Pu Yang

Pengfei Zhou, Shengcong Chen, Di Chen et al.

Robotic manipulation requires models that generate executable actions while anticipating and evaluating their future consequences before physical execution. We present $τ_0$-World Model ($τ_0$-WM), a unified video-action world model that integrates policy learning, video prediction, and action evaluation within a single future-predictive framework. Built on a shared video diffusion backbone, $τ_0$-WM provides two complementary interfaces. First, a video action model jointly predicts future visual latents and continuous action chunks from multi-view observations, language instructions, and robot state. Second, an action-conditioned video simulator rolls out candidate action chunks into multi-view futures and predicts dense task-progress scores. The model is trained on approximately $27{,}300$ hours of real-robot teleoperation, UMI-style interaction, egocentric human videos, and rollout or failure trajectories using modality-specific supervision masks. At inference time, $τ_0$-WM uses test-time computation to sample action candidates, rank them with re-denoising consistency, and invoke simulator-based rectification for low-quality candidates. On challenging long-horizon and fine-grained robotic manipulation tasks, $τ_0$-WM shows superior performance over other relevant baselines.

Pu Yang, Bin Dong

Magnetic Resonance Imaging (MRI) is a widely used medical imaging technique, but its long acquisition time can be a limiting factor in clinical settings. To address this issue, researchers have been exploring ways to reduce the acquisition time while maintaining the reconstruction quality. Previous works have focused on finding either sparse samplers with a fixed reconstructor or finding reconstructors with a fixed sampler. However, these approaches do not fully utilize the potential of joint learning of samplers and reconstructors. In this paper, we propose an alternating training framework for jointly learning a good pair of samplers and reconstructors via deep reinforcement learning (RL). In particular, we consider the process of MRI sampling as a sampling trajectory controlled by a sampler, and introduce a novel sparse-reward Partially Observed Markov Decision Process (POMDP) to formulate the MRI sampling trajectory. Compared to the dense-reward POMDP used in existing works, the proposed sparse-reward POMDP is more computationally efficient and has a provable advantage. Moreover, the proposed framework, called L2SR (Learning to Sample and Reconstruct), overcomes the training mismatch problem that arises in previous methods that use dense-reward POMDP. By alternately updating samplers and reconstructors, L2SR learns a pair of samplers and reconstructors that achieve state-of-the-art reconstruction performances on the fastMRI dataset. Codes are available at \url{https://github.com/yangpuPKU/L2SR-Learning-to-Sample-and-Reconstruct}.

Yuchen Yang, Yaru Zhao, Pu Yang et al.

While Mixture-of-Experts (MoE) architectures substantially bolster the expressive power of large-language models, their prohibitive memory footprint severely impedes the practical deployment on resource-constrained edge devices, especially when model behavior must be preserved without relying on lossy quantization. In this paper, we present ZipMoE, an efficient and semantically lossless on-device MoE serving system. ZipMoE exploits the synergy between the hardware properties of edge devices and the statistical redundancy inherent to MoE parameters via a caching-scheduling co-design with provable performance guarantee. Fundamentally, our design shifts the paradigm of on-device MoE inference from an I/O-bound bottleneck to a compute-centric workflow that enables efficient parallelization. We implement a prototype of ZipMoE and conduct extensive experiments on representative edge computing platforms using popular open-source MoE models and real-world workloads. Our evaluation reveals that ZipMoE achieves up to $72.77\%$ inference latency reduction and up to $6.76\times$ higher throughput than the state-of-the-art systems.

Yi Wang, Xinchen Li, Pengwei Xie et al.

