Creating detailed 3D human avatars with fitted garments traditionally requires specialized expertise and labor-intensive workflows. While recent advances in generative AI have enabled text-to-3D human and clothing synthesis, existing methods fall short in offering accessible, integrated pipelines for generating CG-ready 3D avatars with physically compatible outfits; here we use the term CG-ready for models following a technical aesthetic common in computer graphics (CG) and adopt standard CG polygonal meshes and strands representations (rather than radiance field representations like NeRF and 3DGS) that can be directly integrated into conventional CG pipelines and support downstream tasks such as physical simulation. To bridge this gap, we introduce Tailor, an integrated text-to-3D framework that generates high-fidelity, customizable 3D avatars dressed in simulation-ready garments. Tailor consists of three stages. (1) Semantic Parsing: we employ a large language model to interpret textual descriptions and translate them into parameterized human avatars and semantically matched garment templates. (2) Geometry-Aware Garment Generation: we propose topology-preserving deformation with novel geometric losses to generate body-aligned garments under text control. (3) Consistent Texture Synthesis: we propose a novel multi-view diffusion process optimized for garment texturing, which enforces view consistency, preserves photorealistic details, and optionally supports symmetric texture generation common in garments. Through comprehensive quantitative and qualitative evaluations, we demonstrate that Tailor outperforms state-of-the-art methods in fidelity, usability, and diversity.

Recent advances in large language model assistants have made them indispensable, raising significant concerns over managing their safety. Automated red teaming offers a promising alternative to the labor-intensive and error-prone manual probing for vulnerabilities, providing more consistent and scalable safety evaluations. However, existing approaches often compromise diversity by focusing on maximizing attack success rate. Additionally, methods that decrease the cosine similarity from historical embeddings with semantic diversity rewards lead to novelty stagnation as history grows. To address these issues, we introduce DiveR-CT, which relaxes conventional constraints on the objective and semantic reward, granting greater freedom for the policy to enhance diversity. Our experiments demonstrate DiveR-CT's marked superiority over baselines by 1) generating data that perform better in various diversity metrics across different attack success rate levels, 2) better-enhancing resiliency in blue team models through safety tuning based on collected data, 3) allowing dynamic control of objective weights for reliable and controllable attack success rates, and 4) reducing susceptibility to reward overoptimization. Overall, our method provides an effective and efficient approach to LLM red teaming, accelerating real-world deployment. WARNING: This paper contains examples of potentially harmful text.

Reinforcement learning with verifiable rewards (RLVR) has shown promise in enhancing the reasoning capabilities of large language models by learning directly from rule-based outcome rewards. Recent RLVR works that operate under the zero setting avoid supervision in labeling the reasoning process, but still depend on manually curated collections of questions and answers for training. The scarcity of high-quality, human-produced examples raises concerns about the long-term scalability of relying on human supervision, a challenge already evident in the domain of language model pretraining. Furthermore, in a hypothetical future where AI surpasses human intelligence, tasks provided by humans may offer limited learning potential for a superintelligent system. To address these concerns, we propose a new RLVR paradigm called Absolute Zero, in which a single model learns to propose tasks that maximize its own learning progress and improves reasoning by solving them, without relying on any external human or distillation data. Under this paradigm, we introduce the Absolute Zero Reasoner (AZR), a system that self-evolves its training curriculum and reasoning ability. AZR uses a code executor to both validate self-proposed code reasoning tasks and verify answers, serving as an unified source of verifiable feedback to guide open-ended yet grounded learning. Despite being trained entirely without external data, AZR achieves overall SOTA performance on coding and mathematical reasoning tasks, outperforming existing zero-setting models that rely on tens of thousands of in-domain human-curated examples. Furthermore, we demonstrate that AZR can be effectively applied across different model scales and is compatible with various model classes.

Some robust point cloud registration approaches with controllable pose refinement magnitude, such as ICP and its variants, are commonly used to improve 6D pose estimation accuracy. However, the effectiveness of these methods gradually diminishes with the advancement of deep learning techniques and the enhancement of initial pose accuracy, primarily due to their lack of specific design for pose refinement. In this paper, we propose Point Cloud Completion and Keypoint Refinement with Fusion Data (PCKRF), a new pose refinement pipeline for 6D pose estimation. The pipeline consists of two steps. First, it completes the input point clouds via a novel pose-sensitive point completion network. The network uses both local and global features with pose information during point completion. Then, it registers the completed object point cloud with the corresponding target point cloud by our proposed Color supported Iterative KeyPoint (CIKP) method. The CIKP method introduces color information into registration and registers a point cloud around each keypoint to increase stability. The PCKRF pipeline can be integrated with existing popular 6D pose estimation methods, such as the full flow bidirectional fusion network, to further improve their pose estimation accuracy. Experiments demonstrate that our method exhibits superior stability compared to existing approaches when optimizing initial poses with relatively high precision. Notably, the results indicate that our method effectively complements most existing pose estimation techniques, leading to improved performance in most cases. Furthermore, our method achieves promising results even in challenging scenarios involving textureless and symmetrical objects.

