Xingyu Wang

Kaiyu Huang, Xingyu Wang, Mingze Kong et al.

In real-world deployments of large language models (LLMs), balancing inference quality and computational cost has become a central challenge. Existing approaches tackle this trade-off along two largely independent dimensions: model routing, which switches among models of different scales to match request complexity, and test-time scaling (TTS), which adjusts inference-time compute within a fixed model for fine-grained control. However, this decoupled design introduces inherent limitations. Model routing yields coarse-grained, discrete performance changes due to the sparse set of model scales, while single-model TTS often encounters capacity ceilings and exhibits diminishing returns as compute increases. Moreover, treating the two mechanisms separately restricts adaptability in dynamic inference environments. To overcome these limitations, we introduce Unified Inference Scaling (UIS), which unifies model routing and TTS in a single optimization space. Building on this formulation, we propose UniScale, an online framework that models adaptive UIS as a contextual multi-armed bandit problem and learns inference policies via LinUCB. The framework incorporates efficiency-aware learning and cost modeling to ensure stable and scalable optimization over high-dimensional action spaces. Evaluation shows that UniScale effectively exploits the synergy in the UIS space to deliver a fine-grained and consistently better quality-cost trade-off across diverse, dynamic inference scenarios.

Xingyu Wang, Roman Samulyak, Xiangmin Jiao et al.

We propose a new adaptive Particle-in-Cloud (AP-Cloud) method for obtaining optimal numerical solutions to the Vlasov-Poisson equation. Unlike the traditional particle-in-cell (PIC) method, which is commonly used for solving this problem, the AP-Cloud adaptively selects computational nodes or particles to deliver higher accuracy and efficiency when the particle distribution is highly non-uniform. Unlike other adaptive techniques for PIC, our method balances the errors in PDE discretization and Monte Carlo integration, and discretizes the differential operators using a generalized finite difference (GFD) method based on a weighted least square formulation. As a result, AP-Cloud is independent of the geometric shapes of computational domains and is free of artificial parameters. Efficient and robust implementation is achieved through an octree data structure with 2:1 balance. We analyze the accuracy and convergence order of AP-Cloud theoretically, and verify the method using an electrostatic problem of a particle beam with halo. Simulation results show that the AP-Cloud method is substantially more accurate and faster than the traditional PIC, and it is free of artificial forces that are typical for some adaptive PIC techniques.

Tongxuan Liu, Xingyu Wang, Weizhe Huang et al.

In recent years, Large Language Models (LLMs) have demonstrated remarkable capabilities across diverse NLP tasks. Extensive research has explored how to enhance the logical reasoning abilities such as Chain-of-Thought, Chain-of-Thought with Self-Consistency, Tree-Of-Thoughts, and multi-agent debates. In the context of multi-agent debates, significant performance improvements can be achieved with an increasing number of agents and debate rounds. However, the escalation in the number of agents and debate rounds can drastically raise the tokens cost of debates, thereby limiting the scalability of the multi-agent debate technique. To better harness the advantages of multi-agent debates in logical reasoning tasks, this paper proposes a method to significantly reduce token cost in multi-agent debates. This approach involves dividing all agents into multiple debate groups, with agents engaging in debates within their respective groups and sharing interim debate results between groups. Comparative experiments across multiple datasets have demonstrated that this method can reduce the total tokens by up to 51.7% during debates and while potentially enhancing accuracy by as much as 25%. Our method significantly enhances the performance and efficiency of interactions in the multi-agent debate.

Tongxuan Liu, Wenjiang Xu, Weizhe Huang et al.

Large Language Models (LLMs) have demonstrated remarkable capabilities across various tasks but their performance in complex logical reasoning tasks remains unsatisfactory. Although some prompting methods, such as Chain-of-Thought, can improve the reasoning ability of LLMs to some extent, they suffer from an unfaithful issue where derived conclusions may not align with the generated reasoning chain. To address this issue, some studies employ the approach of propositional logic to further enhance logical reasoning abilities of LLMs. However, the potential omissions in the extraction of logical expressions in these methods can cause information loss in the logical reasoning process, thereby generating incorrect results. To this end, we propose Logic-of-Thought (LoT) prompting which employs propositional logic to generate expanded logical information descriptions and utilizes them as an additional augmentation to original contexts, thereby ensuring information completeness and enhancing logical reasoning ability. LoT is orthogonal to existing prompting methods and can be seamlessly integrated with them. Extensive experiments demonstrate that LoT boosts the performance of various prompting methods with a striking margin across five logical reasoning tasks. In particular, LoT enhances Chain-of-Thought's performance on the ReClor dataset by +4.35%, improves Chain-of-Thought with Self-Consistency's performance on the RuleTaker dataset by +3.52%, and boosts performance of Tree-of-Thoughts on the ProofWriter dataset by +8%.

