Quang Minh Tran

Author directory

Also published as: Quang-Minh Tran


2026

Direct Preference Optimization (DPO) is well suited to verifiable reasoning tasks because preference pairs can be constructed automatically from answer correctness. However, in iterative DPO, correctness alone does not determine whether a pair provides a useful training signal. We distinguish two regimes according to the origin of the chosen response: On-policy pairs, where both the chosen and rejected responses are sampled from the current policy, and Fallback pairs, where the policy fails to generate a correct response and the chosen completion is supplied by a dataset-provided reference solution. Based on this distinction, we propose support-aware filtering, which applies margin-based filtering to On-policy pairs and a likelihood-based support gate to Fallback pairs. Experiments on GSM8K and MATH-500 with Qwen2.5-style 0.5B, 1.5B, and 3B models show that support-aware filtering yields its clearest gains on MATH-500 at 1.5B and 3B, while GSM8K and the 0.5B MATH-500 setting are more mixed. Overall, our findings suggest that preference margins in iterative verifiable DPO are pair-origin dependent, and that chosen-response support is a useful signal for stabilizing fallback supervision.

2024

Task-oriented dialogue (TOD) systems are introduced to solve specific tasks, which focus on training multiple tasks such as language understanding, tracking states, and generating appropriate responses to help users achieve their specific goals. Currently, one of the remaining challenges in this emergent research field is the capability to produce more robust architectures fine-tuned for end-to-end TOD systems. In this study, we consider this issue by exploiting the ability of pre-trained models to provide synthesis responses, which are then used as the input for the fine-tuned process. The main idea is to overcome the gap between the training process and inference process during fine-tuning end-to-end TOD systems. The experiment on Multiwoz datasets shows the effectiveness of our model compared with strong baselines in this research field. The source code is available for further exploitation.

2023

We propose an ensemble method that combines several pre-trained language models to enhance entity recognition in legal text. Our approach achieved a 90.9873% F1 score on the private test set, ranking 2nd on the leaderboard for SemEval 2023 Task 6, Subtask B - Legal Named Entities Extraction.