Hanyang Wang

Peng Xia, Jianwen Chen, Hanyang Wang et al.

Large Language Model (LLM) agents have shown stunning results in complex tasks, yet they often operate in isolation, failing to learn from past experiences. Existing memory-based methods primarily store raw trajectories, which are often redundant and noise-heavy. This prevents agents from extracting high-level, reusable behavioral patterns that are essential for generalization. In this paper, we propose SkillRL, a framework that bridges the gap between raw experience and policy improvement through automatic skill discovery and recursive evolution. Our approach introduces an experience-based distillation mechanism to build a hierarchical skill library SkillBank, an adaptive retrieval strategy for general and task-specific heuristics, and a recursive evolution mechanism that allows the skill library to co-evolve with the agent's policy during reinforcement learning. These innovations significantly reduce the token footprint while enhancing reasoning utility. Experimental results on ALFWorld, WebShop and seven search-augmented tasks demonstrate that SkillRL achieves state-of-the-art performance, outperforming strong baselines over 15.3% and maintaining robustness as task complexity increases. Code is available at this https://github.com/aiming-lab/SkillRL.

Shengjun Zhang, Zhang Zhang, Simin Huang et al.

Recent advancements in video-based world models have demonstrated an unprecedented ability to synthesize high-fidelity visual sequences. However, a fundamental gap persists between visually plausible video generation and the functional requirements of a world model, particularly in maintaining a stable and reasonable internal state over extended temporal horizons. While existing benchmarks primarily emphasize visual quality, motion coherence, and text-video alignment, they largely overlook memory, the core capability of a world model to preserve consistency across long-term horizons and complex interactions. To address this gap, we present \textbf{MBench}, a comprehensive benchmark dedicated to quantifying and evaluating the memory capability of video world models. We systematically decompose the memory capability of video world models into three hierarchical and complementary core dimensions: entity consistency, environment consistency, and causal consistency, which are further refined into 12 quantifiable sub-dimensions for comprehensive characterization of long-term memory. Our benchmark is built upon rigorously curated real-captured long videos, and evaluated by rule-based quantitative matrices and VLM to enable objective and comprehensive consistency assessment. Extensive evaluations of mainstream state-of-the-art video world models reveal critical systemic limitations of existing methods in long-term state retention, providing a standardized benchmark and clear research direction to advance the field.

Taiyi Su, Jian Zhu, Tianjian Wang et al.

Real-world household robots require Vision-Language-Action (VLA) foundation models that can acquire reusable manipulation skills across diverse objects, task conditions, and household environments. Deformable-object folding is a representative challenge, requiring robots to handle clothing items from random initial states across varying categories, geometries, materials, and scenes. However, existing VLA systems commonly train separate policies for different object categories, while naively mixed multi-task training often suffers from task interference and degraded performance. To move beyond category-specific folding policies, we introduce DeMaVLA, a VLA foundation model for generalizable Deformable Manipulation. DeMaVLA adopts a VLM backbone with an action expert and formulates continuous action generation using flow matching. To improve efficiency, the action expert is constructed by pruning every other transformer layer while preserving layer-wise alignment with the VLM backbone, reducing training and inference cost. DeMaVLA is first pre-trained on approximately 5,000 hours of selected real-world dual-arm demonstrations to acquire general manipulation priors. It is then post-trained on mixed folding data that aggregates self-collected demonstrations and corrective trajectories from real-robot failures across multiple folding tasks through a human-in-the-loop Data Aggregation~(DAgger) pipeline. Experiments show that DeMaVLA achieves competitive performance on RoboTwin and strong real-world results on our household folding benchmark. These results highlight the value of scalable real-world data, efficient action generation, and corrective learning for general-purpose VLA policies in deformable-object manipulation.

Fangfu Liu, Wenqiang Sun, Hanyang Wang et al.

