Chao Song

Wenhui Ren, Weikang Li, Shibo Xu et al. · tsinghua

Quantum computing promises to enhance machine learning and artificial intelligence. Different quantum algorithms have been proposed to improve a wide spectrum of machine learning tasks. Yet, recent theoretical works show that, similar to traditional classifiers based on deep classical neural networks, quantum classifiers would suffer from the vulnerability problem: adding tiny carefully-crafted perturbations to the legitimate original data samples would facilitate incorrect predictions at a notably high confidence level. This will pose serious problems for future quantum machine learning applications in safety and security-critical scenarios. Here, we report the first experimental demonstration of quantum adversarial learning with programmable superconducting qubits. We train quantum classifiers, which are built upon variational quantum circuits consisting of ten transmon qubits featuring average lifetimes of 150 $μ$s, and average fidelities of simultaneous single- and two-qubit gates above 99.94% and 99.4% respectively, with both real-life images (e.g., medical magnetic resonance imaging scans) and quantum data. We demonstrate that these well-trained classifiers (with testing accuracy up to 99%) can be practically deceived by small adversarial perturbations, whereas an adversarial training process would significantly enhance their robustness to such perturbations. Our results reveal experimentally a crucial vulnerability aspect of quantum learning systems under adversarial scenarios and demonstrate an effective defense strategy against adversarial attacks, which provide a valuable guide for quantum artificial intelligence applications with both near-term and future quantum devices.

Lixia Wu, Jianlin Liu, Junhong Lou et al.

Text-based delivery addresses, as the data foundation for logistics systems, contain abundant and crucial location information. How to effectively encode the delivery address is a core task to boost the performance of downstream tasks in the logistics system. Pre-trained Models (PTMs) designed for Natural Language Process (NLP) have emerged as the dominant tools for encoding semantic information in text. Though promising, those NLP-based PTMs fall short of encoding geographic knowledge in the delivery address, which considerably trims down the performance of delivery-related tasks in logistic systems such as Cainiao. To tackle the above problem, we propose a domain-specific pre-trained model, named G2PTL, a Geography-Graph Pre-trained model for delivery address in Logistics field. G2PTL combines the semantic learning capabilities of text pre-training with the geographical-relationship encoding abilities of graph modeling. Specifically, we first utilize real-world logistics delivery data to construct a large-scale heterogeneous graph of delivery addresses, which contains abundant geographic knowledge and delivery information. Then, G2PTL is pre-trained with subgraphs sampled from the heterogeneous graph. Comprehensive experiments are conducted to demonstrate the effectiveness of G2PTL through four downstream tasks in logistics systems on real-world datasets. G2PTL has been deployed in production in Cainiao's logistics system, which significantly improves the performance of delivery-related tasks. The code of G2PTL is available at https://huggingface.co/Cainiao-AI/G2PTL.

Jiansheng Fang, Jingwen Wang, Anwei Li et al.

In the management of lung nodules, we are desirable to predict nodule evolution in terms of its diameter variation on Computed Tomography (CT) scans and then provide a follow-up recommendation according to the predicted result of the growing trend of the nodule. In order to improve the performance of growth trend prediction for lung nodules, it is vital to compare the changes of the same nodule in consecutive CT scans. Motivated by this, we screened out 4,666 subjects with more than two consecutive CT scans from the National Lung Screening Trial (NLST) dataset to organize a temporal dataset called NLSTt. In specific, we first detect and pair regions of interest (ROIs) covering the same nodule based on registered CT scans. After that, we predict the texture category and diameter size of the nodules through models. Last, we annotate the evolution class of each nodule according to its changes in diameter. Based on the built NLSTt dataset, we propose a siamese encoder to simultaneously exploit the discriminative features of 3D ROIs detected from consecutive CT scans. Then we novelly design a spatial-temporal mixer (STM) to leverage the interval changes of the same nodule in sequential 3D ROIs and capture spatial dependencies of nodule regions and the current 3D ROI. According to the clinical diagnosis routine, we employ hierarchical loss to pay more attention to growing nodules. The extensive experiments on our organized dataset demonstrate the advantage of our proposed method. We also conduct experiments on an in-house dataset to evaluate the clinical utility of our method by comparing it against skilled clinicians.

Yang Yu, Qi Liu, Kai Zhang et al.