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Qiqing Huang, Xingyu Wang, Wanda Guo et al.

Modern 5G user equipment (UE) processes Radio Resource Control (RRC) configuration messages during early control-plane exchanges, before authentication and integrity protection are established. Prior work for testing 5G UEs has largely focused on constructing syntactically invalid inputs. In contrast, we show that syntactically valid but semantically inconsistent messages, which violate specification-level field constraints or cross-field dependencies, can drive baseband implementations into invalid states, triggering assertion failures or modem crashes. These findings reveal semantic inconsistencies in pre-authentication signaling as a critical yet underexplored attack surface in 5G UE implementations. To address this gap, we present Constraint-Guided Semantic Testing (ConSeT), a framework that systematically extracts specification-level constraints and leverages them to generate targeted semantic violations for testing 5G UEs. ConSeT decodes RRC messages into structured fields, derives schema-based rules, infers cross-field dependencies using a Large Language Model (LLM) in an evidence-bounded manner, and produces syntactically valid test cases that intentionally violate semantic constraints. We evaluate ConSeT on both commercial and open-source 5G UEs. On commercial smartphones, it uncovers 7 previously unknown vulnerabilities through responsible disclosure, including 3 high-severity CVEs, affecting 64 chipset models and over 542 commercially available smartphone models. On the open-source OAI UE, ConSeT additionally triggers 29 distinct crash sites.

Xingyu Wang, Pengxiang Ding, Jingkai Xu et al.

Recent advances in visuomotor policy learning have enabled robots to perform control directly from visual inputs. Yet, extending such end-to-end learning from single-arm to bimanual manipulation remains challenging due to the need for both independent perception and coordinated interaction between arms. Existing methods typically favor one side -- either decoupling the two arms to avoid interference or enforcing strong cross-arm coupling for coordination -- thus lacking a unified treatment. We propose CUBic, a Coordinated and Unified framework for Bimanual perception and control that reformulates bimanual coordination as a unified perceptual modeling problem. CUBic learns a shared tokenized representation bridging perception and control, where independence and coordination emerge intrinsically from structure rather than from hand-crafted coupling. Our approach integrates three components: unidirectional perception aggregation, bidirectional perception coordination through two codebooks with shared mapping, and a unified perception-to-control diffusion policy. Extensive experiments on the RoboTwin benchmark show that CUBic consistently surpasses standard baselines, achieving marked improvements in coordination accuracy and task success rates over state-of-the-art visuomotor baselines.

Tongtong Wu, Weigang Wu, Xingyu Wang et al.

Large Language Models (LLMs) have made tremendous strides in code generation, but existing research fails to account for the dynamic nature of software development, marked by frequent library updates. This gap significantly limits LLMs' deployment in realistic settings. In this paper, we propose two novel tasks aimed at bridging this gap: version-specific code completion (VSCC) and version-aware code migration (VACM). In conjunction, we introduce VersiCode, a comprehensive Python dataset specifically designed to evaluate LLMs on these two tasks, together with a novel evaluation metric, Critical Diff Check (CDC@1), which assesses code generation against evolving API requirements. We conduct an extensive evaluation on VersiCode, which reveals that version-controllable code generation is indeed a significant challenge, even for GPT-4o and other strong frontier models. We believe the novel tasks, dataset, and metric open up a new, important research direction that will further enhance LLMs' real-world applicability. The code and resources can be found at https://github.com/wutong8023/VersiCode.

Kun Wu, Chengkai Hou, Jiaming Liu et al.