Advancements in 3D scene reconstruction have transformed 2D images from the real world into 3D models, producing realistic 3D results from hundreds of input photos. Despite great success in dense-view reconstruction scenarios, rendering a detailed scene from insufficient captured views is still an ill-posed optimization problem, often resulting in artifacts and distortions in unseen areas. In this paper, we propose ReconX, a novel 3D scene reconstruction paradigm that reframes the ambiguous reconstruction challenge as a temporal generation task. The key insight is to unleash the strong generative prior of large pre-trained video diffusion models for sparse-view reconstruction. However, 3D view consistency struggles to be accurately preserved in directly generated video frames from pre-trained models. To address this, given limited input views, the proposed ReconX first constructs a global point cloud and encodes it into a contextual space as the 3D structure condition. Guided by the condition, the video diffusion model then synthesizes video frames that are both detail-preserved and exhibit a high degree of 3D consistency, ensuring the coherence of the scene from various perspectives. Finally, we recover the 3D scene from the generated video through a confidence-aware 3D Gaussian Splatting optimization scheme. Extensive experiments on various real-world datasets show the superiority of our ReconX over state-of-the-art methods in terms of quality and generalizability.

Hanyang Wang, Hao Zhou, Sibo Cheng

Despite the success of various methods in addressing the issue of spatial reconstruction of dynamical systems with sparse observations, spatio-temporal prediction for sparse fields remains a challenge. Existing Kriging-based frameworks for spatio-temporal sparse field prediction fail to meet the accuracy and inference time required for nonlinear dynamic prediction problems. In this paper, we introduce the Dynamical System Prediction from Sparse Observations using Voronoi Tessellation (DSOVT) framework, an innovative methodology based on Voronoi tessellation which combines convolutional encoder-decoder (CED) and long short-term memory (LSTM) and utilizing Convolutional Long Short-Term Memory (ConvLSTM). By integrating Voronoi tessellations with spatio-temporal deep learning models, DSOVT is adept at predicting dynamical systems with unstructured, sparse, and time-varying observations. CED-LSTM maps Voronoi tessellations into a low-dimensional representation for time series prediction, while ConvLSTM directly uses these tessellations in an end-to-end predictive model. Furthermore, we incorporate physics constraints during the training process for dynamical systems with explicit formulas. Compared to purely data-driven models, our physics-based approach enables the model to learn physical laws within explicitly formulated dynamics, thereby enhancing the robustness and accuracy of rolling forecasts. Numerical experiments on real sea surface data and shallow water systems clearly demonstrate our framework's accuracy and computational efficiency with sparse and time-varying observations.

Hanyang Wang, Yiyang Liu, Jiawei Chi et al.

Classifier-Free Guidance (CFG) has emerged as a central approach for enhancing semantic alignment in flow-based diffusion models. In this paper, we explore a unified framework called CFG-Ctrl, which reinterprets CFG as a control applied to the first-order continuous-time generative flow, using the conditional-unconditional discrepancy as an error signal to adjust the velocity field. From this perspective, we summarize vanilla CFG as a proportional controller (P-control) with fixed gain, and typical follow-up variants develop extended control-law designs derived from it. However, existing methods mainly rely on linear control, inherently leading to instability, overshooting, and degraded semantic fidelity especially on large guidance scales. To address this, we introduce Sliding Mode Control CFG (SMC-CFG), which enforces the generative flow toward a rapidly convergent sliding manifold. Specifically, we define an exponential sliding mode surface over the semantic prediction error and introduce a switching control term to establish nonlinear feedback-guided correction. Moreover, we provide a Lyapunov stability analysis to theoretically support finite-time convergence. Experiments across text-to-image generation models including Stable Diffusion 3.5, Flux, and Qwen-Image demonstrate that SMC-CFG outperforms standard CFG in semantic alignment and enhances robustness across a wide range of guidance scales. Project Page: https://hanyang-21.github.io/CFG-Ctrl