User modeling, which aims to capture users' characteristics or interests, heavily relies on task-specific labeled data and suffers from the data sparsity issue. Several recent studies tackled this problem by pre-training the user model on massive user behavior sequences with a contrastive learning task. Generally, these methods assume different views of the same behavior sequence constructed via data augmentation are semantically consistent, i.e., reflecting similar characteristics or interests of the user, and thus maximizing their agreement in the feature space. However, due to the diverse interests and heavy noise in user behaviors, existing augmentation methods tend to lose certain characteristics of the user or introduce noisy behaviors. Thus, forcing the user model to directly maximize the similarity between the augmented views may result in a negative transfer. To this end, we propose to replace the contrastive learning task with a new pretext task: Augmentation-Adaptive SelfSupervised Ranking (AdaptSSR), which alleviates the requirement of semantic consistency between the augmented views while pre-training a discriminative user model. Specifically, we adopt a multiple pairwise ranking loss which trains the user model to capture the similarity orders between the implicitly augmented view, the explicitly augmented view, and views from other users. We further employ an in-batch hard negative sampling strategy to facilitate model training. Moreover, considering the distinct impacts of data augmentation on different behavior sequences, we design an augmentation-adaptive fusion mechanism to automatically adjust the similarity order constraint applied to each sample based on the estimated similarity between the augmented views. Extensive experiments on both public and industrial datasets with six downstream tasks verify the effectiveness of AdaptSSR.

Liang Wang, Chao Song, Zhiyuan Liu et al.

Generative tasks about molecules, including but not limited to molecule generation, are crucial for drug discovery and material design, and have consistently attracted significant attention. In recent years, diffusion models have emerged as an impressive class of deep generative models, sparking extensive research and leading to numerous studies on their application to molecular generative tasks. Despite the proliferation of related work, there remains a notable lack of up-to-date and systematic surveys in this area. Particularly, due to the diversity of diffusion model formulations, molecular data modalities, and generative task types, the research landscape is challenging to navigate, hindering understanding and limiting the area's growth. To address this, this paper conducts a comprehensive survey of diffusion model-based molecular generative methods. We systematically review the research from the perspectives of methodological formulations, data modalities, and task types, offering a novel taxonomy. This survey aims to facilitate understanding and further flourishing development in this area. The relevant papers are summarized at: https://github.com/AzureLeon1/awesome-molecular-diffusion-models.

Xixi Wu, Qianguo Sun, Ruiyang Zhang et al.

Reinforcement Learning (RL) is essential for evolving Large Language Models (LLMs) into autonomous agents capable of long-horizon planning, yet a practical recipe for scaling RL in complex, multi-turn environments remains elusive. This paper presents a systematic empirical study using TravelPlanner, a challenging testbed requiring tool orchestration to satisfy multifaceted constraints. We decompose the agentic RL design space along 5 axes: reward shaping, model scaling, data composition, algorithm selection, and environmental stability. Our controlled experiments yield 7 key takeaways, e.g., (1) reward and algorithm choices are scale-dependent as smaller models benefit from staged rewards and enhanced exploration, whereas larger models converge efficiently with simpler dense rewards, (2) ~ 1K training samples with a balanced difficulty mixture mark a sweet spot for both in-domain and out-of-domain performance, and (3) environmental stability is critical to prevent policy degradation. Based on our distilled recipe, our RL-trained models achieve state-of-the-art performance on TravelPlanner, significantly outperforming leading LLMs.

Ruiyang Zhang, Qianguo Sun, Chao Song et al.

Large models are increasingly becoming autonomous agents that interact with real-world environments and use external tools to augment their static capabilities. However, most recent progress has focused on text-only large language models, which are limited to a single modality and therefore have narrower application scenarios. On the other hand, multimodal large models, while offering stronger perceptual capabilities, remain limited to static knowledge and lack the ability to access and leverage up-to-date web information. In this paper, we propose VSearcher, turning static multimodal model into multimodal search agent capable of long-horizon, multi-turn tool use in real-world web environments, including text search, image search, and web browsing, via reinforcement learning. Specifically, we introduce Iterative Injection Data Synthesis pipeline to generate large-scale, complex multimodal QA questions, which are further filtered with comprehensive metrics to ensure high quality and sufficient difficulty. We then adopt an SFT-then-RL training pipeline to turn base multimodal models to agent capable of multi-turn tool calling in real-world web environments. Besides, we propose a multimodal search benchmark MM-SearchExam dedicated to evaluating search capabilities of multimodal search agents, which proves highly challenging for recent proprietary models. Extensive evaluations across multiple multimodal search benchmarks reveal effectiveness of our method. VSearcher achieves superior performance compared to recent multimodal search agents and even surpasses several proprietary models on multimodal web search tasks.

Chao Song, Zhiyuan Liu, Han Huang et al.

