Yuan Lü

Minyang Tian, Luyu Gao, Shizhuo Dylan Zhang et al. · princeton, uw

Since language models (LMs) now outperform average humans on many challenging tasks, it has become increasingly difficult to develop challenging, high-quality, and realistic evaluations. We address this issue by examining LMs' capabilities to generate code for solving real scientific research problems. Incorporating input from scientists and AI researchers in 16 diverse natural science sub-fields, including mathematics, physics, chemistry, biology, and materials science, we created a scientist-curated coding benchmark, SciCode. The problems in SciCode naturally factorize into multiple subproblems, each involving knowledge recall, reasoning, and code synthesis. In total, SciCode contains 338 subproblems decomposed from 80 challenging main problems. It offers optional descriptions specifying useful scientific background information and scientist-annotated gold-standard solutions and test cases for evaluation. Claude3.5-Sonnet, the best-performing model among those tested, can solve only 4.6% of the problems in the most realistic setting. We believe that SciCode demonstrates both contemporary LMs' progress towards becoming helpful scientific assistants and sheds light on the development and evaluation of scientific AI in the future.

Jiazheng Sun, Te Yang, Jiayang Niu et al.

Large multi-modal models (LMMs) have advanced mobile GUI agents. However, existing methods struggle with real-world scenarios involving diverse app interfaces and evolving user needs. End-to-end methods relying on model's commonsense often fail on long-tail apps, and agents without user interaction act unilaterally, harming user experience. To address these limitations, we propose Fairy, an interactive multi-agent mobile assistant capable of continuously accumulating app knowledge and self-evolving during usage. Fairy enables cross-app collaboration, interactive execution, and continual learning through three core modules:(i) a Global Task Planner that decomposes user tasks into sub-tasks from a cross-app view; (ii) an App-Level Executor that refines sub-tasks into steps and actions based on long- and short-term memory, achieving precise execution and user interaction via four core agents operating in dual loops; and (iii) a Self-Learner that consolidates execution experience into App Map and Tricks. To evaluate Fairy, we introduce RealMobile-Eval, a real-world benchmark with a comprehensive metric suite, and LMM-based agents for automated scoring. Experiments show that Fairy with GPT-4o backbone outperforms the previous SoTA by improving user requirement completion by 33.7% and reducing redundant steps by 58.5%, showing the effectiveness of its interaction and self-learning.

Yifan Hao, Yanxin Lu, Hanning Zhang et al.

As overparameterized models become increasingly prevalent, training loss alone offers limited insight into generalization performance. While smoothness has been linked to improved generalization across various settings, directly enforcing smoothness in neural networks remains challenging. To address this, we introduce Distributional Input Projection Networks (DIPNet), a novel framework that projects inputs into learnable distributions at each layer. This distributional representation induces a smoother loss landscape with respect to the input, promoting better generalization. We provide theoretical analysis showing that DIPNet reduces both local smoothness measures and the Lipschitz constant of the network, contributing to improved generalization performance. Empirically, we validate DIPNet across a wide range of architectures and tasks, including Vision Transformers (ViTs), Large Language Models (LLMs), ResNet and MLPs. Our method consistently enhances test performance under standard settings, adversarial attacks, out-of-distribution inputs, and reasoning benchmarks. We demonstrate that the proposed input projection strategy can be seamlessly integrated into existing models, providing a general and effective approach for boosting generalization performance in modern deep learning.