Sirui Xie

Peiyu Yu, Sirui Xie, Xiaojian Ma et al.

Latent space Energy-Based Models (EBMs), also known as energy-based priors, have drawn growing interests in generative modeling. Fueled by its flexibility in the formulation and strong modeling power of the latent space, recent works built upon it have made interesting attempts aiming at the interpretability of text modeling. However, latent space EBMs also inherit some flaws from EBMs in data space; the degenerate MCMC sampling quality in practice can lead to poor generation quality and instability in training, especially on data with complex latent structures. Inspired by the recent efforts that leverage diffusion recovery likelihood learning as a cure for the sampling issue, we introduce a novel symbiosis between the diffusion models and latent space EBMs in a variational learning framework, coined as the latent diffusion energy-based model. We develop a geometric clustering-based regularization jointly with the information bottleneck to further improve the quality of the learned latent space. Experiments on several challenging tasks demonstrate the superior performance of our model on interpretable text modeling over strong counterparts.

Aoyang Qin, Feng Gao, Qing Li et al.

Conventional imitation learning assumes access to the actions of demonstrators, but these motor signals are often non-observable in naturalistic settings. Additionally, sequential decision-making behaviors in these settings can deviate from the assumptions of a standard Markov Decision Process (MDP). To address these challenges, we explore deep generative modeling of state-only sequences with non-Markov Decision Process (nMDP), where the policy is an energy-based prior in the latent space of the state transition generator. We develop maximum likelihood estimation to achieve model-based imitation, which involves short-run MCMC sampling from the prior and importance sampling for the posterior. The learned model enables \textit{decision-making as inference}: model-free policy execution is equivalent to prior sampling, model-based planning is posterior sampling initialized from the policy. We demonstrate the efficacy of the proposed method in a prototypical path planning task with non-Markovian constraints and show that the learned model exhibits strong performances in challenging domains from the MuJoCo suite.

Peiyu Yu, Dinghuai Zhang, Hengzhi He et al.

Noise Contrastive Estimation (NCE) has fueled major breakthroughs in representation learning and generative modeling. Yet a long-standing challenge remains: accurately estimating ratios between distributions that differ substantially, which significantly limits the applicability of NCE on modern high-dimensional and multimodal datasets. We revisit this problem from a less explored perspective: the magnitude of the noise distribution. Specifically, we show that with a virtually scaled (\ie, artificially increased) noise magnitude, the gradient of the NCE objective can closely align with that of Maximum Likelihood, enabling a trajectory-wise approximation from NCE to MLE, and faster convergence both theoretically and empirically. Building on this insight, we introduce ``Noisier'' NCE, a simple drop-in modification to vanilla NCE that incurs little to no extra computational cost, while effectively handling density-ratio estimation in challenging regimes where traditional MLE and NCE struggle. Beyond improving classical density-ratio learning, ``Noisier'' NCE proves broadly applicable: it achieves strong results across image modeling, anomaly detection, and offline black-box optimization. On CIFAR-10 and ImageNet64x64 datasets, it yields 10-step and even 1-step samplers that match or surpass state-of-the-art methods, while cutting training iterations by up to half.

Peiyu Yu, Yaxuan Zhu, Sirui Xie et al.

Latent space Energy-Based Models (EBMs), also known as energy-based priors, have drawn growing interests in the field of generative modeling due to its flexibility in the formulation and strong modeling power of the latent space. However, the common practice of learning latent space EBMs with non-convergent short-run MCMC for prior and posterior sampling is hindering the model from further progress; the degenerate MCMC sampling quality in practice often leads to degraded generation quality and instability in training, especially with highly multi-modal and/or high-dimensional target distributions. To remedy this sampling issue, in this paper we introduce a simple but effective diffusion-based amortization method for long-run MCMC sampling and develop a novel learning algorithm for the latent space EBM based on it. We provide theoretical evidence that the learned amortization of MCMC is a valid long-run MCMC sampler. Experiments on several image modeling benchmark datasets demonstrate the superior performance of our method compared with strong counterparts

