Positional bias (PB), manifesting as non-uniform sensitivity across different contextual locations, significantly impairs long-context comprehension and processing capabilities. Previous studies have addressed PB either by modifying the underlying architectures or by employing extensive contextual awareness training. However, the former approach fails to effectively eliminate the substantialperformance disparities, while the latter imposes significant data and computational overhead. To address PB effectively, we introduce Pos2Distill, a position to position knowledge distillation framework. Pos2Distill transfers the superior capabilities from advantageous positions to less favorable ones, thereby reducing the huge performance gaps. The conceptual principle is to leverage the inherent, position-induced disparity to counteract the PB itself. We identify distinct manifestations of PB under retrieval and reasoning paradigms, thereby designing two specialized instantiations: Pos2Distill-R1 and Pos2Distill-R2 respectively, both grounded in this core principle. By employing the Pos2Distill approach, we achieve enhanced uniformity and significant performance gains across all contextual positions in long-context retrieval and reasoning tasks. Crucially, both specialized systems exhibit strong cross-task generalization mutually, while achieving superior performance on their respective tasks.

Self-taught reasoners (STaRs) enhance the mathematical reasoning abilities of large language models (LLMs) by leveraging self-generated responses for self-training. Recent studies have incorporated reward models to guide response selection or decoding, aiming to obtain higher-quality data. However, they typically allocate a uniform sampling budget across all problems, overlooking the varying utility of problems at different difficulty levels. In this work, we conduct an empirical study and find that problems near the boundary of the LLM’s reasoning capability offer significantly greater learning utility than both easy and overly difficult ones. To identify and exploit such problems, we propose HS-STaR, a Hierarchical Sampling framework for Self-Taught Reasoners. Given a fixed sampling budget, HS-STaR first performs lightweight pre-sampling with a reward-guided difficulty estimation strategy to efficiently identify boundary-level problems. Subsequently, it dynamically reallocates the remaining budget toward these high-utility problems during a re-sampling phase, maximizing the generation of valuable training data. Extensive experiments across multiple reasoning benchmarks and backbone LLMs demonstrate that HS-STaR significantly outperforms other baselines without requiring additional sampling budget.

This paper introduces the system developed by the HITSZ-HLT team for SemEval-2025 Task 8: DataBench, Question-Answering over Tabular Data.The primary objective of Table Question Answering (TableQA) is to provide accurate answers to user queries by interpreting and understanding tabular data. To address this, we propose the Multi-turn Interactive Code GeneratiOn(MICO) framework. Specifically, MICO employs code generation as proxy task for TableQA and integrates feedback from the execution of the generated code via multi-turn dialogue process, thereby guiding the model towards self-correction.Experimental results demonstrate the effectiveness of our framework, achieving notable performance with a rank of 4/38 on the DataBench and 5/38 on the DataBench lite.

SemEval-2024 Task 8 introduces the challenge of identifying machine-generated texts from diverse Large Language Models (LLMs) in various languages and domains. The task comprises three subtasks: binary classification in monolingual and multilingual (Subtask A), multi-class classification (Subtask B), and mixed text detection (Subtask C). This paper focuses on Subtask A & B. To tackle this task, this paper proposes two methods: 1) using traditional machine learning (ML) with natural language preprocessing (NLP) for feature extraction, and 2) fine-tuning LLMs for text classification. For fine-tuning, we use the train datasets provided by the task organizers. The results show that transformer models like LoRA-RoBERTa and XLM-RoBERTa outperform traditional ML models, particularly in multilingual subtasks. However, traditional ML models performed better than transformer models for the monolingual task, demonstrating the importance of considering the specific characteristics of each subtask when selecting an appropriate approach.

This paper describes the system developed by the HITSZ-HLT team for WASSA-2024 Shared Task 2, which addresses two closely linked sub-tasks: Cross-lingual Emotion Detection and Binary Trigger Word Detection in tweets. The main goal of Shared Task 2 is to simultaneously identify the emotions expressed and detect the trigger words across multiple languages. To achieve this, we introduce a Language-agnostic Multi Task Learning (LaMTL) framework that integrates emotion prediction and emotion trigger word detection tasks. By fostering synergistic interactions between task-specific and task-agnostic representations, the LaMTL aims to mutually enhance emotional cues, ultimately improving the performance of both tasks. Additionally, we leverage large-scale language models to translate the training dataset into multiple languages, thereby fostering the formation of language-agnostic representations within the model, significantly enhancing the model’s ability to transfer and perform well across multilingual data. Experimental results demonstrate the effectiveness of our framework across both tasks, with a particular highlight on its success in achieving second place in sub-task 2.