Guowei Xu

Guowei Xu, Zhenting Qi, Huangyuan Su et al.

Search has been proposed as an effective method for self-improving language models and agentic systems, both for post-training sample generation and for inference. However, widely used methods such as best-of-N sampling and tree search face two fundamental limitations: they are guided by sparse verification signals, and they construct candidates primarily through autoregressive expansion, restricting exploration to regions with substantial model probability mass. To address these, we propose Bidirectional Evolutionary Search (BES), a search framework that couples forward candidate evolution with backward goal decomposition. In the forward search, BES augments standard expansion with evolution operators that recombine partial trajectories to generate candidates that are difficult to obtain from a single model rollout. In the backward search, BES recursively decomposes the original task into checkable subgoals, producing dense intermediate feedback that guides forward search. We provide theoretical motivation showing that candidates generated by expansion-only search are confined to a narrow entropy shell while evolutionary operators can escape it, and that backward search can exponentially reduce the number of required samples to find a correct answer. Experiments show that on challenging post-training tasks where mainstream post-training algorithms fail to improve, BES enables consistent gains, and on three open problem solving benchmarks at inference time, BES outperforms existing open-source frameworks in both average and best-case performance. Code and trained models are available at https://github.com/Embodied-Minds-Lab/BES.

Zhenting Qi, Huangyuan Su, Ao Qu et al.

How can a population of agents self-orchestrate and self-adapt into stronger collective intelligence without centralized control? Inspired by Friedrich Hayek's economic theory of decentralized coordination in markets, we study this question through an agent economy in which agents compete via auctions for the right to act, exchange payments, and accumulate wealth from environmental rewards. These simple economic signals induce decentralized credit assignment, driving planning without global orchestration or explicit communication protocols. The population evolves through economic selection: effective agents accumulate wealth and are mutated via exploitation, while ineffective ones go bankrupt and are replaced via exploration. We show that, initialized with weak agents, the economy produces emergent multi-step reasoning strategies and outperforms stronger monolithic baselines across five agentic tasks, including mathematical reasoning, financial research, scientific research, accelerator design, and distributed-system optimization. We further provide theoretical insights into how economic dynamics shape agent behaviors, linking local incentives to long-term global performance. Our results suggest a new path to multi-agent intelligence: rather than engineering coordination, we can design decentralized incentive structures under which it automatically emerges.

Guowei Xu, Jiale Tao, Wen Li et al.

In the realm of stochastic human motion prediction (SHMP), researchers have often turned to generative models like GANS, VAEs and diffusion models. However, most previous approaches have struggled to accurately predict motions that are both realistic and coherent with past motion due to a lack of guidance on the latent distribution. In this paper, we introduce Semantic Latent Directions (SLD) as a solution to this challenge, aiming to constrain the latent space to learn meaningful motion semantics and enhance the accuracy of SHMP. SLD defines a series of orthogonal latent directions and represents the hypothesis of future motion as a linear combination of these directions. By creating such an information bottleneck, SLD excels in capturing meaningful motion semantics, thereby improving the precision of motion predictions. Moreover, SLD offers controllable prediction capabilities by adjusting the coefficients of the latent directions during the inference phase. Expanding on SLD, we introduce a set of motion queries to enhance the diversity of predictions. By aligning these motion queries with the SLD space, SLD is further promoted to more accurate and coherent motion predictions. Through extensive experiments conducted on widely used benchmarks, we showcase the superiority of our method in accurately predicting motions while maintaining a balance of realism and diversity. Our code and pretrained models are available at https://github.com/GuoweiXu368/SLD-HMP.