In this paper, we introduce RoboMIND (Multi-embodiment Intelligence Normative Data for Robot Manipulation), a dataset containing 107k demonstration trajectories across 479 diverse tasks involving 96 object classes. RoboMIND is collected through human teleoperation and encompasses comprehensive robotic-related information, including multi-view observations, proprioceptive robot state information, and linguistic task descriptions. To ensure data consistency and reliability for imitation learning, RoboMIND is built on a unified data collection platform and a standardized protocol, covering four distinct robotic embodiments: the Franka Emika Panda, the UR5e, the AgileX dual-arm robot, and a humanoid robot with dual dexterous hands. Our dataset also includes 5k real-world failure demonstrations, each accompanied by detailed causes, enabling failure reflection and correction during policy learning. Additionally, we created a digital twin environment in the Isaac Sim simulator, replicating the real-world tasks and assets, which facilitates the low-cost collection of additional training data and enables efficient evaluation. To demonstrate the quality and diversity of our dataset, we conducted extensive experiments using various imitation learning methods for single-task settings and state-of-the-art Vision-Language-Action (VLA) models for multi-task scenarios. By leveraging RoboMIND, the VLA models achieved high manipulation success rates and demonstrated strong generalization capabilities. To the best of our knowledge, RoboMIND is the largest multi-embodiment teleoperation dataset collected on a unified platform, providing large-scale and high-quality robotic training data. Our project is at https://x-humanoid-robomind.github.io/.

Xingyu Wang, Tao Wang

Test-Time Adaptation (TTA) is essential for enabling deep learning models to handle real-world data distribution shifts. However, current approaches face significant limitations: backpropagation-based methods are not suitable for low-end deployment devices, due to their high computation and memory requirements, as well as their tendency to modify model weights during adaptation; while traditional backpropagation-free techniques exhibit constrained adaptation capabilities. In this work, we propose Forward-Only Zeroth-Order Optimization (FOZO), a novel and practical backpropagation-free paradigm for TTA. FOZO leverages a memory-efficient zeroth-order prompt optimization, which is led by objectives optimizing both intermediate feature statistics and prediction entropy. To ensure efficient and stable adaptation over the out-of-distribution data stream, we introduce a dynamically decaying perturbation scale during zeroth-order gradient estimation and theoretically prove its convergence under the TTA data stream assumption. Extensive continual adaptation experiments on ImageNet-C, ImageNet-R, and ImageNet-Sketch demonstrate FOZO's superior performance, achieving 59.52% Top-1 accuracy on ImageNet-C (5K, level 5) and outperforming main gradient-based methods and SOTA forward-only FOA (58.13%). Furthermore, FOZO exhibits strong generalization on quantized (INT8) models. These findings demonstrate that FOZO is a highly competitive solution for TTA deployment in resource-limited scenarios.

Yijun Shen, Delong Chen, Fan Liu et al.

While densely annotated image captions significantly facilitate the learning of robust vision-language alignment, methodologies for systematically optimizing human annotation efforts remain underexplored. We introduce Chain-of-Talkers (CoTalk), an AI-in-the-loop methodology designed to maximize the number of annotated samples and improve their comprehensiveness under fixed budget constraints (e.g., total human annotation time). The framework is built upon two key insights. First, sequential annotation reduces redundant workload compared to conventional parallel annotation, as subsequent annotators only need to annotate the ``residual'' -- the missing visual information that previous annotations have not covered. Second, humans process textual input faster by reading while outputting annotations with much higher throughput via talking; thus a multimodal interface enables optimized efficiency. We evaluate our framework from two aspects: intrinsic evaluations that assess the comprehensiveness of semantic units, obtained by parsing detailed captions into object-attribute trees and analyzing their effective connections; extrinsic evaluation measures the practical usage of the annotated captions in facilitating vision-language alignment. Experiments with eight participants show our Chain-of-Talkers (CoTalk) improves annotation speed (0.42 vs. 0.30 units/sec) and retrieval performance (41.13% vs. 40.52%) over the parallel method.

Xingyu Wang, Chang-Han Rhee

Stochastic gradient descent (SGD) and its variants enable modern artificial intelligence. However, theoretical understanding lags far behind their empirical success. It is widely believed that SGD has a curious ability to avoid sharp local minima in the loss landscape, which are associated with poor generalization. To unravel this mystery and further enhance such capability of SGDs, it is imperative to go beyond the traditional local convergence analysis and obtain a comprehensive understanding of SGDs' global dynamics. In this paper, we develop a set of technical machinery based on the recent large deviations and metastability analysis in Wang and Rhee (2023) and obtain sharp characterization of the global dynamics of heavy-tailed SGDs. In particular, we reveal a fascinating phenomenon in deep learning: by injecting and then truncating heavy-tailed noises during the training phase, SGD can almost completely avoid sharp minima and achieve better generalization performance for the test data. Simulation and deep learning experiments confirm our theoretical prediction that heavy-tailed SGD with gradient clipping finds local minima with a more flat geometry and achieves better generalization performance.