Hanyang Wang, Mingxuan Zhu

Modern reasoning models continue generating long after the answer is already determined. Across five model configurations, two families, and three benchmarks, we find that \textbf{52--88\% of chain-of-thought tokens are produced after the answer is recoverable} from a partial prefix. This post-commitment generation reveals a structural phenomenon: the \textbf{detection--extraction gap}. Free continuations from early prefixes recover the correct answer even at 10\% of the trace, while forced extraction fails on 42\% of these cases. The answer is recoverable from the model state, yet prompt-conditioned decoding fails to extract it. We formalize this mismatch via a total-variation bound between free and forced continuation distributions, yielding quantitative estimates of suffix-induced shift. Exploiting this asymmetry, we propose Black-box Adaptive Early Exit (\BAEE{}), which uses free continuations for both detection and extraction, truncating \textbf{70--78\% of serial generation} while \textbf{improving accuracy by 1--5\,pp} across all models. For thinking-mode models, early exit prevents post-commitment overwriting, yielding gains of up to 5.8\,pp; a cost-optimized variant achieves 68--73\% reduction at a median of 9 API calls. Code is available at https://github.com/EdWangLoDaSc/know2say.

Feng Liu, Hanyang Wang, Siyuan Shen

The study of Dynamic Facial Expression Recognition (DFER) is a nascent field of research that involves the automated recognition of facial expressions in video data. Although existing research has primarily focused on learning representations under noisy and hard samples, the issue of the coexistence of both types of samples remains unresolved. In order to overcome this challenge, this paper proposes a robust method of distinguishing between hard and noisy samples. This is achieved by evaluating the prediction agreement of the model on different sampled clips of the video. Subsequently, methodologies that reinforce the learning of hard samples and mitigate the impact of noisy samples can be employed. Moreover, to identify the principal expression in a video and enhance the model's capacity for representation learning, comprising a key expression re-sampling framework and a dual-stream hierarchical network is proposed, namely Robust Dynamic Facial Expression Recognition (RDFER). The key expression re-sampling framework is designed to identify the key expression, thereby mitigating the potential confusion caused by non-target expressions. RDFER employs two sequence models with the objective of disentangling short-term facial movements and long-term emotional changes. The proposed method has been shown to outperform current State-Of-The-Art approaches in DFER through extensive experimentation on benchmark datasets such as DFEW and FERV39K. A comprehensive analysis provides valuable insights and observations regarding the proposed agreement. This work has significant implications for the field of dynamic facial expression recognition and promotes the further development of the field of noise-consistent robust learning in dynamic facial expression recognition. The code is available from [https://github.com/Cross-Innovation-Lab/RDFER].

Hanyang Wang, Lu Wang, Chaoyun Zhang et al.

Traditional RLHF optimizes language models with coarse, scalar rewards that mask the fine-grained reasons behind success or failure, leading to slow and opaque learning. Recent work augments RL with textual critiques through prompting or reflection, improving interpretability but leaving model parameters untouched. We introduce Text2Grad, a reinforcement-learning paradigm that turns free-form textual feedback into span-level gradients. Given human (or programmatic) critiques, Text2Grad aligns each feedback phrase with the relevant token spans, converts these alignments into differentiable reward signals, and performs gradient updates that directly refine the offending portions of the model's policy. This yields precise, feedback-conditioned adjustments instead of global nudges. Text2Grad is realized through three components: (1) a high-quality feedback-annotation pipeline that pairs critiques with token spans; (2) a fine-grained reward model that predicts span-level reward on answer while generating explanatory critiques; and (3) a span-level policy optimizer that back-propagates natural-language gradients. Across summarization, code generation, and question answering, Text2Grad consistently surpasses scalar-reward RL and prompt-only baselines, providing both higher task metrics and richer interpretability. Our results demonstrate that natural-language feedback, when converted to gradients, is a powerful signal for fine-grained policy optimization. The code for our method is available at https://github.com/microsoft/Text2Grad

Ziwang Fu, Feng Liu, Hanyang Wang et al.