Generalist robot policies increasingly benefit from large-scale pretraining, but offline data alone is insufficient for robust real-world deployment. Deployed robots encounter distribution shifts, long-tail failures, task variations, and human correction opportunities that fixed demonstration datasets cannot fully capture. We present Learning While Deploying (LWD), a fleet-scale offline-to-online reinforcement learning framework for continual post-training of generalist Vision-Language-Action (VLA) policies. Starting from a pretrained VLA policy, LWD closes the loop between deployment, shared physical experience, policy improvement, and redeployment by using autonomous rollouts and human interventions collected across a robot fleet. To stabilize learning from heterogeneous, sparse-reward fleet data, LWD combines Distributional Implicit Value Learning (DIVL) for robust value estimation with Q-learning via Adjoint Matching (QAM) for policy extraction in flow-based VLA action generators. We validate LWD on a fleet of 16 dual-arm robots across eight real-world manipulation tasks, including semantic grocery restocking and 3--5 minute long-horizon tasks. A single generalist policy improves as fleet experience accumulates, reaching an average success rate of 95%, with the largest gains on long-horizon tasks.

Ziju Shen, Naohao Huang, Fanyi Yang et al.

Nowadays, formal theorem provers have made monumental progress on high-school and competition-level mathematics, but few of them generalize to more advanced mathematics. In this paper, we present REAL-Prover, a new open-source stepwise theorem prover for Lean 4 to push this boundary. This prover, based on our fine-tuned large language model (REAL-Prover-v1) and integrated with a retrieval system (Leansearch-PS), notably boosts performance on solving college-level mathematics problems. To train REAL-Prover-v1, we developed HERALD-AF, a data extraction pipeline that converts natural language math problems into formal statements, and a new open-source Lean 4 interactive environment (Jixia-interactive) to facilitate synthesis data collection. In our experiments, our prover using only supervised fine-tune achieves competitive results with a 23.7% success rate (Pass@64) on the ProofNet dataset-comparable to state-of-the-art (SOTA) models. To further evaluate our approach, we introduce FATE-M, a new benchmark focused on algebraic problems, where our prover achieves a SOTA success rate of 56.7% (Pass@64).

Elvis Dohmatob, Yunzhen Feng, Pu Yang et al. · pku

As AI model size grows, neural scaling laws have become a crucial tool to predict the improvements of large models when increasing capacity and the size of original (human or natural) training data. Yet, the widespread use of popular models means that the ecosystem of online data and text will co-evolve to progressively contain increased amounts of synthesized data. In this paper we ask: How will the scaling laws change in the inevitable regime where synthetic data makes its way into the training corpus? Will future models, still improve, or be doomed to degenerate up to total (model) collapse? We develop a theoretical framework of model collapse through the lens of scaling laws. We discover a wide range of decay phenomena, analyzing loss of scaling, shifted scaling with number of generations, the ''un-learning" of skills, and grokking when mixing human and synthesized data. Our theory is validated by large-scale experiments with a transformer on an arithmetic task and text generation using the large language model Llama2.

Pu Yang, Bin Dong

Image captioning is a critical task at the intersection of computer vision and natural language processing, with wide-ranging applications across various domains. For complex tasks such as diagnostic report generation, deep learning models require not only domain-specific image-caption datasets but also the incorporation of relevant general knowledge to provide contextual accuracy. Existing approaches exhibit inherent limitations: specialized models excel in capturing domain-specific details but lack generalization, while vision-language models (VLMs) built on large language models (LLMs) leverage general knowledge but struggle with domain-specific adaptation. To address these limitations, this paper proposes a novel agent-enhanced model collaboration framework, which we call MoColl, designed to effectively integrate domain-specific and general knowledge. Specifically, our approach is to decompose complex image captioning tasks into a series of interconnected question-answer subtasks. A trainable visual question answering (VQA) model is employed as a specialized tool to focus on domain-specific visual analysis, answering task-specific questions based on image content. Concurrently, an LLM-based agent with general knowledge formulates these questions and synthesizes the resulting question-answer pairs into coherent captions. Beyond its role in leveraging the VQA model, the agent further guides its training to enhance its domain-specific capabilities. Experimental results on radiology report generation validate the effectiveness of the proposed framework, demonstrating significant improvements in the quality of generated reports.