Leiyao Cui, Zihang Zhao, Sirui Xie et al. · pku

Autonomous manipulation of articulated objects remains a fundamental challenge for robots in human environments. Vision-based methods can infer hidden kinematics but can yield imprecise estimates on unfamiliar objects. Tactile approaches achieve robust control through contact feedback but require accurate initialization. This suggests a natural synergy: vision for global guidance, touch for local precision. Yet no framework systematically exploits this complementarity for generalized articulated manipulation. Here we present Vi-TacMan, which uses vision to propose grasps and coarse directions that seed a tactile controller for precise execution. By incorporating surface normals as geometric priors and modeling directions via von Mises-Fisher distributions, our approach achieves significant gains over baselines (all p<0.0001). Critically, manipulation succeeds without explicit kinematic models -- the tactile controller refines coarse visual estimates through real-time contact regulation. Tests on more than 50,000 simulated and diverse real-world objects confirm robust cross-category generalization. This work establishes that coarse visual cues suffice for reliable manipulation when coupled with tactile feedback, offering a scalable paradigm for autonomous systems in unstructured environments.

Zhide Zhong, Jiakai Cao, Songen Gu et al.

We present ASSIST, an object-wise neural radiance field as a panoptic representation for compositional and realistic simulation. Central to our approach is a novel scene node data structure that stores the information of each object in a unified fashion, allowing online interaction in both intra- and cross-scene settings. By incorporating a differentiable neural network along with the associated bounding box and semantic features, the proposed structure guarantees user-friendly interaction on independent objects to scale up novel view simulation. Objects in the scene can be queried, added, duplicated, deleted, transformed, or swapped simply through mouse/keyboard controls or language instructions. Experiments demonstrate the efficacy of the proposed method, where scaled realistic simulation can be achieved through interactive editing and compositional rendering, with color images, depth images, and panoptic segmentation masks generated in a 3D consistent manner.

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.

Aoyang Qin, Deqian Kong, Wei Wang et al.

Conventional Reinforcement Learning (RL) algorithms, typically focused on estimating or maximizing expected returns, face challenges when refining offline pretrained models with online experiences. This paper introduces Generative Actor Critic (GAC), a novel framework that decouples sequential decision-making by reframing \textit{policy evaluation} as learning a generative model of the joint distribution over trajectories and returns, $p(τ, y)$, and \textit{policy improvement} as performing versatile inference on this learned model. To operationalize GAC, we introduce a specific instantiation based on a latent variable model that features continuous latent plan vectors. We develop novel inference strategies for both \textit{exploitation}, by optimizing latent plans to maximize expected returns, and \textit{exploration}, by sampling latent plans conditioned on dynamically adjusted target returns. Experiments on Gym-MuJoCo and Maze2D benchmarks demonstrate GAC's strong offline performance and significantly enhanced offline-to-online improvement compared to state-of-the-art methods, even in absence of step-wise rewards.

Shoukang Hu, Sirui Xie, Hehui Zheng et al.

If NAS methods are solutions, what is the problem? Most existing NAS methods require two-stage parameter optimization. However, performance of the same architecture in the two stages correlates poorly. In this work, we propose a new problem definition for NAS, task-specific end-to-end, based on this observation. We argue that given a computer vision task for which a NAS method is expected, this definition can reduce the vaguely-defined NAS evaluation to i) accuracy of this task and ii) the total computation consumed to finally obtain a model with satisfying accuracy. Seeing that most existing methods do not solve this problem directly, we propose DSNAS, an efficient differentiable NAS framework that simultaneously optimizes architecture and parameters with a low-biased Monte Carlo estimate. Child networks derived from DSNAS can be deployed directly without parameter retraining. Comparing with two-stage methods, DSNAS successfully discovers networks with comparable accuracy (74.4%) on ImageNet in 420 GPU hours, reducing the total time by more than 34%. Our implementation is available at https://github.com/SNAS-Series/SNAS-Series.

Sirui Xie, Hehui Zheng, Chunxiao Liu et al.

We propose Stochastic Neural Architecture Search (SNAS), an economical end-to-end solution to Neural Architecture Search (NAS) that trains neural operation parameters and architecture distribution parameters in same round of back-propagation, while maintaining the completeness and differentiability of the NAS pipeline. In this work, NAS is reformulated as an optimization problem on parameters of a joint distribution for the search space in a cell. To leverage the gradient information in generic differentiable loss for architecture search, a novel search gradient is proposed. We prove that this search gradient optimizes the same objective as reinforcement-learning-based NAS, but assigns credits to structural decisions more efficiently. This credit assignment is further augmented with locally decomposable reward to enforce a resource-efficient constraint. In experiments on CIFAR-10, SNAS takes less epochs to find a cell architecture with state-of-the-art accuracy than non-differentiable evolution-based and reinforcement-learning-based NAS, which is also transferable to ImageNet. It is also shown that child networks of SNAS can maintain the validation accuracy in searching, with which attention-based NAS requires parameter retraining to compete, exhibiting potentials to stride towards efficient NAS on big datasets. We have released our implementation at https://github.com/SNAS-Series/SNAS-Series.

