While Large Language Models (LLMs) can now automate many parts of scientific writing, architectural figure generation faces a critical modality bottleneck because image outputs are inherently difficult to edit, verify, and integrate into text-native LLM workflows. To enable fully text-driven synthesis, we propose a new task: generating architectural figures as pure HTML/SVG source code. Furthermore, we introduce FigureCode, a novel benchmark that evaluates structural fidelity and visual properties through code-based evaluation. Our benchmark improves agreement with human judgments by 0.35 Kendall’s τ over prior VLM-based methods. Moreover, the figure-as-code paradigm improves edit-instruction success rate by 30% and retrieval F1 score by 25.3% over baselines, while achieving 66.1% WCAG/ARIA accessibility coverage.

As large language models (LLMs) increasingly process instructions from multiple sources with different privilege ranks, preserving the priority of higher-rank directives over conflicting lower-rank ones is critical for safe and predictable deployment. However, existing evaluations mostly focus on limited conflict settings, leaving it unclear how models handle instruction hierarchy (IH) when multiple conflicts arise. To address this gap, we present the first systematic study of IH under multi-conflict settings and introduce MIH-Bench, a benchmark that scales the number of simultaneous conflicts across six pairwise hierarchy scenarios. Experiments on 21 LLMs show that IH compliance deteriorates sharply as conflicts accumulate, while neither model scale nor generation reliably predicts robustness. We further find that system-involving conflicts are the most vulnerable, a particularly concerning result given the safety-critical role of system prompts. Moreover, reasoning can improve conflict resolution, although the gains vary across models and are bounded by underlying model capability. These results show that current LLMs remain fragile in pairwise multi-conflict settings, highlighting the need for stronger hierarchy-aware alignment. Code and data are available at https://anonymous.4open.science/r/MIH-release-A5B8/

Instruction tuning has emerged as a powerful technique, significantly boosting zero-shot performance on unseen tasks. While recent work has explored cross-lingual generalization by applying instruction tuning to multilingual models, previous studies have primarily focused on English, with a limited exploration of non-English tasks. For in-depth exploration of cross-lingual generalization in instruction tuning, we perform instruction tuning individually for two distinct language meta-datasets. Subsequently, we assess the performance on unseen tasks in the language different from the one used for training. To facilitate this investigation, we introduce a novel non-English meta-dataset named “KORANI” (Korean Natural Instruction), comprising 51 Korean benchmarks. Moreover, we design cross-lingual templates to mitigate discrepancies in language and instruction-format of the template between training and inference within the cross-lingual setting. Our experiments reveal consistent improvements through cross-lingual generalization in both English and Korean, outperforming baseline by average scores of 20.7% and 13.6%, respectively. Remarkably, these enhancements are comparable to those achieved by mono-lingual instruction tuning and even surpass them in some tasks. The result underscores the significance of relevant data acquisition across languages over linguistic congruence with unseen tasks during instruction tuning.

Instruction tuning has been proven effective in enhancing zero-shot generalization across various tasks and in improving the performance of specific tasks. For task-specific improvements, strategically selecting and training on related tasks that provide meaningful supervision is crucial, as this approach enhances efficiency and prevents performance degradation from learning irrelevant tasks. In this light, we introduce a simple yet effective task selection method that leverages instruction information alone to identify relevant tasks, optimizing instruction tuning for specific tasks. Our method is significantly more efficient than traditional approaches, which require complex measurements of pairwise transferability between tasks or the creation of data samples for the target task. Additionally, by aligning the model with the unique instructional template style of the meta-dataset, we enhance its ability to granularly discern relevant tasks, leading to improved overall performance. Experimental results demonstrate that training on a small set of tasks, chosen solely based on the instructions, results in substantial improvements in performance on benchmarks such as P3, Big-Bench, NIV2, and Big-Bench Hard. Significantly, these improvements surpass those achieved by prior task selection methods, highlighting the superiority of our approach.

Language Models (LMs) become outdated as the world changes; they often fail to perform tasks requiring recent factual information which was absent or different during training, a phenomenon called temporal misalignment. This is especially a challenging problem because the research community still lacks a coherent dataset for assessing the adaptability of LMs to frequently-updated knowledge corpus such as Wikipedia. To this end, we introduce TemporalWiki, a lifelong benchmark for ever-evolving LMs that utilizes the difference between consecutive snapshots of English Wikipedia and English Wikidata for training and evaluation, respectively. The benchmark hence allows researchers to periodically track an LM’s ability to retain previous knowledge and acquire updated/new knowledge at each point in time. We also find that training an LM on the diff data through continual learning methods achieves similar or better perplexity than on the entire snapshot in our benchmark with 12 times less computational cost, which verifies that factual knowledge in LMs can be safely updated with minimal training data via continual learning.