Junfeng Liu

Ailin Huang, Chengyuan Yao, Chunrui Han et al.

We present STEP3-VL-10B, a lightweight open-source foundation model designed to redefine the trade-off between compact efficiency and frontier-level multimodal intelligence. STEP3-VL-10B is realized through two strategic shifts: first, a unified, fully unfrozen pre-training strategy on 1.2T multimodal tokens that integrates a language-aligned Perception Encoder with a Qwen3-8B decoder to establish intrinsic vision-language synergy; and second, a scaled post-training pipeline featuring over 1k iterations of reinforcement learning. Crucially, we implement Parallel Coordinated Reasoning (PaCoRe) to scale test-time compute, allocating resources to scalable perceptual reasoning that explores and synthesizes diverse visual hypotheses. Consequently, despite its compact 10B footprint, STEP3-VL-10B rivals or surpasses models 10$\times$-20$\times$ larger (e.g., GLM-4.6V-106B, Qwen3-VL-235B) and top-tier proprietary flagships like Gemini 2.5 Pro and Seed-1.5-VL. Delivering best-in-class performance, it records 92.2% on MMBench and 80.11% on MMMU, while excelling in complex reasoning with 94.43% on AIME2025 and 75.95% on MathVision. We release the full model suite to provide the community with a powerful, efficient, and reproducible baseline.

Junfeng Liu, Christopher Symons, Ranga Raju Vatsavai

Recent advances in machine learning and deep learning have led to the widespread use of Conversational AI in many practical applications. However, it is still very challenging to leverage auxiliary information that can provide conversational context or personalized tuning to improve the quality of conversations. For example, there has only been limited research on using an individuals persona information to improve conversation quality, and even state-of-the-art conversational AI techniques are unable to effectively leverage signals from heterogeneous sources of auxiliary data, such as multi-modal interaction data, demographics, SDOH data, etc. In this paper, we present a novel Persona-Coded Poly-Encoder method that leverages persona information in a multi-stream encoding scheme to improve the quality of response generation for conversations. To show the efficacy of the proposed method, we evaluate our method on two different persona-based conversational datasets, and compared against two state-of-the-art methods. Our experimental results and analysis demonstrate that our method can improve conversation quality over the baseline method Poly-Encoder by 3.32% and 2.94% in terms of BLEU score and HR@1, respectively. More significantly, our method offers a path to better utilization of multi-modal data in conversational tasks. Lastly, our study outlines several challenges and future research directions for advancing personalized conversational AI technology.

Jianqiang Xiao, Weiwen Guo, Junfeng Liu et al.

In the field of computer vision, data augmentation is widely used to enrich the feature complexity of training datasets with deep learning techniques. However, regarding the generalization capabilities of models, the difference in artificial features generated by data augmentation and natural visual features has not been fully revealed. This study introduces the concept of "visual representation variables" to define the possible visual variations in a task as a joint distribution of these variables. We focus on the visual representation variable "illumination", by simulating its distribution degradation and examining how data augmentation techniques enhance model performance on a classification task. Our goal is to investigate the differences in generalization between models trained with augmented data and those trained under real-world illumination conditions. Results indicate that after applying various data augmentation methods, model performance has significantly improved. Yet, a noticeable generalization gap still exists after utilizing various data augmentation methods, emphasizing the critical role of feature diversity in the training set for enhancing model generalization.

Qinghua Zhao, Jiaang Li, Lei Li et al.

Existing works have studied the impacts of the order of words within natural text. They usually analyze it by destroying the original order of words to create a scrambled sequence, and then comparing the models' performance between the original and scrambled sequences. The experimental results demonstrate marginal drops. Considering this findings, different hypothesis about word order is proposed, including ``the order of words is redundant with lexical semantics'', and ``models do not rely on word order''. In this paper, we revisit the aforementioned hypotheses by adding a order reconstruction perspective, and selecting datasets of different spectrum. Specifically, we first select four different datasets, and then design order reconstruction and continuing generation tasks. Empirical findings support that ChatGPT relies on word order to infer, but cannot support or negate the redundancy relations between word order lexical semantics.

PeiFeng Wang, Jonathan Zamora, Junfeng Liu et al.

Humans use natural language to compose common concepts from their environment into plausible, day-to-day scene descriptions. However, such generative commonsense reasoning (GCSR) skills are lacking in state-of-the-art text generation methods. Descriptive sentences about arbitrary concepts generated by neural text generation models (e.g., pre-trained text-to-text Transformers) are often grammatically fluent but may not correspond to human common sense, largely due to their lack of mechanisms to capture concept relations, to identify implicit concepts, and to perform generalizable reasoning about unseen concept compositions. In this paper, we propose an Imagine-and-Verbalize (I&V) method, which learns to imagine a relational scene knowledge graph (SKG) with relations between the input concepts, and leverage the SKG as a constraint when generating a plausible scene description. We collect and harmonize a set of knowledge resources from different domains and modalities, providing a rich auxiliary supervision signal for I&V. The experiments demonstrate the effectiveness of I&V in improving language models on both concept-to-sentence and concept-to-story generation tasks, while enabling the model to learn well from fewer task examples and generate SKGs that make common sense to human annotators.