Deqian Kong, Minglu Zhao, Dehong Xu et al.

We propose a novel class of language models, Latent Thought Models (LTMs), which incorporate explicit latent thought vectors that follow an explicit prior model in latent space. These latent thought vectors guide the autoregressive generation of ground tokens through a Transformer decoder. Training employs a dual-rate optimization process within the classical variational Bayes framework: fast learning of local variational parameters for the posterior distribution of latent vectors (inference-time computation), and slow learning of global decoder parameters. Empirical studies reveal that LTMs possess additional scaling dimensions beyond traditional Large Language Models (LLMs), such as the number of iterations in inference-time computation and number of latent thought vectors. Higher sample efficiency can be achieved by increasing training compute per token, with further gains possible by trading model size for more inference steps. Designed based on these scaling properties, LTMs demonstrate superior sample and parameter efficiency compared to autoregressive models and discrete diffusion models. They significantly outperform these counterparts in validation perplexity and zero-shot language modeling tasks. Additionally, LTMs exhibit emergent few-shot in-context reasoning capabilities that scale with model size, and achieve competitive performance in conditional and unconditional text generation.

Deqian Kong, Dehong Xu, Minglu Zhao et al.

In tasks aiming for long-term returns, planning becomes essential. We study generative modeling for planning with datasets repurposed from offline reinforcement learning. Specifically, we identify temporal consistency in the absence of step-wise rewards as one key technical challenge. We introduce the Latent Plan Transformer (LPT), a novel model that leverages a latent variable to connect a Transformer-based trajectory generator and the final return. LPT can be learned with maximum likelihood estimation on trajectory-return pairs. In learning, posterior sampling of the latent variable naturally integrates sub-trajectories to form a consistent abstraction despite the finite context. At test time, the latent variable is inferred from an expected return before policy execution, realizing the idea of planning as inference. Our experiments demonstrate that LPT can discover improved decisions from sub-optimal trajectories, achieving competitive performance across several benchmarks, including Gym-Mujoco, Franka Kitchen, Maze2D, and Connect Four. It exhibits capabilities in nuanced credit assignments, trajectory stitching, and adaptation to environmental contingencies. These results validate that latent variable inference can be a strong alternative to step-wise reward prompting.

Peiyu Yu, Suraj Kothawade, Sirui Xie et al.

Most post-training methods for text-to-image samplers focus on model weights: either fine-tuning the backbone for alignment or distilling it for few-step efficiency. We take a different route: rescheduling the sampling timeline of a frozen sampler. Instead of a fixed, global schedule, we learn instance-level (prompt- and noise-conditioned) schedules through a single-pass Dirichlet policy. To ensure accurate gradient estimates in high-dimensional policy learning, we introduce a novel reward baseline based on a principled James-Stein estimator; it provably achieves lower estimation errors than commonly used variants and leads to superior performance. Our rescheduled samplers consistently improve text-image alignment including text rendering and compositional control across modern Stable Diffusion and Flux model families. Additionally, a 5-step Flux-Dev sampler with our schedules can attain generation quality comparable to deliberately distilled samplers like Flux-Schnell. We thus position our scheduling framework as an emerging model-agnostic post-training lever that unlocks additional generative potential in pretrained samplers.

Shuwen Qiu, Sirui Xie, Lifeng Fan et al.

Humans communicate with graphical sketches apart from symbolic languages. Primarily focusing on the latter, recent studies of emergent communication overlook the sketches; they do not account for the evolution process through which symbolic sign systems emerge in the trade-off between iconicity and symbolicity. In this work, we take the very first step to model and simulate this process via two neural agents playing a visual communication game; the sender communicates with the receiver by sketching on a canvas. We devise a novel reinforcement learning method such that agents are evolved jointly towards successful communication and abstract graphical conventions. To inspect the emerged conventions, we define three fundamental properties -- iconicity, symbolicity, and semanticity -- and design evaluation methods accordingly. Our experimental results under different controls are consistent with the observation in studies of human graphical conventions. Of note, we find that evolved sketches can preserve the continuum of semantics under proper environmental pressures. More interestingly, co-evolved agents can switch between conventionalized and iconic communication based on their familiarity with referents. We hope the present research can pave the path for studying emergent communication with the modality of sketches.

