Large language models (LLM) have prioritized expanding the context window from which models can incorporate more information. However, training models to handle long contexts presents significant challenges. These include the scarcity of high-quality natural long-context data, the potential for performance degradation on short-context tasks, and the reduced training efficiency associated with attention mechanisms. In this paper, we introduce Untie the Knots (UtK), a novel data augmentation strategy employed during the continue pre-training phase, designed to efficiently enable LLMs to gain long-context capabilities without the need to modify the existing data mixture. In particular, we chunk the documents, shuffle the chunks, and create a complex and knotted structure of long texts; LLMs are then trained to untie these knots and identify relevant segments within seemingly chaotic token sequences. This approach greatly improves the model’s performance by accurately attending to relevant information in long context and the training efficiency is also largely increased. We conduct extensive experiments on models with 7B and 72B parameters, trained on 20 billion tokens, demonstrating that UtK achieves 75% and 84.5% accurracy on RULER at 128K context length, significantly outperforming other long context strategies. The trained models will open-source for further research.

This paper provides an overview of the Hallucination Detection for Scientific Content (SciHal) shared task held in the 2025 ACL Scholarly Document Processing workshop. The task invites participants to detect hallucinated claims in answers to research-oriented questions generated by real-world GenAI-powered research assistants. This task is formulated as a multi-label classification problem, each instance consists of a question, an answer, an extracted claim, and supporting reference abstracts. Participants are asked to label claims under two subtasks: (1) coarse-grained detection with labels Entailment, Contradiction, or Unverifiable; and (2) fine-grained detection with a more detailed taxonomy including 8 types.The dataset consists of 500 research-oriented questions collected over one week from a generative assistant tool. These questions were rewritten using GPT-4o and manually reviewed to address potential privacy or commercial concerns. In total, 10,000 reference abstracts were retrieved, and 4,592 claims were extracted from the assistant’s answers. Each claim is annotated with hallucination labels. The dataset is divided into 3,592 training, 500 validation, and 500 test instances.Subtask 1 saw 88 submissions across 10 teams while subtask 2 saw 39 submissions across 6 teams, resulting in a total of 5 published technical reports. This paper summarizes the task design, dataset, participation, and key findings.