The research team led by Professor Song Dawei at the School of Computer Science and Technology, Beijing Institute of Technology (BIT), has recently made significant strides in the fields of lightweight large language models, value alignment, retrieval enhancement, reasoning optimization, and downstream applications such as machine translation and sentiment analysis.

Four papers from the team have been accepted at the top-tier international conference, the 64th Annual Meeting of the Association for Computational Linguistics (ACL) 2026, following their receipt of the "Outstanding Paper Award" at ACL 2025.

The ACL is a premier international conference in artificial intelligence, computational linguistics, and natural language processing.

ACL 2025 was held in Vienna, Austria, from July 27 to August 1, where Ph.D. student Zhang Chen's paper, "Towards the Law of Capacity Gap in Distilling Language Models", won the "Outstanding Paper Award". This paper introduced the groundbreaking law of capacity gap in model distillation, revealing a near-linear proportional relationship between the optimal sizes of teacher and student models. The application of this law led to the creation of a 3B model that outperformed contemporary baseline models of similar size, setting a new computational-performance Pareto frontier.

ACL 2026 is scheduled to take place from July 2 to July 7 in San Diego, California, the United States. The acceptance rate for the main conference is 19 percent, with the findings section having an acceptance rate of 18 percent.

The team's four accepted papers were authored by master's graduate Li Zelin, Ph.D. student Tian Yanzhi (co-supervised by Dr. Guo Yuhang from the School of Computer Science and Technology), Sui Yi, and Meng Lingang. The accepted papers include: Reward Alignment Optimization: A Direct Point-wise Alignment Approach (Main), Beyond Literal Mapping: Benchmarking and Improving Non-Literal Translation Evaluation (Main), Think Less, Know More: State-Aware Reasoning Compression with Knowledge Guidance for Efficient Reasoning (Findings), and Beyond Polarity: Continuous Affect-Enhanced Multimodal Aspect-Based Sentiment Classification (Findings).