Hard-negative preference optimization should improve support for ground-truth preferred responses while limiting drift from the reference model. In Japanese medical instruction tuning, we find that standard DPO and IPO improve the preferred–dispreferred gap but can leave preferred-response likelihood below the reference model and induce growing sequence-level KL divergence. Motivated by this failure-mode analysis, we propose a stabilized IPO objective that combines preferred-response anchoring with explicit KL control. Ablations show that positive anchoring improves preferred-response support, while KL control suppresses reference-model drift; their combination preserves anchoring gains while further reducing drift. These results suggest that reliable medical preference optimization benefits from both preferred-response anchoring and drift control.

In this paper, we present our approach to addressing the binary classification tasks, Tasks 5 and 6, as part of the Social Media Mining for Health (SMM4H) text classification challenge. Both tasks involved working with imbalanced datasets that featured a scarcity of positive examples. To mitigate this imbalance, we employed a Large Language Model to generate synthetic texts with positive labels, aiming to augment the training data for our text classification models. Unfortunately, this method did not significantly improve model performance. Through clustering analysis using text embeddings, we discovered that the generated texts significantly lacked diversity compared to the raw data. This finding highlights the challenges of using synthetic text generation for enhancing model efficacy in real-world applications, specifically in the context of health-related social media data.

This paper introduces a zero-shot hate detection experiment using a multimodal large model. Although the implemented model comprises an unsupervised method, results demonstrate that its performance is comparable to previous supervised methods. Furthemore, this study proposed experiments with various prompts and demonstrated that simpler prompts, as opposed to the commonly used detailed prompts in large language models, led to better performance for multimodal hate speech event detection tasks. While supervised methods offer high performance, they require significant computational resources for training, and the approach proposed here can mitigate this issue.The code is publicly available at https://github.com/yamagishi0824/zeroshot-hate-detect.