Current benchmarks for evaluating the reasoning capabilities of Large Language Models (LLMs) face significant limitations: task oversimplification, data contamination, and flawed evaluation items. These deficiencies necessitate more rigorous assessment methods. To address these limitations, we introduce PHYBench, a benchmark of 500 original physics problems ranging from high school to Physics Olympiad difficulty. PHYBench addresses data contamination through original content and employs a systematic curation pipeline to eliminate flawed items. Evaluations show that PHYBench activates more tokens and provides stronger differentiation between reasoning models compared to other baselines like AIME 2024, OlympiadBench and GPQA. Even the best-performing model, Gemini 2.5 Pro, achieves only 36.9\% accuracy compared to human experts' 61.9\%. To further enhance evaluation precision, we introduce the Expression Edit Distance (EED) Score for mathematical expression assessment, which improves sample efficiency by 204\% over binary scoring. Moreover, PHYBench effectively elicits multi-step and multi-condition reasoning, providing a platform for examining models' reasoning robustness, preferences, and deficiencies. The benchmark results and dataset are publicly available at https://www.phybench.cn/.

The autonomous database of the next generation aims to apply the reinforcement learning (RL) on tasks like query optimization and performance tuning with little or no human DBAs’ intervention. Despite the promise, to obtain a decent policy model in the domain of database optimization is still challenging — primarily due to the inherent computational overhead involved in the data hungry RL frameworks — in particular on large databases. In the line of mitigating this adverse effect, we propose Mirror in this work. The core to Mirror is a sampling process built in an RL framework together with a transferring process of the policy model from the sampled database to its original counterpart. While being conceptually simple, we identify that the policy transfer between databases involves heavy noise and prediction drifting that cannot be neglectable. Thereby we build a theoretical-guided sampling algorithm in Mirror assisted by a continuous fine-tuning module. The experiments on the PostgreSQL and an industry database PolarDB validate that Mirror has effectively reduced the computational cost while maintaining a satisfactory performance.

We propose a novel end-to-end deep learning framework, capable of 3D human shape reconstruction from a 2D image without the need of a 3D prior parametric model. We employ a “prior-less” representation of the human shape using unordered point clouds. Due to the lack of prior information, comparing the generated and ground truth point clouds to evaluate the reconstruction error is challenging. We solve this problem by proposing an Earth Mover’s Distance (EMD) function to find the optimal mapping between point clouds. Our experimental results show that we are able to obtain a visually accurate estimation of the 3D human shape from a single 2D image, with some inaccuracy for heavily occluded parts.

Many new memory indexing structures have been proposed and outperform current in-memory skip-list index adopted by LevelDB, RocksDB and other key-value systems. However, those new indexes cannot be easily intergrated with key-value systems, because most of them do not consider how the data can be efficiently flushed to disk. Some assumptions, such as fixed size key and value, are unrealistic for real applications. In this paper, we present S3, a scalable in-memory skip-list index for the customized version of RocksDB in Alibaba Cloud. S3 adopts a two-layer structure. In the top layer, a cache-sensitive structure is used to maintain a few guard entries to facilitate the search over the skip-list. In the bottom layer, a semi-ordered skip-list index is built to support highly concurrent insertions and fast lookup and range query. To further improve the performance, we train a neural model to select guard entries intelligently according to the data distribution and query distribution. Experiments on multiple datasets show that S3 achieves a comparable performance to other new memory indexing schemes, and can replace current in-memory skip-list of LevelDB and RocksDB to support huge volume of data.

Human action recognition is crucial to many practical applications, ranging from human-computer interaction to video surveillance. Most approaches either recognize the human action from a fixed view or require the knowledge of view angle, which is usually not available in practical applications. In this paper, we propose a novel end-to-end framework to jointly learn a view-invariance transfer dictionary and a view-invariant classifier. The result of the process is a dictionary that can project real-world 2D video into a view-invariant sparse representation, as well as a classifier to recognize actions with an arbitrary view. The main feature of our algorithm is the use of synthetic data to extract view-invariance between 3D and 2D videos during the pre-training phase. This guarantees the availability of training data, and removes the hassle of obtaining real-world videos in specific viewing angles. Additionally, for better describing the actions in 3D videos, we introduce a new feature set called the 3D dense trajectories to effectively encode extracted trajectory information on 3D videos. Experimental results on the IXMAS, N-UCLA, i3DPost and UWA3DII datasets show improvements over existing algorithms.

Human action recognition is an important problem in robotic vision. Traditional recognition algorithms usually require the knowledge of view angle, which is not always available in robotic applications such as active vision. In this paper, we propose a new framework to recognize actions with arbitrary views. A main feature of our algorithm is that view-invariance is learned from synthetic 2D and 3D training data using transfer dictionary learning. This guarantees the availability of training data, and removes the hassle of obtaining real world video in specific viewing angles. The result of the process is a dictionary that can project real world 2D video into a view-invariant sparse representation. This facilitates the training of a view-invariant classifier. Experimental results on the IXMAS and N-UCLA datasets show significant improvements over existing algorithms.