The audio-video based multimodal emotion recognition has attracted a lot of attention due to its robust performance. Most of the existing methods focus on proposing different cross-modal fusion strategies. However, these strategies introduce redundancy in the features of different modalities without fully considering the complementary properties between modal information, and these approaches do not guarantee the non-loss of original semantic information during intra- and inter-modal interactions. In this paper, we propose a novel cross-modal fusion network based on self-attention and residual structure (CFN-SR) for multimodal emotion recognition. Firstly, we perform representation learning for audio and video modalities to obtain the semantic features of the two modalities by efficient ResNeXt and 1D CNN, respectively. Secondly, we feed the features of the two modalities into the cross-modal blocks separately to ensure efficient complementarity and completeness of information through the self-attention mechanism and residual structure. Finally, we obtain the output of emotions by splicing the obtained fused representation with the original representation. To verify the effectiveness of the proposed method, we conduct experiments on the RAVDESS dataset. The experimental results show that the proposed CFN-SR achieves the state-of-the-art and obtains 75.76% accuracy with 26.30M parameters. Our code is available at https://github.com/skeletonNN/CFN-SR.

Fangfu Liu, Hanyang Wang, Yimo Cai et al.

With the scale capability of increasing training data, model size, and computational cost, video generation has achieved impressive results in digital creation, enabling users to express creativity across various domains. Recently, researchers in Large Language Models (LLMs) have expanded the scaling to test-time, which can significantly improve LLM performance by using more inference-time computation. Instead of scaling up video foundation models through expensive training costs, we explore the power of Test-Time Scaling (TTS) in video generation, aiming to answer the question: if a video generation model is allowed to use non-trivial amount of inference-time compute, how much can it improve generation quality given a challenging text prompt. In this work, we reinterpret the test-time scaling of video generation as a searching problem to sample better trajectories from Gaussian noise space to the target video distribution. Specifically, we build the search space with test-time verifiers to provide feedback and heuristic algorithms to guide searching process. Given a text prompt, we first explore an intuitive linear search strategy by increasing noise candidates at inference time. As full-step denoising all frames simultaneously requires heavy test-time computation costs, we further design a more efficient TTS method for video generation called Tree-of-Frames (ToF) that adaptively expands and prunes video branches in an autoregressive manner. Extensive experiments on text-conditioned video generation benchmarks demonstrate that increasing test-time compute consistently leads to significant improvements in the quality of videos. Project page: https://liuff19.github.io/Video-T1

Hanyang Wang, Fangfu Liu, Jiawei Chi et al.

Recovering 3D scenes from sparse views is a challenging task due to its inherent ill-posed problem. Conventional methods have developed specialized solutions (e.g., geometry regularization or feed-forward deterministic model) to mitigate the issue. However, they still suffer from performance degradation by minimal overlap across input views with insufficient visual information. Fortunately, recent video generative models show promise in addressing this challenge as they are capable of generating video clips with plausible 3D structures. Powered by large pretrained video diffusion models, some pioneering research start to explore the potential of video generative prior and create 3D scenes from sparse views. Despite impressive improvements, they are limited by slow inference time and the lack of 3D constraint, leading to inefficiencies and reconstruction artifacts that do not align with real-world geometry structure. In this paper, we propose VideoScene to distill the video diffusion model to generate 3D scenes in one step, aiming to build an efficient and effective tool to bridge the gap from video to 3D. Specifically, we design a 3D-aware leap flow distillation strategy to leap over time-consuming redundant information and train a dynamic denoising policy network to adaptively determine the optimal leap timestep during inference. Extensive experiments demonstrate that our VideoScene achieves faster and superior 3D scene generation results than previous video diffusion models, highlighting its potential as an efficient tool for future video to 3D applications. Project Page: https://hanyang-21.github.io/VideoScene

Fangfu Liu, Hao Li, Jiawei Chi et al.