Guowei Xu, Ruijie Zheng, Yongyuan Liang et al. · tsinghua

Visual reinforcement learning (RL) has shown promise in continuous control tasks. Despite its progress, current algorithms are still unsatisfactory in virtually every aspect of the performance such as sample efficiency, asymptotic performance, and their robustness to the choice of random seeds. In this paper, we identify a major shortcoming in existing visual RL methods that is the agents often exhibit sustained inactivity during early training, thereby limiting their ability to explore effectively. Expanding upon this crucial observation, we additionally unveil a significant correlation between the agents' inclination towards motorically inactive exploration and the absence of neuronal activity within their policy networks. To quantify this inactivity, we adopt dormant ratio as a metric to measure inactivity in the RL agent's network. Empirically, we also recognize that the dormant ratio can act as a standalone indicator of an agent's activity level, regardless of the received reward signals. Leveraging the aforementioned insights, we introduce DrM, a method that uses three core mechanisms to guide agents' exploration-exploitation trade-offs by actively minimizing the dormant ratio. Experiments demonstrate that DrM achieves significant improvements in sample efficiency and asymptotic performance with no broken seeds (76 seeds in total) across three continuous control benchmark environments, including DeepMind Control Suite, MetaWorld, and Adroit. Most importantly, DrM is the first model-free algorithm that consistently solves tasks in both the Dog and Manipulator domains from the DeepMind Control Suite as well as three dexterous hand manipulation tasks without demonstrations in Adroit, all based on pixel observations.

Guowei Xu, Peng Jin, Ziang Wu et al. · tsinghua

Large language models have demonstrated substantial advancements in reasoning capabilities. However, current Vision-Language Models (VLMs) often struggle to perform systematic and structured reasoning, especially when handling complex visual question-answering tasks. In this work, we introduce LLaVA-CoT, a large VLM designed to conduct autonomous multistage reasoning. Unlike chain-of-thought prompting, LLaVA-CoT independently engages in sequential stages of summarization, visual interpretation, logical reasoning, and conclusion generation. This structured approach enables LLaVA-CoT to achieve marked improvements on reasoning-intensive tasks. To accomplish this, we construct the LLaVA-CoT-100k dataset, integrating samples from various visual question answering sources and providing structured reasoning annotations. Besides, we propose a test-time stage-wise retracing search method (SWIRES), which enables effective and efficient test-time scaling. Remarkably, with only 100k training samples and test-time scaling, LLaVA-CoT not only outperforms its base model by 9.4% on a wide range of multimodal reasoning benchmarks, but also surpasses the performance of larger and even closed-source models, such as Gemini-1.5-pro, GPT-4o-mini, and Llama-3.2-90B-Vision-Instruct. The code, dataset, and pre-trained weights are publicly available at https://github.com/PKU-YuanGroup/LLaVA-CoT.

Yunpeng Bai, Zihan Zhong, Chao Dong et al.

The recent GAN inversion methods have been able to successfully invert the real image input to the corresponding editable latent code in StyleGAN. By combining with the language-vision model (CLIP), some text-driven image manipulation methods are proposed. However, these methods require extra costs to perform optimization for a certain image or a new attribute editing mode. To achieve a more efficient editing method, we propose a new Text-driven image Manipulation framework via Space Alignment (TMSA). The Space Alignment module aims to align the same semantic regions in CLIP and StyleGAN spaces. Then, the text input can be directly accessed into the StyleGAN space and be used to find the semantic shift according to the text description. The framework can support arbitrary image editing mode without additional cost. Our work provides the user with an interface to control the attributes of a given image according to text input and get the result in real time. Ex tensive experiments demonstrate our superior performance over prior works.

Shiting Xu, Guowei Xu, Peilei Jia et al.

Task requirements (TRs) writing is an important question type in Key English Test and Preliminary English Test. A TR writing question may include multiple requirements and a high-quality essay must respond to each requirement thoroughly and accurately. However, the limited teacher resources prevent students from getting detailed grading instantly. The majority of existing automatic essay scoring systems focus on giving a holistic score but rarely provide reasons to support it. In this paper, we proposed an end-to-end framework based on machine reading comprehension (MRC) to address this problem to some extent. The framework not only detects whether an essay responds to a requirement question, but clearly marks where the essay answers the question. Our framework consists of three modules: question normalization module, ELECTRA based MRC module and response locating module. We extensively explore state-of-the-art MRC methods. Our approach achieves 0.93 accuracy score and 0.85 F1 score on a real-world educational dataset. To encourage reproducible results, we make our code publicly available at \url{https://github.com/aied2021TRMRC/AIED_2021_TRMRC_code}.

Guowei Xu, Wenbiao Ding, Weiping Fu et al.