Chi Zhang, Sirui Xie, Baoxiong Jia et al.

Is intelligence realized by connectionist or classicist? While connectionist approaches have achieved superhuman performance, there has been growing evidence that such task-specific superiority is particularly fragile in systematic generalization. This observation lies in the central debate between connectionist and classicist, wherein the latter continually advocates an algebraic treatment in cognitive architectures. In this work, we follow the classicist's call and propose a hybrid approach to improve systematic generalization in reasoning. Specifically, we showcase a prototype with algebraic representation for the abstract spatial-temporal reasoning task of Raven's Progressive Matrices (RPM) and present the ALgebra-Aware Neuro-Semi-Symbolic (ALANS) learner. The ALANS learner is motivated by abstract algebra and the representation theory. It consists of a neural visual perception frontend and an algebraic abstract reasoning backend: the frontend summarizes the visual information from object-based representation, while the backend transforms it into an algebraic structure and induces the hidden operator on the fly. The induced operator is later executed to predict the answer's representation, and the choice most similar to the prediction is selected as the solution. Extensive experiments show that by incorporating an algebraic treatment, the ALANS learner outperforms various pure connectionist models in domains requiring systematic generalization. We further show the generative nature of the learned algebraic representation; it can be decoded by isomorphism to generate an answer.

Peiyu Yu, Sirui Xie, Xiaojian Ma et al.

We present Deep Region Competition (DRC), an algorithm designed to extract foreground objects from images in a fully unsupervised manner. Foreground extraction can be viewed as a special case of generic image segmentation that focuses on identifying and disentangling objects from the background. In this work, we rethink the foreground extraction by reconciling energy-based prior with generative image modeling in the form of Mixture of Experts (MoE), where we further introduce the learned pixel re-assignment as the essential inductive bias to capture the regularities of background regions. With this modeling, the foreground-background partition can be naturally found through Expectation-Maximization (EM). We show that the proposed method effectively exploits the interaction between the mixture components during the partitioning process, which closely connects to region competition, a seminal approach for generic image segmentation. Experiments demonstrate that DRC exhibits more competitive performances on complex real-world data and challenging multi-object scenes compared with prior methods. Moreover, we show empirically that DRC can potentially generalize to novel foreground objects even from categories unseen during training.

Sirui Xie, Xiaojian Ma, Peiyu Yu et al.

Humans learn compositional and causal abstraction, \ie, knowledge, in response to the structure of naturalistic tasks. When presented with a problem-solving task involving some objects, toddlers would first interact with these objects to reckon what they are and what can be done with them. Leveraging these concepts, they could understand the internal structure of this task, without seeing all of the problem instances. Remarkably, they further build cognitively executable strategies to \emph{rapidly} solve novel problems. To empower a learning agent with similar capability, we argue there shall be three levels of generalization in how an agent represents its knowledge: perceptual, conceptual, and algorithmic. In this paper, we devise the very first systematic benchmark that offers joint evaluation covering all three levels. This benchmark is centered around a novel task domain, HALMA, for visual concept development and rapid problem-solving. Uniquely, HALMA has a minimum yet complete concept space, upon which we introduce a novel paradigm to rigorously diagnose and dissect learning agents' capability in understanding and generalizing complex and structural concepts. We conduct extensive experiments on reinforcement learning agents with various inductive biases and carefully report their proficiency and weakness.

Sirui Xie, Feng Gao, Song-Chun Zhu

In searching for a generalizable representation of temporally extended tasks, we spot two necessary constituents: the utility needs to be non-Markovian to transfer temporal relations invariant to a probability shift, the utility also needs to be lifted to abstract out specific grounding objects. In this work, we study learning such utility from human demonstrations. While inverse reinforcement learning (IRL) has been accepted as a general framework of utility learning, its fundamental formulation is one concrete Markov Decision Process. Thus the learned reward function does not specify the task independently of the environment. Going beyond that, we define a domain of generalization that spans a set of planning problems following a schema. We hence propose a new quest, Generalized Inverse Planning, for utility learning in this domain. We further outline a computational framework, Maximum Entropy Inverse Planning (MEIP), that learns non-Markovian utility and associated concepts in a generative manner. The learned utility and concepts form a task representation that generalizes regardless of probability shift or structural change. Seeing that the proposed generalization problem has not been widely studied yet, we carefully define an evaluation protocol, with which we illustrate the effectiveness of MEIP on two proof-of-concept domains and one challenging task: learning to fold from demonstrations.