The reconstruction of indoor scenes from multi-view RGB images is challenging due to the coexistence of flat and texture-less regions alongside delicate and fine-grained regions. Recent methods leverage neural radiance fields aided by predicted surface normal priors to recover the scene geometry. These methods excel in producing complete and smooth results for floor and wall areas. However, they struggle to capture complex surfaces with high-frequency structures due to the inadequate neural representation and the inaccurately predicted normal priors. This work aims to reconstruct high-fidelity surfaces with fine-grained details by addressing the above limitations. To improve the capacity of the implicit representation, we propose a hybrid architecture to represent low-frequency and high-frequency regions separately. To enhance the normal priors, we introduce a simple yet effective image sharpening and denoising technique, coupled with a network that estimates the pixel-wise uncertainty of the predicted surface normal vectors. Identifying such uncertainty can prevent our model from being misled by unreliable surface normal supervisions that hinder the accurate reconstruction of intricate geometries. Experiments on the benchmark datasets show that our method outperforms existing methods in terms of reconstruction quality. Furthermore, the proposed method also generalizes well to real-world indoor scenarios captured by our hand-held mobile phones.

Occlusion is a common issue in 3D reconstruction from RGB-D videos, often blocking the complete reconstruction of objects and presenting an ongoing problem. In this paper, we propose a novel framework, empowered by a 2D diffusion-based in-painting model, to reconstruct complete surfaces for the hidden parts of objects. Specifically, we utilize a pre-trained diffusion model to fill in the hidden areas of 2D images. Then we use these in-painted images to optimize a neural implicit surface representation for each instance for 3D reconstruction. Since creating the in-painting masks needed for this process is tricky, we adopt a human-in-the-loop strategy that involves very little human engagement to generate high-quality masks. Moreover, some parts of objects can be totally hidden because the videos are usually shot from limited perspectives. To ensure recovering these invisible areas, we develop a cascaded network architecture for predicting signed distance field, making use of different frequency bands of positional encoding and maintaining overall smoothness. Besides the commonly used rendering loss, Eikonal loss, and silhouette loss, we adopt a CLIP-based semantic consistency loss to guide the surface from unseen camera angles. Experiments on ScanNet scenes show that our proposed framework achieves state-of-the-art accuracy and completeness in object-level reconstruction from scene-level RGB-D videos. Code: https://github.com/THU-LYJ-Lab/O2-Recon.

This paper tackles the problem of efficient and stable video semantic segmentation. While stability has been under-explored, prevalent work in efficient video semantic segmentation uses the keyframe paradigm. They efficiently process videos by only recomputing the low-level features and reusing high-level features computed at selected keyframes. In addition, the reused features stabilize the predictions across frames, thereby improving video consistency. However, dynamic scenes in the video can easily lead to misalignments between reused and recomputed features, which hampers performance. Moreover, relying on feature reuse to improve prediction consistency is brittle; an erroneous alignment of the features can easily lead to unstable predictions. Therefore, the keyframe paradigm exhibits a dilemma between stability and performance. We address this efficiency and stability challenge using a novel yet simple Temporal Feature Correlation (TFC) module. It uses the cosine similarity between two frames’ low-level features to inform the semantic label’s consistency across frames. Specifically, we selectively reuse label-consistent features across frames through linear interpolation and update others through sparse multi-scale deformable attention. As a result, we no longer directly reuse features to improve stability and thus effectively solve feature misalignment. This work provides a significant step towards efficient and stable video semantic segmentation. On the VSPW dataset, our method significantly improves the prediction consistency of image-based methods while being as fast and accurate.

