This paper describes our system for the SemEval-2026 Task 12 on Abductive Event Reasoning (AER). We systematically address the “over-selection” hallucination pathology in Instruction-tuned Large Language Models (LLMs), where models erroneously align distractors with semantic priors rather than retrieved evidence. Our framework utilizes a 32-billion parameter Qwen2.5 foundational model adapted via Low-Rank Adaptation (LoRA) and evaluated under a Zero-shot Chain-of-Thought (CoT) setting. To mitigate epistemic noise, we propose a Precision-Oriented Decoding (POD) strategy that couples low-temperature sampling (T=0.45) with scaled majority voting (K=9). Following a three-stage empirical evolution—from baseline diagnosis to precision optimization and ensemble analysis—our system achieved a score of 0.802 on the official test set. Our findings demonstrate that in causal reasoning tasks with strict penalization for incorrect predictions, epistemic noise suppression is strictly superior to heuristic recall compensation.

In recent years, large language models (LLMs) have made rapid progress in information retrieval, yet existing research has mainly focused on text or static multimodal settings. Open-domain video shot retrieval, which involves richer temporal structure and more complex semantics, still lacks systematic benchmarks and analysis. To fill this gap, we introduce ShotFinder, a benchmark that formalizes editing requirements as keyframe-oriented shot descriptions and introduces five types of controllable single-factor constraints: Temporal order, Color, Visual style, Audio, and Resolution. We curate 1,210 high-quality samples from YouTube across 20 thematic categories, using large models for generation with human verification. Based on the benchmark, we propose ShotFinder, a text-driven three-stage retrieval and localization pipeline: (1) query expansion via video imagination, (2) candidate video retrieval with a search engine, and (3) description-guided shot localization. Experiments on multiple closed-source and open-source models reveal a significant gap to human performance, with clear imbalance across constraints: temporal localization is relatively tractable, while color and visual style remain major challenges. These results reveal that open-domain video shot retrieval is still a critical capability that multimodal large models have yet to overcome.

The widespread deployment of large language models (LLMs) across diverse domains has underscored the critical need to ensure the credibility and accuracy of their generated content, particularly in the presence of hallucinations. These hallucinations can severely compromise both the practical performance of models and the security of their applications. In response to this issue, SemEval-2025 Task 3 Mu-SHROOM: Multilingual Shared-task on Hallucinations and Related Observable Overgeneration Mistakes introduces a more granular task for hallucination detection. This task seeks to identify hallucinations in text, accurately locate hallucinated segments, and assess their credibility. In this paper, we present a three-stage method for fine-grained hallucination detection and localization. First, we transform the text into a triplet representation, facilitating more precise hallucination analysis. Next, we leverage a large language model to generate fact-reference texts that correspond to the triplets. Finally, we employ a fact alignment strategy to identify and localize hallucinated segments by evaluating the semantic consistency between the extracted triplets and the generated reference texts. We evaluate our method on the unlabelled test set across all languages in Task 3, demonstrating strong detection performance and validating its effectiveness in multilingual contexts.