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

This paper presents the Georgetown University NLP (GUNLP) system developed for SemEval 2026 Task 10: Psycholinguistic Conspiracy Marker Extraction and Detection, addressing the classification of conspiratorial beliefs in Reddit posts (Subtask 2). Our approach leverages COVID-Twitter-BERT v2 (CT-BERT-v2) within a multi-task learning framework that jointly optimizes conspiracy classification and emotion label prediction through a dual-head architecture. To address data scarcity, we enrich the training set using paraphrasing-based data augmentation and GPT-5-generated chain-of-thought emotion annotations, effectively doubling the training corpus to approximately 8,600 examples. We evaluate two input configurations: text only and text concatenated with emotion labels. The emotion-aware configuration achieves the strongest performance with an F1 score of 0.87 on the official development set, outperforming the text-only baseline by five F1 points and demonstrating the value of paraphrased samples and affective auxiliary supervision for conspiracy detection in social media text.

Using large language models (LLMs) to simulate diverse human populations promises to transform computational social science. However, the accuracy of these simulations remains unclear because current evaluations prioritize predicting an average response over capturing the rich diversity of opinions within real human groups. Here, we develop a diagnostic framework measuring both accuracy and viewpoint retention across 10,000 respondent-question pairs from the World Values Survey, spanning twelve countries and six continents. By evaluating nine LLMs across four alignment paradigms (zero-shot, SFT, DPO, GRPO) against ground-truth human responses, we identify a failure mode we term consensus collapse. While training alignment improves point accuracy, we find that it compresses the range of outputs toward a single demographic stereotype. The highest-performing model (Tulu 3 70B DPO, 57.9% accuracy) retains only 50% of real human variance (σr = 0.50). We demonstrate this collapse is within the models’ underlying parameters rather than decoding procedures; temperature scaling up to T = 1.0 alters fewer than 3.7% of predictions. Furthermore, this diversity loss is inequitable, erasing significantly more variation for Global South populations like Nigeria (σr = 0.11) and Ethiopia (σr = 0.16) compared to Western democracies (σr = 0.70–0.87), with accuracy-rewarded GRPO unable to reverse this effect. While mixing an aligned model with an unaligned prior partially mitigates the collapse (reducing W1 from 0.55 to 0.52), substantial gaps remain (Nigeria recovering from σr = 0.14 to 0.47). Ultimately, we find that modern AI alignment methods trade distributional diversity for point accuracy, challenging existing assumptions for using LLMs as substitutes for human populations. Yet, by exposing the mechanics of this collapse, our work provides insight for future methods designed to value and preserve diverse human viewpoints.