Ming Meng, Qi Dong, Jiajie Li et al.

Virtual try-on technology has become increasingly important in the fashion and retail industries, enabling the generation of high-fidelity garment images that adapt seamlessly to target human models. While existing methods have achieved notable progress, they still face significant challenges in maintaining consistency across different poses. Specifically, geometric distortions lead to a lack of spatial consistency, mismatches in garment structure and texture across poses result in semantic inconsistency, and the loss or distortion of fine-grained details diminishes visual fidelity. To address these challenges, we propose HF-VTON, a novel framework that ensures high-fidelity virtual try-on performance across diverse poses. HF-VTON consists of three key modules: (1) the Appearance-Preserving Warp Alignment Module (APWAM), which aligns garments to human poses, addressing geometric deformations and ensuring spatial consistency; (2) the Semantic Representation and Comprehension Module (SRCM), which captures fine-grained garment attributes and multi-pose data to enhance semantic representation, maintaining structural, textural, and pattern consistency; and (3) the Multimodal Prior-Guided Appearance Generation Module (MPAGM), which integrates multimodal features and prior knowledge from pre-trained models to optimize appearance generation, ensuring both semantic and geometric consistency. Additionally, to overcome data limitations in existing benchmarks, we introduce the SAMP-VTONS dataset, featuring multi-pose pairs and rich textual annotations for a more comprehensive evaluation. Experimental results demonstrate that HF-VTON outperforms state-of-the-art methods on both VITON-HD and SAMP-VTONS, excelling in visual fidelity, semantic consistency, and detail preservation.

Xingyu Wang, Mengfan Xu

We study decentralized multi-agent multi-armed bandits in fully heavy-tailed settings, where clients communicate over sparse random graphs with heavy-tailed degree distributions and observe heavy-tailed (homogeneous or heterogeneous) reward distributions with potentially infinite variance. The objective is to maximize system performance by pulling the globally optimal arm with the highest global reward mean across all clients. We are the first to address such fully heavy-tailed scenarios, which capture the dynamics and challenges in communication and inference among multiple clients in real-world systems. In homogeneous settings, our algorithmic framework exploits hub-like structures unique to heavy-tailed graphs, allowing clients to aggregate rewards and reduce noises via hub estimators when constructing UCB indices; under $M$ clients and degree distributions with power-law index $α> 1$, our algorithm attains a regret bound (almost) of order $O(M^{1 -\frac{1}α} \log{T})$. Under heterogeneous rewards, clients synchronize by communicating with neighbors, aggregating exchanged estimators in UCB indices; With our newly established information delay bounds on sparse random graphs, we prove a regret bound of $O(M \log{T})$. Our results improve upon existing work, which only address time-invariant connected graphs, or light-tailed dynamics in dense graphs and rewards.

Gemini Team, Rohan Anil, Sebastian Borgeaud et al.

This report introduces a new family of multimodal models, Gemini, that exhibit remarkable capabilities across image, audio, video, and text understanding. The Gemini family consists of Ultra, Pro, and Nano sizes, suitable for applications ranging from complex reasoning tasks to on-device memory-constrained use-cases. Evaluation on a broad range of benchmarks shows that our most-capable Gemini Ultra model advances the state of the art in 30 of 32 of these benchmarks - notably being the first model to achieve human-expert performance on the well-studied exam benchmark MMLU, and improving the state of the art in every one of the 20 multimodal benchmarks we examined. We believe that the new capabilities of the Gemini family in cross-modal reasoning and language understanding will enable a wide variety of use cases. We discuss our approach toward post-training and deploying Gemini models responsibly to users through services including Gemini, Gemini Advanced, Google AI Studio, and Cloud Vertex AI.