Pu Yang, J. A. Barria

This paper presents a Wavelet Probabilistic Recurrent Convolutional Network (WPRCN) for Multivariate Time Series Classification (MTSC), especially effective in handling non-stationary environments, data scarcity and noise perturbations. We introduce a versatile wavelet probabilistic module designed to extract and analyse the probabilistic features, which can seamlessly integrate with a variety of neural network architectures. This probabilistic module comprises an Adaptive Wavelet Probabilistic Feature Generator (AWPG) and a Channel Attention-based Probabilistic Temporal Convolutional Network (APTCN). Such formulation extends the application of wavelet probabilistic neural networks to deep neural networks for MTSC. The AWPG constructs an ensemble probabilistic model addressing different data scarcities and non-stationarity; it adaptively selects the optimal ones and generates probabilistic features for APTCN. The APTCN analyses the correlations of the features and forms a comprehensive feature space with existing MTSC models for classification. Here, we instantiate the proposed module to work in parallel with a Long Short-Term Memory (LSTM) network and a Causal Fully Convolutional Network (C-FCN), demonstrating its broad applicability in time series analysis. The WPRCN is evaluated on 30 diverse MTS datasets and outperforms all the benchmark algorithms on average accuracy and rank, exhibiting pronounced strength in handling scarce data and physiological data subject to perturbations and non-stationarities.

Pu Yang, Yunzhen Feng, Ziyuan Chen et al.

Modern foundation models often undergo iterative ``bootstrapping'' in their post-training phase: a model generates synthetic data, an external verifier filters out low-quality samples, and the high-quality subset is used for further fine-tuning. Over multiple iterations, the model performance improves, raising a crucial question: How should the total budget for generation and training be allocated across iterations to maximize final performance? In this work, we develop a theoretical framework for analyzing budget allocation strategies. Specifically, we show that constant policies fail to converge with high probability, while increasing policies -- particularly exponential growth policies -- exhibit significant theoretical advantages. Experiments on image denoising with diffusion probabilistic models and math reasoning with large language models show that both exponential and polynomial growth policies consistently outperform constant policies, with exponential policies often providing more stable performance.

Yunzhen Feng, Elvis Dohmatob, Pu Yang et al.

Large Language Models (LLM) are increasingly trained on data generated by other LLM, either because generated text and images become part of the pre-training corpus, or because synthetized data is used as a replacement for expensive human-annotation. This raises concerns about \emph{model collapse}, a drop in model performance when their training sets include generated data. Considering that it is easier for both humans and machines to tell between good and bad examples than to generate high-quality samples, we investigate the use of verification on synthesized data to prevent model collapse. We provide a theoretical characterization using Gaussian mixtures, linear classifiers, and linear verifiers to derive conditions with measurable proxies to assess whether the verifier can effectively select synthesized data that leads to optimal performance. We experiment with two practical tasks -- computing matrix eigenvalues with transformers and news summarization with LLMs -- which both exhibit model collapse when trained on generated data, and show that verifiers, even imperfect ones, can indeed be harnessed to prevent model collapse and that our proposed proxy measure strongly correlates with performance.

Austin Reiter, Menglin Jia, Pu Yang et al.

Many vision-related tasks benefit from reasoning over multiple modalities to leverage complementary views of data in an attempt to learn robust embedding spaces. Most deep learning-based methods rely on a late fusion technique whereby multiple feature types are encoded and concatenated and then a multi layer perceptron (MLP) combines the fused embedding to make predictions. This has several limitations, such as an unnatural enforcement that all features be present at all times as well as constraining only a constant number of occurrences of a feature modality at any given time. Furthermore, as more modalities are added, the concatenated embedding grows. To mitigate this, we propose Deep Multi-Modal Sets: a technique that represents a collection of features as an unordered set rather than one long ever-growing fixed-size vector. The set is constructed so that we have invariance both to permutations of the feature modalities as well as to the cardinality of the set. We will also show that with particular choices in our model architecture, we can yield interpretable feature performance such that during inference time we can observe which modalities are most contributing to the prediction.With this in mind, we demonstrate a scalable, multi-modal framework that reasons over different modalities to learn various types of tasks. We demonstrate new state-of-the-art performance on two multi-modal datasets (Ads-Parallelity [34] and MM-IMDb [1]).