Xing Zhao, Manos Papagelis, Aijun An et al.

The bulk synchronous parallel (BSP) is a celebrated synchronization model for general-purpose parallel computing that has successfully been employed for distributed training of machine learning models. A prevalent shortcoming of the BSP is that it requires workers to wait for the straggler at every iteration. To ameliorate this shortcoming of classic BSP, we propose ELASTICBSP a model that aims to relax its strict synchronization requirement. The proposed model offers more flexibility and adaptability during the training phase, without sacrificing on the accuracy of the trained model. We also propose an efficient method that materializes the model, named ZIPLINE. The algorithm is tunable and can effectively balance the trade-off between quality of convergence and iteration throughput, in order to accommodate different environments or applications. A thorough experimental evaluation demonstrates that our proposed ELASTICBSP model converges faster and to a higher accuracy than the classic BSP. It also achieves comparable (if not higher) accuracy than the other sensible synchronization models.

Xing Zhao, Aijun An, Junfeng Liu et al.

Deep learning is a popular machine learning technique and has been applied to many real-world problems. However, training a deep neural network is very time-consuming, especially on big data. It has become difficult for a single machine to train a large model over large datasets. A popular solution is to distribute and parallelize the training process across multiple machines using the parameter server framework. In this paper, we present a distributed paradigm on the parameter server framework called Dynamic Stale Synchronous Parallel (DSSP) which improves the state-of-the-art Stale Synchronous Parallel (SSP) paradigm by dynamically determining the staleness threshold at the run time. Conventionally to run distributed training in SSP, the user needs to specify a particular staleness threshold as a hyper-parameter. However, a user does not usually know how to set the threshold and thus often finds a threshold value through trial and error, which is time-consuming. Based on workers' recent processing time, our approach DSSP adaptively adjusts the threshold per iteration at running time to reduce the waiting time of faster workers for synchronization of the globally shared parameters, and consequently increases the frequency of parameters updates (increases iteration throughput), which speedups the convergence rate. We compare DSSP with other paradigms such as Bulk Synchronous Parallel (BSP), Asynchronous Parallel (ASP), and SSP by running deep neural networks (DNN) models over GPU clusters in both homogeneous and heterogeneous environments. The results show that in a heterogeneous environment where the cluster consists of mixed models of GPUs, DSSP converges to a higher accuracy much earlier than SSP and BSP and performs similarly to ASP. In a homogeneous distributed cluster, DSSP has more stable and slightly better performance than SSP and ASP, and converges much faster than BSP.

Jun Long, Qunfeng Liu, Xinpan Yuan et al.

Image similarity measures play an important role in nearest neighbor search and duplicate detection for large-scale image datasets. Recently, Minwise Hashing (or Minhash) and its related hashing algorithms have achieved great performances in large-scale image retrieval systems. However, there are a large number of comparisons for image pairs in these applications, which may spend a lot of computation time and affect the performance. In order to quickly obtain the pairwise images that theirs similarities are higher than the specific threshold T (e.g., 0.5), we propose a dynamic threshold filter of Minwise Hashing for image similarity measures. It greatly reduces the calculation time by terminating the unnecessary comparisons in advance. We also find that the filter can be extended to other hashing algorithms, on when the estimator satisfies the binomial distribution, such as b-Bit Minwise Hashing, One Permutation Hashing, etc. In this pager, we use the Bag-of-Visual-Words (BoVW) model based on the Scale Invariant Feature Transform (SIFT) to represent the image features. We have proved that the filter is correct and effective through the experiment on real image datasets.

Yicheng He, Junfeng Liu, Xia Ning

Selecting the right drugs for the right patients is a primary goal of precision medicine. In this manuscript, we consider the problem of cancer drug selection in a learning-to-rank framework. We have formulated the cancer drug selection problem as to accurately predicting 1). the ranking positions of sensitive drugs and 2). the ranking orders among sensitive drugs in cancer cell lines based on their responses to cancer drugs. We have developed a new learning-to-rank method, denoted as pLETORg , that predicts drug ranking structures in each cell line via using drug latent vectors and cell line latent vectors. The pLETORg method learns such latent vectors through explicitly enforcing that, in the drug ranking list of each cell line, the sensitive drugs are pushed above insensitive drugs, and meanwhile the ranking orders among sensitive drugs are correct. Genomics information on cell lines is leveraged in learning the latent vectors. Our experimental results on a benchmark cell line-drug response dataset demonstrate that the new pLETORg significantly outperforms the state-of-the-art method in prioritizing new sensitive drugs.