Actionable and diverse counterfactual explanations incorporating domain knowledge and causal constraints

A new method called SpLap has been introduced for proxy-free deformation of Gaussian splats, utilizing a surface-aware splat graph to enhance the quality of deformations while minimizing computational overhead. This approach overcomes limitations of traditional methods that rely on proxies, which can be of varying quality and add complexity to the deformation process.

Zoo3D has been introduced as the first training-free 3D object detection framework, enabling the construction of 3D bounding boxes through graph clustering of 2D instance masks. This innovative approach allows for the recognition of previously unseen objects without the need for extensive training, marking a significant advancement in 3D object detection technology.

A comprehensive survey on Long Context Language Models (LCLMs) has been published, highlighting the importance of efficiently processing long textual inputs in Natural Language Processing. The survey covers effective strategies for obtaining, training, and evaluating LCLMs, addressing the increasing demand for handling extensive documents and dialogues in various applications.

A new paradigm called Policy World Model (PWM) has been introduced, integrating world modeling and trajectory planning into a unified architecture. This model enhances the planning capabilities of autonomous systems by utilizing learned world knowledge through an action-free future state forecasting scheme, enabling more reliable planning performance through collaborative state-action prediction.

OceanGym has been introduced as the first comprehensive benchmark for underwater embodied agents, aimed at enhancing AI capabilities in challenging oceanic environments characterized by low visibility and dynamic currents. This benchmark includes eight realistic task domains and utilizes Multi-modal Large Language Models (MLLMs) to integrate perception, memory, and decision-making processes.

MXtalTools has been introduced as a flexible Python package designed for data-driven modeling of molecular crystals, enhancing machine learning applications in the molecular solid state. The toolkit includes utilities for dataset curation, model training, crystal parameterization, and high-throughput modeling using CUDA acceleration.

FunReason has been introduced as a novel framework aimed at enhancing the function calling capabilities of large language models (LLMs) through an automated data refinement strategy and a Self-Refinement Multiscale Loss (SRML) approach. This development addresses the challenges of integrating reasoning processes with accurate function execution, which has been a significant hurdle in optimizing LLM performance in real-world applications.

A new model repair module named SafeFix has been introduced to address systematic errors in deep learning models for visual recognition, particularly those stemming from underrepresented semantic subpopulations. This module utilizes a conditional text-to-image model to generate targeted images for failure cases, enhancing the model's performance by ensuring semantic consistency with the original data distribution.