This paper presents our approach for the SemEval-2026 Task 8: MTRAGEval (SubtaskB: Answer Generation), which challenges sys-tems to generate faithful, extractive answers to multi-turn questions based strictly on provided gold-standard reference passages. The primary scientific challenge lies in maintaining high faithfulness and structural consistency while adapting to diverse answer styles across a conversation, as systems must generate responses that vary significantly in length and format without hallucinating. Conventional reference-based generation methods often rely on static prompting or greedy decoding, which fail to capture these dynamic stylistic requirements and lack robustness against generation noise. To address these limitations, we propose a Intent-Aware Parallel Generation and Reranking System powered by a large language model. Experimental results on the official test set demonstrate the effectiveness of our method, achieving competitive performance comparable to SoTA baselines. Ultimately,our approach secured the third place in the competition. The code of the paper is available at: https://github.com/viaviachris/SemEval-2026-Task8

This paper describes our systems for SemEval-2026 Task 11, Disentangling Content and Formal Reasoning in Language Models. We participated in all four subtasks, addressing the Content Effect-a phenomenon where models rely on real-world plausibility rather than logical validity. Existing methods, such as standard Chain-of-Thought (CoT) prompting or single-task Supervised Fine-Tuning (SFT), often struggle to completely decouple content from reasoning due to the inherent probabilistic biases in pre-trained models. To address these limitations, a hybrid neuro-symbolic framework based on the Qwen2.5-14B architecture is proposed, integrating multi-task instruction tuning with a robust neuro-symbolic pipeline. The principal innovation lies in the deployment of a Reflexion mechanism coupled with formal verification: natural language arguments are parsed into First-Order Logic (FOL) and subsequently verified by the Z3 Theorem Prover. Parsing anomalies are automatically rectified through an iterative self-correction module. The proposed system ranked 1st in Subtasks 1 & 2, 2nd in Subtask 4, and 9th in Subtask 3, validating its ability to decouple logical validity from content plausibility.

In this paper, we present our submission to the SemEval-2026 Psycholinguistic Conspiracy Shared Task (Task 10), which consists of two tasks: conspiracy marker extraction and conspiracy detection. For conspiracy marker extraction, we formulate the problem as a token classification task and fine-tune pretrained language models, achieving performance above the official baseline and ranking 6th on the final leaderboard. For conspiracy detection, we apply data preprocessing, regularization, and ensemble inference strategies,resulting in improvements over the baseline and a 10th-place ranking. Overall, our results demonstrate the effectiveness of pretrained language models for both tasks.