Hidetsune at SemEval-2026 Task 11: Adapting Pretrained Reasoning Models with Deep Supervision and Inference Refinement for Content-Independent Validity Classification

AbstractThis paper presents a system that applies training and inference approaches for SemEval2026 Task 11 Subtask 1, which focuses on binary classification for content-independent validity reasoning in syllogistic inference. Building on fine-tuning of relatively standard language models, additional approaches were explored, including layer-wise deep supervision and in-context learning. Furthermore, models that had been previously trained on datasets related to logical reasoning were adapted to thetask through additional fine-tuning. Finally, refinement was performed at the inference stage by adjusting the softmax-based decision threshold of the selected model. The experimental results illustrate how model selection, training strategies, and threshold adjustment affect not only validity accuracy but also robustness against plausibility-driven bias, thereby contributing to improved logical integrity.