METEOR: Evolutionary Journey of Large Language Models from Guidance to Self-Growth

Model evolution enables learning from feedback to refine experiences and update skills, transforming models from having no domain knowledge to becoming domain experts. However, there is currently no unified and effective method for guiding this evolutionary process. To address this gap, we propose the Meteor method, which includes three training phases: weak-to-strong data distillation, iterative training, and self-evolution strategies. Each phase maximizes the model's inherent domain capabilities, allowing it to autonomously refine its domain knowledge and enhance performance. Experiments demonstrate that our approach significantly improves accuracy, completeness, relevance, coherence, and reliability across domain-specific tasks.

Reviving Dormant Memories: Investigating Catastrophic Forgetting in Language Models through Rationale-Guidance Difficulty

Although substantial efforts have been made to mitigate catastrophic forgetting in continual learning, the intrinsic mechanisms are not well understood. In this paper, we discover that when a forgetting model passively receives an externally provided partial appropriate rationale, its performance on the forgotten task can be restored. Furthermore, by simply adding a task-agnostic prefix to the original instruction, the forgetting model can actively generate an appropriate rationale to reach the correct answer. These findings suggest that the model does not actually ``forget'' the task knowledge; instead, the degraded performance can be attributed to the failure of the original instructions in guiding the model to generate the appropriate rationales. Based on this insight, we propose the Rationale-Guidance Difficulty metric to evaluate how effectively a given instruction guides the model in generating appropriate rationales. We apply this metric to optimize the allocation of replay data in replay-based continual learning algorithm. Experimental results demonstrate that our data allocation method effectively mitigates catastrophic forgetting and maintains better model plasticity simultaneously across models.

PSPO*: An Effective Process-supervised Policy Optimization for Reasoning Alignment

Process supervision enhances the performance of large language models in reasoning tasks by providing feedback at each step of chain-of-thought reasoning. However, due to the lack of effective process supervision methods, even advanced large language models are prone to logical errors and redundant reasoning. We claim that the effectiveness of process supervision significantly depends on both the accuracy and the length of reasoning chains. Moreover, we identify that these factors exhibit a nonlinear relationship with the overall reward score of the reasoning process. Inspired by these insights, we propose a novel process supervision paradigm, PSPO*, which systematically outlines the workflow from reward model training to policy optimization, and highlights the importance of nonlinear rewards in process supervision. Based on PSPO*, we develop the PSPO-WRS, which considers the number of reasoning steps in determining reward scores and utilizes an adjusted Weibull distribution for nonlinear reward shaping. Experimental results on six mathematical reasoning datasets demonstrate that PSPO-WRS consistently outperforms current mainstream models.

UniHOI: Learning Fast, Dense and Generalizable 4D Reconstruction for Egocentric Hand Object Interaction Videos

Egocentric Hand Object Interaction (HOI) videos provide valuable insights into human interactions with the physical world, attracting growing interest from the computer vision and robotics communities. A key task in fully understanding the geometry and dynamics of HOI scenes is dense pointclouds sequence reconstruction. However, the inherent motion of both hands and the camera makes this challenging. Current methods often rely on time-consuming test-time optimization, making them impractical for reconstructing internet-scale videos. To address this, we introduce UniHOI, a model that unifies the estimation of all variables necessary for dense 4D reconstruction, including camera intrinsic, camera poses, and video depth, for egocentric HOI scene in a fast feed-forward manner. We end-to-end optimize all these variables to improve their consistency in 3D space. Furthermore, our model could be trained solely on large-scale monocular video dataset, overcoming the limitation of scarce labeled HOI data. We evaluate UniHOI with both in-domain and zero-shot generalization setting, surpassing all baselines in pointclouds sequence reconstruction and long-term 3D scene flow recovery. UniHOI is the first approach to offer fast, dense, and generalizable monocular egocentric HOI scene reconstruction in the presence of motion. Code and trained model will be released in the future.

Multi-Transmotion: Pre-trained Model for Human Motion Prediction

The ability of intelligent systems to predict human behaviors is crucial, particularly in fields such as autonomous vehicle navigation and social robotics. However, the complexity of human motion have prevented the development of a standardized dataset for human motion prediction, thereby hindering the establishment of pre-trained models. In this paper, we address these limitations by integrating multiple datasets, encompassing both trajectory and 3D pose keypoints, to propose a pre-trained model for human motion prediction. We merge seven distinct datasets across varying modalities and standardize their formats. To facilitate multimodal pre-training, we introduce Multi-Transmotion, an innovative transformer-based model designed for cross-modality pre-training. Additionally, we present a novel masking strategy to capture rich representations. Our methodology demonstrates competitive performance across various datasets on several downstream tasks, including trajectory prediction in the NBA and JTA datasets, as well as pose prediction in the AMASS and 3DPW datasets. The code is publicly available: https://github.com/vita-epfl/multi-transmotion

Data Scaling Laws in Imitation Learning for Robotic Manipulation

Data scaling has revolutionized fields like natural language processing and computer vision, providing models with remarkable generalization capabilities. In this paper, we investigate whether similar data scaling laws exist in robotics, particularly in robotic manipulation, and whether appropriate data scaling can yield single-task robot policies that can be deployed zero-shot for any object within the same category in any environment. To this end, we conduct a comprehensive empirical study on data scaling in imitation learning. By collecting data across numerous environments and objects, we study how a policy's generalization performance changes with the number of training environments, objects, and demonstrations. Throughout our research, we collect over 40,000 demonstrations and execute more than 15,000 real-world robot rollouts under a rigorous evaluation protocol. Our findings reveal several intriguing results: the generalization performance of the policy follows a roughly power-law relationship with the number of environments and objects. The diversity of environments and objects is far more important than the absolute number of demonstrations; once the number of demonstrations per environment or object reaches a certain threshold, additional demonstrations have minimal effect. Based on these insights, we propose an efficient data collection strategy. With four data collectors working for one afternoon, we collect sufficient data to enable the policies for two tasks to achieve approximately 90% success rates in novel environments with unseen objects.