Many real-world applications involve the use of Optical Character Recognition (OCR) engines to transform handwritten images into transcripts on which downstream Natural Language Processing (NLP) models are applied. In this process, OCR engines may introduce errors and inputs to downstream NLP models become noisy. Despite that pre-trained models achieve state-of-the-art performance in many NLP benchmarks, we prove that they are not robust to noisy texts generated by real OCR engines. This greatly limits the application of NLP models in real-world scenarios. In order to improve model performance on noisy OCR transcripts, it is natural to train the NLP model on labelled noisy texts. However, in most cases there are only labelled clean texts. Since there is no handwritten pictures corresponding to the text, it is impossible to directly use the recognition model to obtain noisy labelled data. Human resources can be employed to copy texts and take pictures, but it is extremely expensive considering the size of data for model training. Consequently, we are interested in making NLP models intrinsically robust to OCR errors in a low resource manner. We propose a novel robust training framework which 1) employs simple but effective methods to directly simulate natural OCR noises from clean texts and 2) iteratively mines the hard examples from a large number of simulated samples for optimal performance. 3) To make our model learn noise-invariant representations, a stability loss is employed. Experiments on three real-world datasets show that the proposed framework boosts the robustness of pre-trained models by a large margin. We believe that this work can greatly promote the application of NLP models in actual scenarios, although the algorithm we use is simple and straightforward. We make our codes and three datasets publicly available\footnote{https://github.com/tal-ai/Robust-learning-MSSHEM}.

Suning Huang, Zheyu Zhang, Tianhai Liang et al. · tsinghua

Visual deep reinforcement learning (RL) enables robots to acquire skills from visual input for unstructured tasks. However, current algorithms suffer from low sample efficiency, limiting their practical applicability. In this work, we present MENTOR, a method that improves both the architecture and optimization of RL agents. Specifically, MENTOR replaces the standard multi-layer perceptron (MLP) with a mixture-of-experts (MoE) backbone and introduces a task-oriented perturbation mechanism. MENTOR outperforms state-of-the-art methods across three simulation benchmarks and achieves an average of 83% success rate on three challenging real-world robotic manipulation tasks, significantly surpassing the 32% success rate of the strongest existing model-free visual RL algorithm. These results underscore the importance of sample efficiency in advancing visual RL for real-world robotics. Experimental videos are available at https://suninghuang19.github.io/mentor_page/.

Tianying Ji, Yongyuan Liang, Yan Zeng et al. · tsinghua

The varying significance of distinct primitive behaviors during the policy learning process has been overlooked by prior model-free RL algorithms. Leveraging this insight, we explore the causal relationship between different action dimensions and rewards to evaluate the significance of various primitive behaviors during training. We introduce a causality-aware entropy term that effectively identifies and prioritizes actions with high potential impacts for efficient exploration. Furthermore, to prevent excessive focus on specific primitive behaviors, we analyze the gradient dormancy phenomenon and introduce a dormancy-guided reset mechanism to further enhance the efficacy of our method. Our proposed algorithm, ACE: Off-policy Actor-critic with Causality-aware Entropy regularization, demonstrates a substantial performance advantage across 29 diverse continuous control tasks spanning 7 domains compared to model-free RL baselines, which underscores the effectiveness, versatility, and efficient sample efficiency of our approach. Benchmark results and videos are available at https://ace-rl.github.io/.

Yuhang Yang, Fan Zhang, Huaijin Pi et al.

Digital characters are central to modern media, yet generating character videos with long-duration, consistent multi-view appearance and expressive identity remains challenging. Existing approaches either provide insufficient context to preserve identity or leverage non-character-centric information as the memory, leading to suboptimal consistency. Recognizing that character video generation inherently resembles an outside-looking-in scenario. In this work, we propose representing the character visual attributes through a compact set of anchor frames. This design provides stable references for consistency, while reference-based video generation inherently faces challenges of copy-pasting and multi-reference conflicts. To address these, we introduce two mechanisms: Superset Content Anchoring, providing intra- and extra-training clip cues to prevent duplication, and RoPE as Weak Condition, encoding positional offsets to distinguish multiple anchors. Furthermore, we construct a scalable pipeline to extract these anchors from massive videos. Experiments show our method generates high-quality character videos exceeding 10 minutes, and achieves expressive identity and appearance consistency across views, surpassing existing methods.

Zixi Liang, Guowei Xu, Haifeng Wu et al.