Recovering 3D structures with open-vocabulary scene understanding from 2D images is a fundamental but daunting task. Recent developments have achieved this by performing per-scene optimization with embedded language information. However, they heavily rely on the calibrated dense-view reconstruction paradigm, thereby suffering from severe rendering artifacts and implausible semantic synthesis when limited views are available. In this paper, we introduce a novel generative framework, coined LangScene-X, to unify and generate 3D consistent multi-modality information for reconstruction and understanding. Powered by the generative capability of creating more consistent novel observations, we can build generalizable 3D language-embedded scenes from only sparse views. Specifically, we first train a TriMap video diffusion model that can generate appearance (RGBs), geometry (normals), and semantics (segmentation maps) from sparse inputs through progressive knowledge integration. Furthermore, we propose a Language Quantized Compressor (LQC), trained on large-scale image datasets, to efficiently encode language embeddings, enabling cross-scene generalization without per-scene retraining. Finally, we reconstruct the language surface fields by aligning language information onto the surface of 3D scenes, enabling open-ended language queries. Extensive experiments on real-world data demonstrate the superiority of our LangScene-X over state-of-the-art methods in terms of quality and generalizability. Project Page: https://liuff19.github.io/LangScene-X.

Hanyang Wang, Yuxuan Yang, Hongjun Wang et al.

The intelligent fault diagnosis of rotating mechanical equipment usually requires a large amount of labeled sample data. However, in practical industrial applications, acquiring enough data is both challenging and expensive in terms of time and cost. Moreover, different types of rotating mechanical equipment with different unique mechanical properties, require separate training of diagnostic models for each case. To address the challenges of limited fault samples and the lack of generalizability in prediction models for practical engineering applications, we propose a Multi-Attention Meta Transformer method for few-shot unsupervised rotating machinery fault diagnosis (MMT-FD). This framework extracts potential fault representations from unlabeled data and demonstrates strong generalization capabilities, making it suitable for diagnosing faults across various types of mechanical equipment. The MMT-FD framework integrates a time-frequency domain encoder and a meta-learning generalization model. The time-frequency domain encoder predicts status representations generated through random augmentations in the time-frequency domain. These enhanced data are then fed into a meta-learning network for classification and generalization training, followed by fine-tuning using a limited amount of labeled data. The model is iteratively optimized using a small number of contrastive learning iterations, resulting in high efficiency. To validate the framework, we conducted experiments on a bearing fault dataset and rotor test bench data. The results demonstrate that the MMT-FD model achieves 99\% fault diagnosis accuracy with only 1\% of labeled sample data, exhibiting robust generalization capabilities.

Hanyang Wang, Juergen Branke, Matthias Poloczek

Many real-world black-box optimization problems have multiple conflicting objectives. Rather than attempting to approximate the entire set of Pareto-optimal solutions, interactive preference learning allows to focus the search on the most relevant subset. However, few previous studies have exploited the fact that utility functions are usually monotonic. In this paper, we address the Bayesian Optimization with Preference Exploration (BOPE) problem and propose using a neural network ensemble as a utility surrogate model. This approach naturally integrates monotonicity and supports pairwise comparison data. Our experiments demonstrate that the proposed method outperforms state-of-the-art approaches and exhibits robustness to noise in utility evaluations. An ablation study highlights the critical role of monotonicity in enhancing performance.

Hanyang Wang, Juergen Branke, Matthias Poloczek

In many real-world optimization problems, we have prior information about what objective function values are achievable. In this paper, we study the scenario that we have either exact knowledge of the minimum value or a, possibly inexact, lower bound on its value. We propose bound-aware Bayesian optimization (BABO), a Bayesian optimization method that uses a new surrogate model and acquisition function to utilize such prior information. We present SlogGP, a new surrogate model that incorporates bound information and adapts the Expected Improvement (EI) acquisition function accordingly. Empirical results on a variety of benchmarks demonstrate the benefit of taking prior information about the optimal value into account, and that the proposed approach significantly outperforms existing techniques. Furthermore, we notice that even in the absence of prior information on the bound, the proposed SlogGP surrogate model still performs better than the standard GP model in most cases, which we explain by its larger expressiveness.

Fangfu Liu, Hanyang Wang, Shunyu Yao et al.