Xingyu Wang, Sewoong Oh, Chang-Han Rhee

The empirical success of deep learning is often attributed to SGD's mysterious ability to avoid sharp local minima in the loss landscape, as sharp minima are known to lead to poor generalization. Recently, empirical evidence of heavy-tailed gradient noise was reported in many deep learning tasks, and it was shown in Şimşekli (2019a,b) that SGD can escape sharp local minima under the presence of such heavy-tailed gradient noise, providing a partial solution to the mystery. In this work, we analyze a popular variant of SGD where gradients are truncated above a fixed threshold. We show that it achieves a stronger notion of avoiding sharp minima: it can effectively eliminate sharp local minima entirely from its training trajectory. We characterize the dynamics of truncated SGD driven by heavy-tailed noises. First, we show that the truncation threshold and width of the attraction field dictate the order of the first exit time from the associated local minimum. Moreover, when the objective function satisfies appropriate structural conditions, we prove that as the learning rate decreases, the dynamics of heavy-tailed truncated SGD closely resemble those of a continuous-time Markov chain that never visits any sharp minima. Real data experiments on deep learning confirm our theoretical prediction that heavy-tailed SGD with gradient clipping finds a "flatter" local minima and achieves better generalization.

Xingyu Wang, Lida Zhang, Diego Klabjan

Certain type of documents such as tweets are collected by specifying a set of keywords. As topics of interest change with time it is beneficial to adjust keywords dynamically. The challenge is that these need to be specified ahead of knowing the forthcoming documents and the underlying topics. The future topics should mimic past topics of interest yet there should be some novelty in them. We develop a keyword-based topic model that dynamically selects a subset of keywords to be used to collect future documents. The generative process first selects keywords and then the underlying documents based on the specified keywords. The model is trained by using a variational lower bound and stochastic gradient optimization. The inference consists of finding a subset of keywords where given a subset the model predicts the underlying topic-word matrix for the unknown forthcoming documents. We compare the keyword topic model against a benchmark model using viral predictions of tweets combined with a topic model. The keyword-based topic model outperforms this sophisticated baseline model by 67%.

Zhaoyang Qiu, Shugeng Chen, Ian Daly et al.

Stroke is the leading cause of serious and long-term disability worldwide. Some studies have shown that motor imagery (MI) based BCI has a positive effect in poststroke rehabilitation. It could help patients promote the reorganization processes in the damaged brain regions. However, offline motor imagery and conventional online motor imagery with feedback (such as rewarding sounds and movements of an avatar) could not reflect the true intention of the patients. In this study, both virtual limbs and functional electrical stimulation (FES) were used as feedback to provide patients a closed-loop sensorimotor integration for motor rehabilitation. The FES system would activate if the user was imagining hand movement of instructed side. Ten stroke patients (7 male, aged 22-70 years, mean 49.5+-15.1) were involved in this study. All of them participated in BCI-FES rehabilitation training for 4 weeks.The average motor imagery accuracies of the ten patients in the last week were 71.3%, which has improved 3% than that in the first week. Five patients' Fugl-Meyer Assessment (FMA) scores have been raised. Patient 6, who has have suffered from stroke over two years, achieved the greatest improvement after rehabilitation training (pre FMA: 20, post FMA: 35). In the aspect of brain patterns, the active patterns of the five patients gradually became centralized and shifted to sensorimotor areas (channel C3 and C4) and premotor area (channel FC3 and FC4).In this study, motor imagery based BCI and FES system were combined to provided stoke patients with a closed-loop sensorimotor integration for motor rehabilitation. Result showed evidences that the BCI-FES system is effective in restoring upper extremities motor function in stroke. In future work, more cases are needed to demonstrate its superiority over conventional therapy and explore the potential role of MI in poststroke rehabilitation.

Xingyu Wang, Diego Klabjan

This paper considers the problem of inverse reinforcement learning in zero-sum stochastic games when expert demonstrations are known to be not optimal. Compared to previous works that decouple agents in the game by assuming optimality in expert strategies, we introduce a new objective function that directly pits experts against Nash Equilibrium strategies, and we design an algorithm to solve for the reward function in the context of inverse reinforcement learning with deep neural networks as model approximations. In our setting the model and algorithm do not decouple by agent. In order to find Nash Equilibrium in large-scale games, we also propose an adversarial training algorithm for zero-sum stochastic games, and show the theoretical appeal of non-existence of local optima in its objective function. In our numerical experiments, we demonstrate that our Nash Equilibrium and inverse reinforcement learning algorithms address games that are not amenable to previous approaches using tabular representations. Moreover, with sub-optimal expert demonstrations our algorithms recover both reward functions and strategies with good quality.

