Automatic Speech Recognition (ASR) transcription accuracy remains highly sensitive to audio segmentation strategies, yet most benchmarks assume oracle timestamps unavailable in deployment. We systematically evaluate how audio segmentation affects Whisper’s performance on 10 hours of Persian YouTube content, comparing transcript-aligned (oracle) versus silence-based (realistic) approaches across contrasting acoustic conditions. Results reveal striking content-type dependency: podcast content benefits from timestamp segmentation (33% lower mean WER), while entertainment content favors silence-based segmentation (8% lower mean WER). This finding demonstrates that optimal segmentation must be content-aware, with silence detection better capturing natural boundaries in acoustically heterogeneous media while avoiding mid-utterance splits. We publicly release our evaluation framework, 10 hours of audio with gold transcripts, and segmentation results here: https://github.com/ri164-bolleit/persian-youtube-whisper-benchmark

We present the GUIR system for SemEval-2026 Task 7, Everyday Knowledge Across Diverse Languages and Cultures, which probes the extent to which general-purpose LLMs encode cultural knowledge without any culture-specific supervision or fine-tuning. Our system addresses two tracks built on the BLEnD benchmark. For the short-answer question (SAQ) track, we employ zero-shot prompting with gpt-4.1, achieving 55.5% accuracy across 61 language locales. For the multiple-choice question (MCQ) track, we propose a three-stage pipeline: (1) zero-shot chain-of-thought inference with gpt-5-mini, (2) cross-locale majority voting to correct inconsistent predictions, and (3) a multi-agent debate protocol in which three LLM instances argue and adjudicate over residual errors. This pipeline achieves 97.47% overall accuracy across 30 locales, ranking first among all submitted systems on the MCQ track. We further conduct a targeted human evaluation on the Persian locale, revealing that BLEnD’s lemma-matching scorer systematically underestimates model performance, with human annotators scoring the system 18 percentage points higher than the lemma-matching evaluation. This reveals the need for better evaluation of morphologically rich languages like Persian.

Uniform Meaning Representation (UMR) is a recently developed graph-based semantic representation, which expands on Abstract Meaning Representation (AMR) in a number of ways, in particular through the inclusion of document-level information and multilingual flexibility. In order to effectively adopt and leverage UMR for downstream tasks, efforts must be placed toward developing a UMR technological ecosystem. Though only a small amount of UMR annotations have been produced to date, in this work, we investigate the first approaches to producing text from multilingual UMR graphs. Exploiting the structural similarity between UMR and AMR graphs and the wide availability of AMR technologies, we introduce (1) a baseline approach which passes UMR graphs to AMR-to-text generation models, (2) a pipeline conversion of UMR to AMR, then using AMR-to-text generation models, and (3) a fine-tuning approach for both foundation models and AMR-to-text generation models with UMR data. Our best performing models achieve multilingual BERTscores of 0.825 for English and 0.882 for Chinese, a promising indication of the effectiveness of fine-tuning approaches for UMR-to-text generation even with limited UMR data.

LLM jailbreaks are a widespread safety challenge. Given this problem has not yet been tractable, we suggest targeting a key failure mechanism: the failure of safety to generalize across semantically equivalent inputs. We further focus the target by requiring desirable tractability properties of attacks to study: explainability, transferability between models, and transferability between goals. We perform red-teaming within this framework by uncovering new vulnerabilities to multi-turn, multi-image, and translation-based attacks. These attacks are semantically equivalent by our design to their single-turn, single-image, or untranslated counterparts, enabling systematic comparisons; we show that the different structures yield different safety outcomes. We then demonstrate the potential for this framework to enable new defenses by proposing a Structure Rewriting Guardrail, which converts an input to a structure more conducive to safety assessment. This guardrail significantly improves refusal of harmful inputs, without over-refusing benign ones. Thus, by framing this intermediate challenge—more tractable than universal defenses but essential for long-term safety—we highlight a critical milestone for AI safety research.