In recent years, there has been rapid development in 3D generation models, opening up new possibilities for applications such as simulating the dynamic movements of 3D objects and customizing their behaviors. However, current 3D generative models tend to focus only on surface features such as color and shape, neglecting the inherent physical properties that govern the behavior of objects in the real world. To accurately simulate physics-aligned dynamics, it is essential to predict the physical properties of materials and incorporate them into the behavior prediction process. Nonetheless, predicting the diverse materials of real-world objects is still challenging due to the complex nature of their physical attributes. In this paper, we propose \textbf{Physics3D}, a novel method for learning various physical properties of 3D objects through a video diffusion model. Our approach involves designing a highly generalizable physical simulation system based on a viscoelastic material model, which enables us to simulate a wide range of materials with high-fidelity capabilities. Moreover, we distill the physical priors from a video diffusion model that contains more understanding of realistic object materials. Extensive experiments demonstrate the effectiveness of our method with both elastic and plastic materials. Physics3D shows great potential for bridging the gap between the physical world and virtual neural space, providing a better integration and application of realistic physical principles in virtual environments. Project page: https://liuff19.github.io/Physics3D.

Fangfu Liu, Hanyang Wang, Weiliang Chen et al.

Recent years have witnessed the strong power of 3D generation models, which offer a new level of creative flexibility by allowing users to guide the 3D content generation process through a single image or natural language. However, it remains challenging for existing 3D generation methods to create subject-driven 3D content across diverse prompts. In this paper, we introduce a novel 3D customization method, dubbed Make-Your-3D that can personalize high-fidelity and consistent 3D content from only a single image of a subject with text description within 5 minutes. Our key insight is to harmonize the distributions of a multi-view diffusion model and an identity-specific 2D generative model, aligning them with the distribution of the desired 3D subject. Specifically, we design a co-evolution framework to reduce the variance of distributions, where each model undergoes a process of learning from the other through identity-aware optimization and subject-prior optimization, respectively. Extensive experiments demonstrate that our method can produce high-quality, consistent, and subject-specific 3D content with text-driven modifications that are unseen in subject image.

Ziwang Fu, Feng Liu, Hanyang Wang et al.

Learning modality-fused representations and processing unaligned multimodal sequences are meaningful and challenging in multimodal emotion recognition. Existing approaches use directional pairwise attention or a message hub to fuse language, visual, and audio modalities. However, those approaches introduce information redundancy when fusing features and are inefficient without considering the complementarity of modalities. In this paper, we propose an efficient neural network to learn modality-fused representations with CB-Transformer (LMR-CBT) for multimodal emotion recognition from unaligned multimodal sequences. Specifically, we first perform feature extraction for the three modalities respectively to obtain the local structure of the sequences. Then, we design a novel transformer with cross-modal blocks (CB-Transformer) that enables complementary learning of different modalities, mainly divided into local temporal learning,cross-modal feature fusion and global self-attention representations. In addition, we splice the fused features with the original features to classify the emotions of the sequences. Finally, we conduct word-aligned and unaligned experiments on three challenging datasets, IEMOCAP, CMU-MOSI, and CMU-MOSEI. The experimental results show the superiority and efficiency of our proposed method in both settings. Compared with the mainstream methods, our approach reaches the state-of-the-art with a minimum number of parameters.

Feng Liu, HanYang Wang, Jiahao Zhang et al.

The image synthesis technique is relatively well established which can generate facial images that are indistinguishable even by human beings. However, all of these approaches uses gradients to condition the output, resulting in the outputting the same image with the same input. Also, they can only generate images with basic expression or mimic an expression instead of generating compound expression. In real life, however, human expressions are of great diversity and complexity. In this paper, we propose an evolutionary algorithm (EA) assisted GAN, named EvoGAN, to generate various compound expressions with any accurate target compound expression. EvoGAN uses an EA to search target results in the data distribution learned by GAN. Specifically, we use the Facial Action Coding System (FACS) as the encoding of an EA and use a pre-trained GAN to generate human facial images, and then use a pre-trained classifier to recognize the expression composition of the synthesized images as the fitness function to guide the search of the EA. Combined random searching algorithm, various images with the target expression can be easily sythesized. Quantitative and Qualitative results are presented on several compound expressions, and the experimental results demonstrate the feasibility and the potential of EvoGAN.