The recent surge in research interest in applying large language models (LLMs) to decision-making tasks has flourished by leveraging the extensive world knowledge embedded in LLMs. While there is a growing demand to tailor LLMs for custom decision-making tasks, finetuning them for specific tasks is resource-intensive and may diminish the model's generalization capabilities. Moreover, state-of-the-art language models like GPT-4 and Claude are primarily accessible through API calls, with their parametric weights remaining proprietary and unavailable to the public. This scenario emphasizes the growing need for new methodologies that allow learning from agent experiences without requiring parametric updates. To address these problems, we introduce the Experiential Learning (ExpeL) agent. Our agent autonomously gathers experiences and extracts knowledge using natural language from a collection of training tasks. At inference, the agent recalls its extracted insights and past experiences to make informed decisions. Our empirical results highlight the robust learning efficacy of the ExpeL agent, indicating a consistent enhancement in its performance as it accumulates experiences. We further explore the emerging capabilities and transfer learning potential of the ExpeL agent through qualitative observations and additional experiments.

Depth estimation is a crucial task in computer vision, applicable to various domains such as 3D reconstruction, robotics, and autonomous driving. In particular, thermal-based depth estimation has unique advantages, including night-time vision. However, the existing depth estimation method remains challenging in robust generalization due to limited data resources and spectral differences between thermal and RGB images. In this paper, we present a self-supervised approach to enhance thermal-based depth estimation by leveraging pre-trained visual models initially designed for RGB data. In detail, we design a novel two-stage training strategy, incorporating Low-rank Adapters and Convolutional Adapters, which not only significantly improves accuracy and robustness but also enables impressive zero-shot generalization capabilities. Our method outperforms existing thermal-based depth estimation models, opening new possibilities for cross-modal applications in computer vision and robotics research.

Few-shot object detection (FSOD) aims to detect novel targets with only a few instances of the associated samples. Although combinations of distillation techniques and meta-learning paradigms have been acknowledged as the primary strategies for FSOD tasks, the existing distillation methods exhibit inherent biases and sensitivity to novel class variability. A critical hurdle for FSOD distillation is the difficulty in ensuring appropriate knowledge learned from the teacher model during the fine-tuning stage. Furthermore, coarse distillation procedures risk misalignment between the learned and actual distributions. This misalignment could potentially negate the benefits of positive cases and impede the detector’s evolution. To address these deficiencies, we propose a novel self-distillation paradigm exclusively for the fine-tuning stage (SD-FSOD). Our methods integrate a Distribution Prototype Extractor (DPE) and Self-Distillation Memory (SDM), promoting feature distribution consistency during distillation. In detail, the DPE module reliably initializes the weights of the detector, ensuring a robust class distribution for the distillation process. Meanwhile, the SDM module utilizes decoupling techniques to divide the distillation tasks into two sub-task branches, allowing the student model to independently learn and share precise features through isolated distillation processes. The synergistic integration of feature calibration techniques and the continuous self-distillation paradigm distinctly enhances the fine-tuning process, which shows the superiority of the FSOD self-distillation methodologies. The extensive experiments on the PASCAL VOC and MS COCO datasets demonstrate that our proposed approach produces significant improvements and achieves state-of-the-art (SOTA) performance.

The generation of stylistic 3D facial animations driven by speech presents a significant challenge as it requires learning a many-to-many mapping between speech, style, and the corresponding natural facial motion. However, existing methods either employ a deterministic model for speech-to-motion mapping or encode the style using a one-hot encoding scheme. Notably, the one-hot encoding approach fails to capture the complexity of the style and thus limits generalization ability. In this paper, we propose DiffPoseTalk, a generative framework based on the diffusion model combined with a style encoder that extracts style embeddings from short reference videos. During inference, we employ classifier-free guidance to guide the generation process based on the speech and style. In particular, our style includes the generation of head poses, thereby enhancing user perception. Additionally, we address the shortage of scanned 3D talking face data by training our model on reconstructed 3DMM parameters from a high-quality, in-the-wild audio-visual dataset. Extensive experiments and user study demonstrate that our approach outperforms state-of-the-art methods. The code and dataset are at https://diffposetalk.github.io.

Neural implicit representations have revolutionized dense multi-view surface reconstruction, yet their performance significantly diminishes with sparse input views. A few pioneering works have sought to tackle this challenge by leveraging additional geometric priors or multi-scene generalizability. However, they are still hindered by the imperfect choice of input views, using images under empirically determined viewpoints. We propose PVP-Recon , a novel and effective sparse-view surface reconstruction method that progressively plans the next best views to form an optimal set of sparse viewpoints for image capturing. PVP-Recon starts initial surface reconstruction with as few as 3 views and progressively adds new views which are determined based on a novel warping score that reflects the information gain of each newly added view. This progressive view planning progress is interleaved with a neural SDF-based reconstruction module that utilizes multi-resolution hash features, enhanced by a progressive training scheme and a directional Hessian loss. Quantitative and qualitative experiments on three benchmark datasets show that our system achieves high-quality reconstruction with a constrained input budget and outperforms existing baselines.

