Large language models (LLMs) have revolutionized numerous fields of research, driving significant advancements in natural language processing, machine translation, and beyond. Although the extensive number of parameters contributes a lot to the great success, existing studies indicate that not all model parameters hold equal importance, which further leads to redundancy during the parameter update process. Recent works for reducing redundant parameter updates for LLMs either lack task-specific data information, may leading to suboptimal model performance, or discard transformer components or insignificant parameters, limiting the model's scalability across different tasks and potentially compromising the LLM structure. To address these issues and further enhance the performance of LLMs, we propose Gradient-Mask Tuning (GMT), a method that selectively updates parameters based on gradient information, which is specific to the target tasks. Specifically, after calculating gradients during back propagation, we measure their absolute values and mask those with small absolute values. Our empirical results in various training paradigms like SFT and DPO for various domains of tasks demonstrate that GMT not only preserves the original network structure but also enhances the potential performance of LLMs. Further analysis indicates that GMT exhibits insensitivity to mask ratio and possesses computational efficiency comparable to vanilla training approach.

It is well-known that a diverse corpus is critical for training large language models, which are typically constructed from a mixture of various domains.In general, previous efforts resort to either sampling training data from different domains with static proportions or dynamically adjusting these proportions during training to optimise pretraining performance.However, few methods addressed the complexity of domain-adaptive continual pre-training.To fill this gap, we propose Velocitune, a novel framework that dynamically assesses learning velocity and adjusts data proportions accordingly, favouring slower learning domains while de-emphasising faster learning ones, which is guided by a scaling law to estimate the desired learning goal for each domain with a less associated cost.To evaluate the effectiveness of Velocitune, we conduct experiments on a dataset focused on reasoning tasks with CodeLlama, as well as on a corpus of system commands using Llama3 and Mistral.Velocitune achieves performance gains in both math and code reasoning tasks and command-line generation benchmarks.Further analysis reveals that key factors driving the effectiveness of Velocitune include target estimation and data ordering.* The first two authors contribute equally.

Diffusion models have achieved significant progress in both image and video generation while still suffering from huge computation costs. As an effective solution, rectified flow aims to rectify the diffusion process of diffusion models into a straight line for few-step and even one-step generation. However, in this paper, we suggest that the original training pipeline of reflow is not optimal and introduce two techniques to improve it. Firstly, we introduce progressive reflow, which progressively reflows the diffusion models in local timesteps until the whole diffusion progresses, reducing the difficulty of flow matching. Second, we introduce aligned v-prediction, which highlights the importance of direction matching in flow matching over magnitude matching. Experimental results on SDv1.5 and SDXL demonstrate the effectiveness of our method, for example, conducting on SDv1.5 achieves an FID of 10.70 on MSCOCO2014 validation set with only 4 sampling steps, close to our teacher model (32 DDIM steps, FID = 10.05). Our codes will be released at Github.

Jie Wu, Haoling Li, Xin Zhang, Xiao Liu, Yangyu Huang, Jianwen Luo, Yizhen Zhang, Zuchao Li, Ruihang Chu, Yujiu Yang, Scarlett Li. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Existing methods for code generation use code snippets as seed data, restricting the complexity and diversity of the synthesized data. In this paper, we introduce a novel feature tree-based synthesis framework, which revolves around hierarchical code features derived from high-level abstractions of code. The feature tree is constructed from raw data and refined iteratively to increase the quantity and diversity of the extracted features, which captures and recognizes more complex patterns and relationships within the code. By adjusting the depth and breadth of the sampled subtrees, our framework provides precise control over the complexity of the generated code, enabling functionalities that range from function-level operations to multi-file scenarios. We fine-tuned widely-used base models to obtain EpiCoder series, achieving state-of-the-art performance on multiple benchmarks at both the function and file levels. In particular, empirical evidence indicates that our approach shows significant potential in the synthesizing of repository-level code data. Our code and data are publicly available.

Inspired by the impressive reasoning capabilities demonstrated by reinforcement learning approaches like DeepSeek-R1, recent emerging research has begun exploring the use of reinforcement learning (RL) to enhance vision-language models (VLMs) for multimodal reasoning tasks. However, most existing multimodal reinforcement learning approaches remain limited to spatial reasoning within single-image contexts, yet still struggle to generalize to more complex and real-world scenarios involving multi-image positional reasoning, where understanding the relationships across images is crucial. To address this challenge, we propose a general reinforcement learning approach PeRL tailored for interleaved multimodal tasks, and a multi-stage strategy designed to enhance the exploration-exploitation trade-off, thereby improving learning efficiency and task performance. Specifically, we introduce permutation of image sequences to simulate varied positional relationships to explore more spatial and positional diversity. Furthermore, we design a rollout filtering mechanism for resampling to focus on trajectories that contribute most to learning optimal behaviors to exploit learned policies effectively. We evaluate our model on 5 widely-used multi-image benchmarks and 3 single-image benchmarks. Our experiments confirm that PeRL trained model consistently surpasses R1-related and interleaved VLM baselines by a large margin, achieving state-of-the-art performance on multi-image benchmarks, while preserving comparable performance on single-image tasks.

Neural networks (NNs) and decision trees (DTs) are both popular models of machine learning, yet coming with mutually exclusive advantages and limitations. To bring the best of the two worlds, a variety of approaches are proposed to integrate NNs and DTs explicitly or implicitly. In this survey, these approaches are organized in a school which we term neural trees (NTs). This survey aims to present a comprehensive review of NTs and explore in detail how they enhance the model interpretability. Our first contribution is a detailed taxonomy of NTs, which characterizes the seamless integration and co-evolution of NNs and DTs. Subsequently, we analyze NTs in terms of their interpretability and performance and suggest potential solutions to the remaining challenges. Finally, this survey concludes with a discussion about other considerations like conditional computation and promising directions toward this field. A list of papers reviewed in this survey, along with their corresponding codes, is available at: https://github.com/ zju-vipa/awesome-neural-trees.

Yifan Wang, Yafei Liu, Chufan Shi, Haoling Li, Chen Chen, Haonan Lu, Yujiu Yang. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

The faint vision at low illumination affects the performance of intelligent surveillance systems and induces criminals to sin under the cover of darkness. Night-to-day translation is an ideal way to handle this problem, but hard to achieve due to the lack of information at night. We propose a novel approach that combines DCGAN(deep convolutional generative adversarial network) and image processing algorithms to find out the mapping from night to day. Images from night domain are enhanced with MSRCP(multi-scale retinex with chromaticity preservation) algorithm before they’re put into DCGAN to generate bright and clear daylight images without paired supervision. At the same time emerging issues of image atomization and local over-exposure are handled to ensure the quality of output. The experimental results show that our approach can be applied to different conditions and dig out sensitive information from the darkness.