Designing enzyme backbones with substrate-specific functionality is a critical challenge in computational protein engineering. Current generative models excel in protein design but face limitations in binding data, substrate-specific control, and flexibility for de novo enzyme backbone generation. To address this, we introduce EnzyBind, a dataset with 11,100 experimentally validated enzyme-substrate pairs specifically curated from PDBbind. Building on this, we propose EnzyControl, a method that enables functional and substrate-specific control in enzyme backbone generation. Our approach generates enzyme backbones conditioned on MSA-annotated catalytic sites and their corresponding substrates, which are automatically extracted from curated enzyme-substrate data. At the core of EnzyControl is EnzyAdapter, a lightweight, modular component integrated into a pretrained motif-scaffolding model, allowing it to become substrate-aware. A two-stage training paradigm further refines the model's ability to generate accurate and functional enzyme structures. Experiments show that our EnzyControl achieves the best performance across structural and functional metrics on EnzyBind and EnzyBench benchmarks, with particularly notable improvements of 13\% in designability and 13\% in catalytic efficiency compared to the baseline models. The code is released at https://github.com/Vecteur-libre/EnzyControl.

Rui Wang, Qianguo Sun, Chao Song et al.

DPO (Direct Preference Optimization) has become a widely used offline preference optimization algorithm due to its simplicity and training stability. However, DPO is prone to overfitting and collapse. To address these challenges, we propose Linear Preference Optimization (LPO), a novel alignment framework featuring three key innovations. First, we introduce gradient decoupling by replacing the log-sigmoid function with an absolute difference loss, thereby isolating the optimization dynamics. Second, we improve stability through an offset constraint combined with a positive regularization term to preserve the chosen response quality. Third, we implement controllable rejection suppression using gradient separation with straightforward estimation and a tunable coefficient that linearly regulates the descent of the rejection probability. Through extensive experiments, we demonstrate that LPO consistently improves performance on various tasks, including general text tasks, math tasks, and text-to-speech (TTS) tasks. These results establish LPO as a robust and tunable paradigm for preference alignment, and we release the source code, models, and training data publicly.

Chao Song, Zhihao Ye, Qiqiang Lin et al.

With the productive evolution of large language models (LLMs) in the field of natural language processing (NLP), tons of effort has been made to effectively fine-tune common pre-trained LLMs to fulfill a variety of tasks in one or multiple specific domain. In practice, there are two prevailing ways, in which the adaptation can be achieved: (i) Multiple Independent Models: Pre-trained LLMs are fine-tuned a few times independently using the corresponding training samples from each task. (ii) An Integrated Model: Samples from all tasks are employed to fine-tune a pre-trianed LLM unitedly. To address the high computing cost and seesawing issue simultaneously, we propose a unified framework that implements a 1 + N mutli-task fine-tuning pattern in LLMs using a novel Customized Gate Control (CGC) Low-rank Adaptation (LoRA) algorithm. Our work aims to take an advantage of both MTL (i.e., CGC) and PEFT (i.e., LoRA) scheme. For a given cluster of tasks, we design an innovative layer that contains two types of experts as additional trainable parameters to make LoRA be compatible with MTL. To comprehensively evaluate the proposed framework, we conduct well-designed experiments on two public datasets. The experimental results demonstrate that the unified framework with CGC-LoRA modules achieves higher evaluation scores than all benchmarks on both two datasets.

Yunhao Wang, Ziting Li, Shuai Chen et al.

Aligning large-scale vision-language models (VLMs) for complex reasoning via reinforcement learning is often hampered by the limitations of existing policy optimization algorithms, such as static training schedules and the rigid, uniform clipping mechanism in Proximal Policy Optimization (PPO). In this work, we introduce Adaptive Curriculum Policy Optimization (ACPO), a novel framework that addresses these challenges through a dual-component adaptive learning strategy. First, ACPO employs a dynamic curriculum that orchestrates a principled transition from a stable, near on-policy exploration phase to an efficient, off-policy exploitation phase by progressively increasing sample reuse. Second, we propose an Advantage-Aware Adaptive Clipping (AAAC) mechanism that replaces the fixed clipping hyperparameter with dynamic, sample-wise bounds modulated by the normalized advantage of each token. This allows for more granular and robust policy updates, enabling larger gradients for high-potential samples while safeguarding against destructive ones. We conduct extensive experiments on a suite of challenging multimodal reasoning benchmarks, including MathVista, LogicVista, and MMMU-Pro. Results demonstrate that ACPO consistently outperforms strong baselines such as DAPO and PAPO, achieving state-of-the-art performance, accelerated convergence, and superior training stability.

Lei Liu, Chao Song, Liangsheng He et al.