Training practical agents usually involve offline and online reinforcement learning (RL) to balance the policy's performance and interaction costs. In particular, online fine-tuning has become a commonly used method to correct the erroneous estimates of out-of-distribution data learned in the offline training phase. However, even limited online interactions can be inaccessible or catastrophic for high-stake scenarios like healthcare and autonomous driving. In this work, we introduce an interaction-free training scheme dubbed Offline-with-Action-Preferences (OAP). The main insight is that, compared to online fine-tuning, querying the preferences between pre-collected and learned actions can be equally or even more helpful to the erroneous estimate problem. By adaptively encouraging or suppressing policy constraint according to action preferences, OAP could distinguish overestimation from beneficial policy improvement and thus attains a more accurate evaluation of unseen data. Theoretically, we prove a lower bound of the behavior policy's performance improvement brought by OAP. Moreover, comprehensive experiments on the D4RL benchmark and state-of-the-art algorithms demonstrate that OAP yields higher (29% on average) scores, especially on challenging AntMaze tasks (98% higher).

Offline-to-online reinforcement learning (RL) is a training paradigm that combines pre-training on a pre-collected dataset with fine-tuning in an online environment. However, the incorporation of online fine-tuning can intensify the well-known distributional shift problem. Existing solutions tackle this problem by imposing a policy constraint on the policy improvement objective in both offline and online learning. They typically advocate a single balance between policy improvement and constraints across diverse data collections. This one-size-fits-all manner may not optimally leverage each collected sample due to the significant variation in data quality across different states. To this end, we introduce Family Offline-to-Online RL (FamO2O), a simple yet effective framework that empowers existing algorithms to determine state-adaptive improvement-constraint balances. FamO2O utilizes a universal model to train a family of policies with different improvement/constraint intensities, and a balance model to select a suitable policy for each state. Theoretically, we prove that state-adaptive balances are necessary for achieving a higher policy performance upper bound. Empirically, extensive experiments show that FamO2O offers a statistically significant improvement over various existing methods, achieving state-of-the-art performance on the D4RL benchmark. Codes are available at https://github.com/LeapLabTHU/FamO2O.

Unsupervised reinforcement learning aims at learning a generalist policy in a reward-free manner for fast adaptation to downstream tasks. Most of the existing methods propose to provide an intrinsic reward based on surprise. Maximizing or minimizing surprise drives the agent to either explore or gain control over its environment. However, both strategies rely on a strong assumption: the entropy of the environment's dynamics is either high or low. This assumption may not always hold in real-world scenarios, where the entropy of the environment's dynamics may be unknown. Hence, choosing between the two objectives is a dilemma. We propose a novel yet simple mixture of policies to address this concern, allowing us to optimize an objective that simultaneously maximizes and minimizes the surprise. Concretely, we train one mixture component whose objective is to maximize the surprise and another whose objective is to minimize the surprise. Hence, our method does not make assumptions about the entropy of the environment's dynamics. We call our method a $\textbf{M}\text{ixture }\textbf{O}\text{f }\textbf{S}\text{urprise}\textbf{S}$ (MOSS) for unsupervised reinforcement learning. Experimental results show that our simple method achieves state-of-the-art performance on the URLB benchmark, outperforming previous pure surprise maximization-based objectives. Our code is available at: https://github.com/LeapLabTHU/MOSS.

In environmental perception of autonomous driving, zero-shot semantic segmentation that can make prediction of new categories without using any labeled training samples is considered as a challenging task. One key step in this task is to transfer knowledge across categories via auxiliary semantic word embeddings. In this paper, we propose a feature enhanced projection network (FEPNet) that takes full advantage of transferred knowledge to enrich semantic representations. In FEPNet, two projection layers are added to a segmentation network so as to map features into seen (S) and unseen (U) category spaces, respectively. During training, U-space features are transferred to S-space using similarity relations to enhance the representation of seen categories. In the inference stage, the representation of unseen categories is also strengthened by incorporating features transferred from S-space. Moreover, a novel strategy is proposed to effectively alleviate prediction bias by performing segmentation independently in separate areas that contain seen and unseen categories. We conduct extensive experiments on three benchmark datasets. The experimental results show that our FEPNet achieves new state-of-the-art results compared to existing approaches.