Learning-based methods have become increasingly popular in 3D indoor scene synthesis (ISS), showing superior performance over traditional optimization-based approaches. These learning-based methods typically model distributions on simple yet explicit scene representations using generative models. However, due to the oversimplified explicit representations that overlook detailed information and the lack of guidance from multimodal relationships within the scene, most learning-based methods struggle to generate indoor scenes with realistic object arrangements and styles. In this paper, we introduce a new method, Scene Implicit Neural Field (S-INF), for indoor scene synthesis, aiming to learn meaningful representations of multimodal relationships, to enhance the realism of indoor scene synthesis. S-INF assumes that the scene layout is often related to the object-detailed information. It disentangles the multimodal relationships into scene layout relationships and detailed object relationships, fusing them later through implicit neural fields (INFs). By learning specialized scene layout relationships and projecting them into S-INF, we achieve a realistic generation of scene layout. Additionally, S-INF captures dense and detailed object relationships through differentiable rendering, ensuring stylistic consistency across objects. Through extensive experiments on the benchmark 3D-FRONT dataset, we demonstrate that our method consistently achieves state-of-the-art performance under different types of ISS.

Guowei Xu, Mert Yuksekgonul, James Zou

In this paper, we identify a sparse reward subsystem within the hidden states of Large Language Models (LLMs), drawing an analogy to the biological reward subsystem in the human brain. We demonstrate that this subsystem contains value neurons that represent the model's internal expectation of state value, and through intervention experiments, we establish the importance of these neurons for reasoning. Our experiments reveal that these value neurons are robust across diverse datasets, model scales, and architectures; furthermore, they exhibit significant transferability across different datasets and models fine-tuned from the same base model. By examining cases where value predictions and actual rewards diverge, we identify dopamine neurons within the reward subsystem which encode reward prediction errors (RPE). These neurons exhibit high activation when the reward is higher than expected and low activation when the reward is lower than expected.

Guowei Xu, Yuxuan Bian, Ailing Zeng et al.

This paper introduces OmniMotion-X, a versatile multimodal framework for whole-body human motion generation, leveraging an autoregressive diffusion transformer in a unified sequence-to-sequence manner. OmniMotion-X efficiently supports diverse multimodal tasks, including text-to-motion, music-to-dance, speech-to-gesture, and global spatial-temporal control scenarios (e.g., motion prediction, in-betweening, completion, and joint/trajectory-guided synthesis), as well as flexible combinations of these tasks. Specifically, we propose the use of reference motion as a novel conditioning signal, substantially enhancing the consistency of generated content, style, and temporal dynamics crucial for realistic animations. To handle multimodal conflicts, we introduce a progressive weak-to-strong mixed-condition training strategy. To enable high-quality multimodal training, we construct OmniMoCap-X, the largest unified multimodal motion dataset to date, integrating 28 publicly available MoCap sources across 10 distinct tasks, standardized to the SMPL-X format at 30 fps. To ensure detailed and consistent annotations, we render sequences into videos and use GPT-4o to automatically generate structured and hierarchical captions, capturing both low-level actions and high-level semantics. Extensive experimental evaluations confirm that OmniMotion-X significantly surpasses existing methods, demonstrating state-of-the-art performance across multiple multimodal tasks and enabling the interactive generation of realistic, coherent, and controllable long-duration motions.

Guowei Xu, Wenxin Xu, Jiawang Zhao et al. · tsinghua

Diffusion models have recently shown strong potential in language modeling, offering faster generation compared to traditional autoregressive approaches. However, applying supervised fine-tuning (SFT) to diffusion models remains challenging, as they lack precise probability estimates at each denoising step. While the diffusion mechanism enables the model to reason over entire sequences, it also makes the generation process less predictable and often inconsistent. This highlights the importance of controlling key tokens that guide the direction of generation. To address this issue, we propose WeFT, a weighted SFT method for diffusion language models, where tokens are assigned different weights based on their entropy. Derived from diffusion theory, WeFT delivers substantial gains: training on s1K, s1K-1.1, and 3k samples from open-r1, it achieves relative improvements of 39%, 64%, and 83% over standard SFT on four widely used reasoning benchmarks (Sudoku, Countdown, GSM8K, and MATH-500). The code and models will be made publicly available.