This study proposes a high-performance dual-parameter full waveform inversion framework (FWI) for ground-penetrating radar (GPR), accelerated through the hybrid compilation of CUDA kernel functions and PyTorch. The method leverages the computational efficiency of GPU programming while preserving the flexibility and usability of Python-based deep learning frameworks. By integrating customized CUDA kernels into PyTorch's automatic differentiation mechanism, the framework enables accurate and efficient inversion of both dielectric permittivity and electrical conductivity. Experimental evaluations on synthetic data and real wavefield data demonstrate that the proposed method achieves dual-parameter FWI for GPR data while maintaining high accuracy. Moreover, the framework is flexible and extensible, supporting optional regularization strategies such as total variation and multi-scale inversion. These features make the proposed approach a practical and scalable framework for rapid GPR-based subsurface imaging in applications including civil engineering, environmental monitoring, and geophysical exploration.

Chao Song, Tariq Alkhalifah, Umair Bin Waheed et al.

Full-waveform inversion (FWI) is an advanced technique for reconstructing high-resolution subsurface physical parameters by progressively minimizing the discrepancy between observed and predicted seismic data. However, conventional FWI encounters challenges in real data applications, primarily due to its conventional objective of direct measurements of the data misfit. Accurate estimation of the source wavelet is essential for effective data fitting, alongside the need for low-frequency data and a reasonable initial model to prevent cycle skipping. Additionally, wave equation solvers often struggle to accurately simulate the amplitude of observed data in real applications. To address these challenges, we introduce a correlation-based source-independent objective function for FWI that aims to mitigate source uncertainty and amplitude dependency, which effectively enhances its practicality for real data applications. We develop a deep-learning framework constrained by this new objective function with a velocity-distribution supported deep image prior, which reparameterizes velocity inversion into trainable parameters within an autoencoder, thereby reducing the nonlinearity in the conventional FWI's objective function. We demonstrate the superiority of our proposed method using synthetic data from benchmark velocity models and, more importantly, two real datasets. These examples highlight its effectiveness and practicality even under challenging conditions, such as missing low frequencies, a crude initial velocity model, and an incorrect source wavelet.

Ke Wang, Weikang Li, Shibo Xu et al.

Quantum nonlocality describes a stronger form of quantum correlation than that of entanglement. It refutes Einstein's belief of local realism and is among the most distinctive and enigmatic features of quantum mechanics. It is a crucial resource for achieving quantum advantages in a variety of practical applications, ranging from cryptography and certified random number generation via self-testing to machine learning. Nevertheless, the detection of nonlocality, especially in quantum many-body systems, is notoriously challenging. Here, we report an experimental certification of genuine multipartite Bell correlations, which signal nonlocality in quantum many-body systems, up to 24 qubits with a fully programmable superconducting quantum processor. In particular, we employ energy as a Bell correlation witness and variationally decrease the energy of a many-body system across a hierarchy of thresholds, below which an increasing Bell correlation depth can be certified from experimental data. As an illustrating example, we variationally prepare the low-energy state of a two-dimensional honeycomb model with 73 qubits and certify its Bell correlations by measuring an energy that surpasses the corresponding classical bound with up to 48 standard deviations. In addition, we variationally prepare a sequence of low-energy states and certify their genuine multipartite Bell correlations up to 24 qubits via energies measured efficiently by parity oscillation and multiple quantum coherence techniques. Our results establish a viable approach for preparing and certifying multipartite Bell correlations, which provide not only a finer benchmark beyond entanglement for quantum devices, but also a valuable guide towards exploiting multipartite Bell correlation in a wide spectrum of practical applications.

Weicong Ding, Hanlin Tang, Jingshuo Feng et al.

Real-world recommendation systems often consist of two phases. In the first phase, multiple predictive models produce the probability of different immediate user actions. In the second phase, these predictions are aggregated according to a set of 'strategic parameters' to meet a diverse set of business goals, such as longer user engagement, higher revenue potential, or more community/network interactions. In addition to building accurate predictive models, it is also crucial to optimize this set of 'strategic parameters' so that primary goals are optimized while secondary guardrails are not hurt. In this setting with multiple and constrained goals, this paper discovers that a probabilistic strategic parameter regime can achieve better value compared to the standard regime of finding a single deterministic parameter. The new probabilistic regime is to learn the best distribution over strategic parameter choices and sample one strategic parameter from the distribution when each user visits the platform. To pursue the optimal probabilistic solution, we formulate the problem into a stochastic compositional optimization problem, in which the unbiased stochastic gradient is unavailable. Our approach is applied in a popular social network platform with hundreds of millions of daily users and achieves +0.22% lift of user engagement in a recommendation task and +1.7% lift in revenue in an advertising optimization scenario comparing to using the best deterministic parameter strategy.