Author name cluster

Xiaohui Song

Possible papers associated with this exact author name in Arrow. This page groups case-insensitive exact name matches and is not a full identity disambiguation profile.

5 papers

2 author rows

AAAI Conference 2026 Conference Paper

DRIFT: Difference-Aware Reinforcement Through Iterative Fine-Tuning for Language Model

Wenjie Liao
Xiaohui Song
Haonan Lu

Self-play fine-tuning has emerged as a promising approach to improve Large Language Models (LLMs) without additional human annotations. However, existing methods struggle with complex generation tasks requiring long context understanding, where models produce partially correct outputs interleaved with errors. Traditional approaches train on entire sequences uniformly, failing to distinguish between well-predicted and erroneous regions, leading to diluted learning signals and slow convergence. We propose DRIFT (Difference-aware Reinforcement through Iterative Fine-Tuning), a novel self-play framework that selectively trains on prediction differences. DRIFT introduces two key innovations: (1) Difference-Aware Masking (DAM) that identifies and masks common subsequences between model outputs and ground truth, focusing training exclusively on error regions; (2) Occurrence-Aware Loss (OAL) that provides position-invariant vocabulary supervision, complementing the position-sensitive adversarial loss. This dual mechanism enables models to correct both positional and lexical errors effectively. Theoretically, we prove that DRIFT converges when masked distributions align. Empirically, we evaluate DRIFT on diverse summarization benchmarks using Qwen2.5-3B and LLaMA-3.1-8B models. Results show that DRIFT significantly outperforms both supervised fine-tuning (SFT) and self-play fine-tuning (SPIN), achieving up to 16\% improvement on SAMSum dialogue summarization tasks while maintaining general capabilities. Notably, DRIFT breaks the performance ceiling of continued SFT and demonstrates superior efficiency compared to holistic self-play methods, validating that targeted optimization on prediction differences is crucial for structured text generation tasks.

PDF Details DOI

ICML Conference 2025 Conference Paper

HealthGPT: A Medical Large Vision-Language Model for Unifying Comprehension and Generation via Heterogeneous Knowledge Adaptation

Tianwei Lin 0001
Wenqiao Zhang
Sijing Li
Yuqian Yuan
Binhe Yu
Haoyuan Li 0002
Wanggui He
Hao Jiang 0014

We present HealthGPT, a powerful Medical Large Vision-Language Model (Med-LVLM) that integrates medical visual comprehension and generation capabilities within a unified autoregressive paradigm. Our bootstrapping philosophy is to progressively adapt heterogeneous comprehension and generation knowledge to pre-trained Large Language Models (LLMs). This is achieved through a novel heterogeneous low-rank adaptation (H-LoRA) technique, which is complemented by a tailored hierarchical visual perception (HVP) approach and a three-stage learning strategy (TLS). To effectively learn the HealthGPT, we devise a comprehensive medical domain-specific comprehension and generation dataset called VL-Health. Experimental results demonstrate exceptional performance and scalability of HealthGPT in medical visual unified tasks. Our project can be accessed at https: //github. com/DCDmllm/HealthGPT.

Details

EAAI Journal 2024 Journal Article

Evolutionary algorithm incorporating reinforcement learning for energy-conscious flexible job-shop scheduling problem with transportation and setup times

Guohui Zhang
Shaofeng Yan
Xiaohui Song
Deyu Zhang
Shenghui Guo

Flexible job-shop scheduling is considerably important in the modern intelligent manufacturing factory. In a real job shop, transportation and setup times account for a large percentage of the total processing flow, and with today's companies demanding higher delivery times, the feasibility and punctuality of scheduling will be considerably reduced if these time constraints are ignored. Recently, several companies have become green in their manufacturing processes. However, transportation, setup, and delivery times have rarely been combined with energy efficiency. To solve this problem, we employed an integer programming approach to develop a complete mathematical model of the problem and simultaneously optimized four objectives: maximum completion time, total energy consumption, workload of critical machines, and penalties for earliness/tardiness. Subsequently, an evolutionary algorithm incorporating reinforcement learning was proposed to solve the model. The algorithm had the following features: (1) four initialization strategies were designed to obtain high-quality populations; (2) a reinforcement learning-based parameter-adaptive strategy was proposed to guide the population to select the best parameters; (3) a critical path-based neighborhood structure with transportation and setup times was designed, and according to the objectives of this study, four additional neighborhood structures were designed; (4) a reference point-based non-dominated sorting selection was presented to guide the solution toward the Pareto-optimal front; and (5) an external archive was proposed to enhance the utilization of abandoned historical solutions. Finally, the effectiveness of this algorithm was demonstrated using 33 benchmark instances of variants and comparison experiments.

Details DOI

ECAI Conference 2024 Conference Paper

JOSAL: Joint Learning Framework for Open-Set Active Learning

Jun Xie
Xiaohui Song
Yangjie Cao
Zhi Liu 0002
Weiping Wang
Hongli Xu

Previous research in active learning has primarily focused on selecting examples from closed-set data, which consists solely of unlabeled examples from the target classes. However, this approach overlooks the more prevalent scenario of open-set data in real-world applications. Open-set data encompasses examples from both target classes and non-target classes. To fill this gap, we propose a novel framework called JOSAL, which enhances the accuracy of the classifier by precisely selecting the target class examples from open-set data. The JOSAL framework introduces the concept of joint learning, where the Sampler and Classifier components perform sampling and classification tasks, respectively, by sharing example features extracted from a pre-trained Encoder. To maximize the classification accuracy of the Classifier, the framework adopts a novel joint learning strategy. This strategy initially prioritizes optimizing the Sampler and gradually shifts the optimization attention to the Classifier. The experimental results demonstrate that, compared to baselines, our approach exhibits stronger sampling precision and achieves higher classification accuracy. To the best of our knowledge, this is the first work to address the open-set active learning problem using the joint learning paradigm.

Details

EAAI Journal 2024 Journal Article

Unsupervised video forecasting with flow parsing mechanism of human visual system

Beibei Jin
Xiaohui Song
Jindong Li
Pengfei Zhang

Video forecasting aims to predict future video frames based on past observed video frames, and unlike object recognition or object classification, it does not require manual labeling of the data set. The explosive growth of Internet video data provides a huge space for its development. At present, it has become a research hotspot in the field of computer vision, and has broad application prospects in the field of automatic driving or robot navigation. However, due to the high dimensional characteristics and the complex spatial–temporal logic of video data, current methods still face the challenges of blurry and inconsistent prediction. The cognitive ability of “flow parsing mechanism” helps humans adapt to new situations systematically. Inspired by this, a deep flow parsing network for future video forecasting is proposed in this paper, which is designed to predict future scenes by parsing optical flow into rigid flow and residual flow. The rigid flow represents the scene dynamics due to observer’s ego-motion, while the residual flow corresponds to the movement of the other objects in the scene. With this procedure, the model exhibits much more comprehensive understanding over the environment and achieves top performance on competitive driving datasets, demonstrating its effectiveness and generalizability.

Details DOI