate six diagnostic tasks spanning surface recognition to mechanism-guided generation. Experiments on representative closed-source and open-source LLMs reveal a substantial surface–mechanism gap: although the best model achieves 82.8% detection accuracy and 88.2% containment localization accuracy, it remains below the 4/5 threshold used in our evaluation on implicit meaning explanation, mechanism explanation, rhetorical effect explanation, and mechanism-consistent generation. These findings suggest that recognizing puns or generating fluent pun-like text does not necessarily indicate robust mechanism-level pun understanding.

Model editing provides a promising mechanism for updating large language models (LLMs) without expensive retraining. Existing approaches, particularly locate-and-edit methods based on least-squares optimization, aim to introduce targeted knowledge changes while preserving pre-trained behavior. In this work, we show that this objective is fundamentally fragile under standard single-edit evaluation protocols. We first develop a unified theoretical framework that characterizes activation-based editing as a constrained intervention on intermediate representations. Within this framework, we demonstrate that least-squares edits cannot, in general, isolate target updates from unrelated activations, giving rise to unavoidable interference that accumulates with successive edits. Crucially, this degradation can remain undetected in single-edit settings when assessed using conventional success and locality metrics. To expose such hidden instabilities, we introduce an uncertainty-based evaluation protocol that combines structured semantic perturbations with uncertainty quantification based on Sampling with Perturbation for UQ. By measuring edit-induced growth in aleatoric and epistemic uncertainty, our method reveals local knowledge conflicts that are invisible to existing benchmarks. Extensive experiments across multiple models, datasets, and editing algorithms show that both least-squares and other parameter-update-based methods consistently increase post-edit uncertainty. Together, our results suggest that current evaluation practices substantially overestimate the reliability of single-edit model editing, and that uncertainty-based diagnostics are necessary for assessing edit stability.

While Large Language Models (LLMs) achieve high accuracy on established Classical Chinese Poetry benchmarks, it remains challenging to distinguish transferable Linguistic-Aesthetic Reasoning from reliance on familiar pre-training patterns. To address this issue, we introduce Neo-Classic, an evaluation benchmark that combines a constructionist Out-of-Sample (OOS) dataset with a suite of reverse understanding probes. Unlike traditional benchmarks that rely on verification or generation over historical corpora, Neo-Classic comprises strictly metrical poetry authored by contemporary experts, reducing the possibility of direct retrieval. We evaluate state-of-the-art models, including Qwen3-Max, Gemini-3-Pro, and DeepSeek-V3.2, across five behavioral probes designed to test hierarchical constraint satisfaction. Our results reveal two primary limitations. First, a performance gap of 20%–50% emerges when models transition from historical to contemporary texts. Second, models exhibit substantial difficulties in discourse-level ordering tasks, with standard accuracy remaining low (0–13%). Although expert-level guidance improves the performance of reasoning-enhanced models to 36%, a notable gap with human experts persists. These findings suggest that while current LLMs capture local formal patterns, they struggle with global hierarchical planning required for robust Linguistic-Aesthetic Reasoning.