Detecting hallucinations in Large Language Model (LLM) outputs is critical for safe deployment, yet existing methods require white-box access to internals, depend on multiple generations or auxiliary judge models, or collapse token-level uncertainty into scalar summaries that discard temporal structure. We introduce HALT, a lightweight grey-box detector that treats the top-20 log-probabilities of an LLM generation as a multivariate time series and processes them with a small bidirectional GRU (5M parameters), in a single forward pass and using no model internals. We evaluate on HUB, a unified protocol reorganizing existing hallucination datasets into ten capability clusters spanning factual and logical failure modes. HALT matches or surpasses Lettuce, a fine-tuned ModernBERT-base detector, on macro-F₁ while using ∼30× fewer parameters; a controlled comparison against a text classifier trained on HALT’s own data confirms this advantage comes from the log-probability signal itself rather than from differences in training data, and the same recipe applies across eight LLM families from 360M to 70B parameters.

This paper presents our submissions to WMT 22 shared task in the Unsupervised and Very Low Resource Supervised Machine Translation tasks. The task revolves around translating between German ↔ Upper Sorbian (de ↔ hsb), German ↔ Lower Sorbian (de ↔ dsb) and Upper Sorbian ↔ Lower Sorbian (hsb ↔ dsb) in both unsupervised and supervised manner. For the unsupervised system, we trained an unsupervised phrase-based statistical machine translation (UPBSMT) system on each pair independently. We pretrained a De-Salvic mBART model on the following languages Polish (pl), Czech (cs), German (de), Upper Sorbian (hsb), Lower Sorbian (dsb). We then fine-tuned our mBART on the synthetic parallel data generated by the (UPBSMT) model along with authentic parallel data (de ↔ pl, de ↔ cs). We further fine-tuned our unsupervised system on authentic parallel data (hsb ↔ dsb, de ↔ dsb, de ↔ hsb) to submit our supervised low-resource system.

Online presence on social media platforms such as Facebook and Twitter has become a daily habit for internet users. Despite the vast amount of services the platforms offer for their users, users suffer from cyber-bullying, which further leads to mental abuse and may escalate to cause physical harm to individuals or targeted groups. In this paper, we present our submission to the Arabic Hate Speech 2022 Shared Task Workshop (OSACT5 2022) using the associated Arabic Twitter dataset. The Shared Task consists of 3 Sub-tasks, Sub-task A focuses on detecting whether the tweet is Offensive or not. Then, For offensive Tweets, Sub-task B focuses on detecting whether the tweet is Hate Speech or not. Finally, For Hate Speech Tweets, Sub-task C focuses on detecting the fine-grained type of hate speech among six different classes. Transformer models proved their efficiency in classification tasks, but with the problem of over-fitting when fine-tuned on a small or an imbalanced dataset. We overcome this limitation by investigating multiple training paradigms such as Contrastive learning and Multi-task learning along with classification fine-tuning and an ensemble of our top 5 performers. Our proposed solution achieved 0.841, 0.817, and 0.476 macro F1-average in sub-tasks A, B, and C respectively.