Jing Jin, Brendan Z. Allison, Yu Zhang et al.

Objective: Amyotrophic lateral sclerosis (ALS) is a rare disease, but is also one of the most common motor neuron diseases, and people of all races and ethnic backgrounds are affected. There is currently no cure. Brain computer interfaces (BCIs) can establish a communication channel directly between the brain and an external device by recognizing brain activities that reflect user intent. Therefore, this technology could help ALS patients in promoting functional independence through BCI-based speller systems and motor assistive devices. Methods: In this paper, two kinds of ERP-based speller systems were tested on 18 ALS patients to: (1) assess performance when they spelled 42 characters online continuously, without a break; and (2) to compare performance between a matrix-based speller paradigm (MS-P, mean visual angle 6 degree) and a new speller paradigm that used a larger visual angle called the large visual angle speller paradigm (LS-P, mean visual angle 8 degree). Results: Although results showed that there were no significant differences between the two paradigms in accuracy trend over continuous use (p>0.05), the fatigue during the LS-P condition was significantly lower than that of MS-P (p<0.05). Results also showed that continuous use slightly reduced the performance of this ERP-based BCI. Conclusion: 15 subjects obtained higher than 80% feedback accuracy (online output accuracy) and 9 subjects obtained higher than 90% feedback accuracy in one of the two paradigms, thus validating the BCI approaches in this study. Significance: Most ALS subjects in this study could spell effectively after continuous use of an ERP-based BCI. The new LS-P display may be easier for subjects to use, resulting in lower fatigue.

Zhaoyang Qiu, Brendan Z. Allison, Jing Jin et al.

Motor imagery (MI) is a mental representation of motor behavior that has been widely used as a control method for a brain-computer interface (BCI), allowing communication for the physically impaired. The performance of MI based BCI mainly depends on the subject's ability to self-modulate EEG signals. Proper training can help naive subjects learn to modulate brain activity proficiently. However, training subjects typically involves abstract motor tasks and is time-consuming. To improve the performance of naive subjects during motor imagery, a novel paradigm was presented that would guide naive subjects to modulate brain activity effectively. In this new paradigm, pictures of the left or right hand were used as cues for subjects to finish the motor imagery task. Fourteen healthy subjects (11 male, aged 22-25 years, mean 23.6+/-1.16) participated in this study. The task was to imagine writing a Chinese character. Specifically, subjects could imagine hand movements following the sequence of writing strokes in the Chinese character. This paradigm was meant to find an effective and familiar action for most Chinese people, to provide them with a specific, extensively practiced task and help them modulate brain activity. Results showed that the writing task paradigm yielded significantly better performance than the traditional arrow paradigm (p<0.001). Questionnaire replies indicated that most subjects thought the new paradigm was easier and more comfortable. The proposed new motor imagery paradigm could guide subjects to help them modulate brain activity effectively. Results showed that there were significant improvements using new paradigm, both in classification accuracy and usability.

Yu Zhang, Guoxu Zhou, Jing Jin et al.

Canonical correlation analysis (CCA) has been one of the most popular methods for frequency recognition in steady-state visual evoked potential (SSVEP)-based brain-computer interfaces (BCIs). Despite its efficiency, a potential problem is that using pre-constructed sine-cosine waves as the required reference signals in the CCA method often does not result in the optimal recognition accuracy due to their lack of features from the real EEG data. To address this problem, this study proposes a novel method based on multiset canonical correlation analysis (MsetCCA) to optimize the reference signals used in the CCA method for SSVEP frequency recognition. The MsetCCA method learns multiple linear transforms that implement joint spatial filtering to maximize the overall correlation among canonical variates, and hence extracts SSVEP common features from multiple sets of EEG data recorded at the same stimulus frequency. The optimized reference signals are formed by combination of the common features and completely based on training data. Experimental study with EEG data from ten healthy subjects demonstrates that the MsetCCA method improves the recognition accuracy of SSVEP frequency in comparison with the CCA method and other two competing methods (multiway CCA (MwayCCA) and phase constrained CCA (PCCA)), especially for a small number of channels and a short time window length. The superiority indicates that the proposed MsetCCA method is a new promising candidate for frequency recognition in SSVEP-based BCIs.