Guowei Xu, Mert Yuksekgonul, Carlos Guestrin et al. · tsinghua

Large language models (LLMs) are increasingly used in learning algorithms, evaluations, and optimization tasks. Recent studies have shown that using LLM-based optimizers to automatically optimize model prompts, demonstrations, predictions themselves, or other components can significantly enhance the performance of AI systems, as demonstrated by frameworks such as DSPy and TextGrad. However, optimizers built on language models themselves are usually designed by humans with manual design choices; optimizers themselves are not optimized. Moreover, these optimizers are general purpose by design, to be useful to a broad audience, and are not tailored for specific tasks. To address these challenges, we propose metaTextGrad, which focuses on designing a meta-optimizer to further enhance existing optimizers and align them to be good optimizers for a given task. Our approach consists of two key components: a meta prompt optimizer and a meta structure optimizer. The combination of these two significantly improves performance across multiple benchmarks, achieving an average absolute performance improvement of up to 6% compared to the best baseline.

Wentao Wang, Guowei Xu, Wenbiao Ding et al.

Representation learning has been proven to play an important role in the unprecedented success of machine learning models in numerous tasks, such as machine translation, face recognition and recommendation. The majority of existing representation learning approaches often require a large number of consistent and noise-free labels. However, due to various reasons such as budget constraints and privacy concerns, labels are very limited in many real-world scenarios. Directly applying standard representation learning approaches on small labeled data sets will easily run into over-fitting problems and lead to sub-optimal solutions. Even worse, in some domains such as education, the limited labels are usually annotated by multiple workers with diverse expertise, which yields noises and inconsistency in such crowdsourcing settings. In this paper, we propose a novel framework which aims to learn effective representations from limited data with crowdsourced labels. Specifically, we design a grouping based deep neural network to learn embeddings from a limited number of training samples and present a Bayesian confidence estimator to capture the inconsistency among crowdsourced labels. Furthermore, to expedite the training process, we develop a hard example selection procedure to adaptively pick up training examples that are misclassified by the model. Extensive experiments conducted on three real-world data sets demonstrate the superiority of our framework on learning representations from limited data with crowdsourced labels, comparing with various state-of-the-art baselines. In addition, we provide a comprehensive analysis on each of the main components of our proposed framework and also introduce the promising results it achieved in our real production to fully understand the proposed framework.

Wenbiao Ding, Guowei Xu, Tianqiao Liu et al.

Spoken language proficiency is critically important for children's growth and personal development. Due to the limited and imbalanced educational resources in China, elementary students barely have chances to improve their oral language skills in classes. Verbal fluency tasks (VFTs) were invented to let the students practice their spoken language proficiency after school. VFTs are simple but concrete math related questions that ask students to not only report answers but speak out the entire thinking process. In spite of the great success of VFTs, they bring a heavy grading burden to elementary teachers. To alleviate this problem, we develop Dolphin, a spoken language proficiency assessment system for Chinese elementary education. Dolphin is able to automatically evaluate both phonological fluency and semantic relevance of students' VFT answers. We conduct a wide range of offline and online experiments to demonstrate the effectiveness of Dolphin. In our offline experiments, we show that Dolphin improves both phonological fluency and semantic relevance evaluation performance when compared to state-of-the-art baselines on real-world educational data sets. In our online A/B experiments, we test Dolphin with 183 teachers from 2 major cities (Hangzhou and Xi'an) in China for 10 weeks and the results show that VFT assignments grading coverage is improved by 22\%.

Guowei Xu, Wenbiao Ding, Jiliang Tang et al.

Learning representation has been proven to be helpful in numerous machine learning tasks. The success of the majority of existing representation learning approaches often requires a large amount of consistent and noise-free labels. However, labels are not accessible in many real-world scenarios and they are usually annotated by the crowds. In practice, the crowdsourced labels are usually inconsistent among crowd workers given their diverse expertise and the number of crowdsourced labels is very limited. Thus, directly adopting crowdsourced labels for existing representation learning algorithms is inappropriate and suboptimal. In this paper, we investigate the above problem and propose a novel framework of \textbf{R}epresentation \textbf{L}earning with crowdsourced \textbf{L}abels, i.e., "RLL", which learns representation of data with crowdsourced labels by jointly and coherently solving the challenges introduced by limited and inconsistent labels. The proposed representation learning framework is evaluated in two real-world education applications. The experimental results demonstrate the benefits of our approach on learning representation from limited labeled data from the crowds, and show RLL is able to outperform state-of-the-art baselines. Moreover, detailed experiments are conducted on RLL to fully understand its key components and the corresponding performance.