START: A Generalized State Space Model with Saliency-Driven Token-Aware Transformation

Domain Generalization (DG) aims to enable models to generalize to unseen target domains by learning from multiple source domains. Existing DG methods primarily rely on convolutional neural networks (CNNs), which inherently learn texture biases due to their limited receptive fields, making them prone to overfitting source domains. While some works have introduced transformer-based methods (ViTs) for DG to leverage the global receptive field, these methods incur high computational costs due to the quadratic complexity of self-attention. Recently, advanced state space models (SSMs), represented by Mamba, have shown promising results in supervised learning tasks by achieving linear complexity in sequence length during training and fast RNN-like computation during inference. Inspired by this, we investigate the generalization ability of the Mamba model under domain shifts and find that input-dependent matrices within SSMs could accumulate and amplify domain-specific features, thus hindering model generalization. To address this issue, we propose a novel SSM-based architecture with saliency-based token-aware transformation (namely START), which achieves state-of-the-art (SOTA) performances and offers a competitive alternative to CNNs and ViTs. Our START can selectively perturb and suppress domain-specific features in salient tokens within the input-dependent matrices of SSMs, thus effectively reducing the discrepancy between different domains. Extensive experiments on five benchmarks demonstrate that START outperforms existing SOTA DG methods with efficient linear complexity. Our code is available at https://github.com/lingeringlight/START.

GravMAD: Grounded Spatial Value Maps Guided Action Diffusion for Generalized 3D Manipulation

Robots' ability to follow language instructions and execute diverse 3D tasks is vital in robot learning. Traditional imitation learning-based methods perform well on seen tasks but struggle with novel, unseen ones due to variability. Recent approaches leverage large foundation models to assist in understanding novel tasks, thereby mitigating this issue. However, these methods lack a task-specific learning process, which is essential for an accurate understanding of 3D environments, often leading to execution failures. In this paper, we introduce GravMAD, a sub-goal-driven, language-conditioned action diffusion framework that combines the strengths of imitation learning and foundation models. Our approach breaks tasks into sub-goals based on language instructions, allowing auxiliary guidance during both training and inference. During training, we introduce Sub-goal Keypose Discovery to identify key sub-goals from demonstrations. Inference differs from training, as there are no demonstrations available, so we use pre-trained foundation models to bridge the gap and identify sub-goals for the current task. In both phases, GravMaps are generated from sub-goals, providing flexible 3D spatial guidance compared to fixed 3D positions. Empirical evaluations on RLBench show that GravMAD significantly outperforms state-of-the-art methods, with a 28.63% improvement on novel tasks and a 13.36% gain on tasks encountered during training. These results demonstrate GravMAD's strong multi-task learning and generalization in 3D manipulation. Video demonstrations are available at: https://gravmad.github.io.

3D Gaussian Splatting: Survey, Technologies, Challenges, and Opportunities

3D Gaussian Splatting (3DGS) has emerged as a prominent technique with the potential to become a mainstream method for 3D representations. It can effectively transform multi-view images into explicit 3D Gaussian representations through efficient training, and achieve real-time rendering of novel views. This survey aims to analyze existing 3DGS-related works from multiple intersecting perspectives, including related tasks, technologies, challenges, and opportunities. The primary objective is to provide newcomers with a rapid understanding of the field and to assist researchers in methodically organizing existing technologies and challenges. Specifically, we delve into the optimization, application, and extension of 3DGS, categorizing them based on their focuses or motivations. Additionally, we summarize and classify nine types of technical modules and corresponding improvements identified in existing works. Based on these analyses, we further examine the common challenges and technologies across various tasks, proposing potential research opportunities.

Learn to Preserve and Diversify: Parameter-Efficient Group with Orthogonal Regularization for Domain Generalization

Domain generalization (DG) aims to avoid the performance degradation of the model when the distribution shift between the limited training data and unseen test data occurs. Recently, foundation models with enormous parameters have been pre-trained with huge datasets, demonstrating strong generalization ability and showing promising direction for solving the DG problem. However, fully Fine-Tuning (FT) the foundation models results in unsatisfactory out-of-distribution accuracy due to the destroyed pre-trained generalized features. Recently, Parameter-Efficient Fine-Tuning (PEFT) alleviates the above problem by fine-tuning a small portion of the model parameters while keeping the rest frozen, which achieves better generalization performance compared to FT. Nevertheless, PEFT still suffers from the issue of overfitting to the training domains. To address the above issue, we propose Parameter-Efficient Group with Orthogonal regularization (PEGO) for vision transformers, which effectively preserves the generalization ability of the pre-trained network and learns more diverse knowledge compared with conventional PEFT. Specifically, we inject a group of trainable Low-Rank Adaptation (LoRA) modules into the pre-trained model and propose an orthogonal regularization loss to enhance the generalization ability of the model. Our framework achieves SOTA performance on five DG benchmarks, while only requiring training a small number of parameters without adding additional testing cost.