Sirui Xie, Shoukang Hu, Xinjiang Wang et al.

Controversy exists on whether differentiable neural architecture search methods discover wiring topology effectively. To understand how wiring topology evolves, we study the underlying mechanism of several existing differentiable NAS frameworks. Our investigation is motivated by three observed searching patterns of differentiable NAS: 1) they search by growing instead of pruning; 2) wider networks are more preferred than deeper ones; 3) no edges are selected in bi-level optimization. To anatomize these phenomena, we propose a unified view on searching algorithms of existing frameworks, transferring the global optimization to local cost minimization. Based on this reformulation, we conduct empirical and theoretical analyses, revealing implicit inductive biases in the cost's assignment mechanism and evolution dynamics that cause the observed phenomena. These biases indicate strong discrimination towards certain topologies. To this end, we pose questions that future differentiable methods for neural wiring discovery need to confront, hoping to evoke a discussion and rethinking on how much bias has been enforced implicitly in existing NAS methods.

Junning Huang, Sirui Xie, Jiankai Sun et al.

Autonomous driving systems have a pipeline of perception, decision, planning, and control. The decision module processes information from the perception module and directs the execution of downstream planning and control modules. On the other hand, the recent success of deep learning suggests that this pipeline could be replaced by end-to-end neural control policies, however, safety cannot be well guaranteed for the data-driven neural networks. In this work, we propose a hybrid framework to learn neural decisions in the classical modular pipeline through end-to-end imitation learning. This hybrid framework can preserve the merits of the classical pipeline such as the strict enforcement of physical and logical constraints while learning complex driving decisions from data. To circumvent the ambiguous annotation of human driving decisions, our method learns high-level driving decisions by imitating low-level control behaviors. We show in the simulation experiments that our modular driving agent can generalize its driving decision and control to various complex scenarios where the rule-based programs fail. It can also generate smoother and safer driving trajectories than end-to-end neural policies.

Peiwen Lin, Peng Sun, Guangliang Cheng et al.

Designing a lightweight semantic segmentation network often requires researchers to find a trade-off between performance and speed, which is always empirical due to the limited interpretability of neural networks. In order to release researchers from these tedious mechanical trials, we propose a Graph-guided Architecture Search (GAS) pipeline to automatically search real-time semantic segmentation networks. Unlike previous works that use a simplified search space and stack a repeatable cell to form a network, we introduce a novel search mechanism with new search space where a lightweight model can be effectively explored through the cell-level diversity and latencyoriented constraint. Specifically, to produce the cell-level diversity, the cell-sharing constraint is eliminated through the cell-independent manner. Then a graph convolution network (GCN) is seamlessly integrated as a communication mechanism between cells. Finally, a latency-oriented constraint is endowed into the search process to balance the speed and performance. Extensive experiments on Cityscapes and CamVid datasets demonstrate that GAS achieves the new state-of-the-art trade-off between accuracy and speed. In particular, on Cityscapes dataset, GAS achieves the new best performance of 73.5% mIoU with speed of 108.4 FPS on Titan Xp.

Sirui Xie, Junning Huang, Lanxin Lei et al.

Reinforcement learning agents need exploratory behaviors to escape from local optima. These behaviors may include both immediate dithering perturbation and temporally consistent exploration. To achieve these, a stochastic policy model that is inherently consistent through a period of time is in desire, especially for tasks with either sparse rewards or long term information. In this work, we introduce a novel on-policy temporally consistent exploration strategy - Neural Adaptive Dropout Policy Exploration (NADPEx) - for deep reinforcement learning agents. Modeled as a global random variable for conditional distribution, dropout is incorporated to reinforcement learning policies, equipping them with inherent temporal consistency, even when the reward signals are sparse. Two factors, gradients' alignment with the objective and KL constraint in policy space, are discussed to guarantee NADPEx policy's stable improvement. Our experiments demonstrate that NADPEx solves tasks with sparse reward while naive exploration and parameter noise fail. It yields as well or even faster convergence in the standard mujoco benchmark for continuous control.