Papers

Large language model (LLM) agents have evolved towards greater autonomy with the advancement of model context protocols. Self-controlled agents, such as Codex and Claude Code, highlight the need for novel organizational frameworks that facilitate agent-level autonomy. In this paper, we propose a tree-based orchestration system, TrAgent, which utilizes a PUCT-style search to dynamically allocate agent actions while maintaining autonomy. This approach offers three key benefits: (i) full agent autonomy for critical tasks like planning and tool use, (ii) a generalized mechanism for inter-agent experience sharing, and (iii) scalability as the number of agents increases. We demonstrate the system’s effectiveness through the general matrix multiplication kernel optimization, achieving 80\% of the performance of the cuBLAS code. Additionally, the system exhibits a scaling phenomenon as the number of agents increases. Our approach provides a solution for organizing increasingly autonomous agents.

Determining the crystal structures of inorganic crystalline materials is crucial as the structures encode essential information about their physical, chemical, and mechanical properties. Powder X-ray diffraction is one of the most widely used structural characterization techniques. However, determining crystal structure directly from experimental powder X-ray diffraction patterns can be challenging and requires significant crystallographic knowledge, which still heavily relies on manual inspection by human experts. Even the state-of-the-art databases contain thousands of entries with incomplete or implausible crystal structure information. In this work, we trained a diffusion model based on equivariant graph neural networks that can infer atomic coordinates from powder X-ray diffraction patterns. Starting from a random guess, our model iteratively refines atom coordinates until it reaches a chemically reasonable structure that matches the target diffraction pattern. Our approach is both efficient and accurate. It takes on average 0.6 seconds to solve the atomic positions per crystal structure, which is several orders of magnitude faster than previous approaches. The success rate reaches 82.3% and 81.6% on the simulated and experimental diffraction datasets, respectively. We revisited energetically unfavorable crystal structures in the database and demonstrated that our model can propose more plausible structure solutions for 39 entries. We also suggested 912 complete crystal structure models for entries in the database lacking all or partial atomic positions, including entries that contain light elements, are natural minerals, or exhibit chemical disorder lattice sites. We demonstrated that conditional equivariant generative model can tackle the structure determination problem and provide high-quality structure models for inorganic crystalline materials, paving the way for automated structural analysis of diffraction patterns in autonomous materials development loops.

Modeling complex human behavior, such as voter decisions in national elections, is a long-standing challenge for computational social science. Traditional agent-based models (ABMs) are limited by oversimplified rules, while large-scale statistical models often lack interpretability. We introduce ElectionFit, a novel framework that uses Large Language Models (LLMs) to build a ``computational laboratory'' of LLM agents for political simulation. Each agent is instantiated with a high-fidelity demographic profile and dynamic contextual information (e.g., candidate policies), enabling it to perform nuanced, generative reasoning to simulate a voting decision. We deployed this framework as a testbed on the 2024 U.S. Presidential Election, focusing on seven key swing states. Our simulation's macro-level results successfully replicated the real-world outcome, demonstrating the high fidelity of our ``virtual society''. The primary contribution is not only the prediction, but also the framework's utility as an interpretable research tool. ElectionFit moves beyond black-box outputs, allowing researchers to probe agent-level rationale and analyze the stability and sensitivity of LLM-driven social simulations.

Solar energy is a clean and renewable resource, and the low-rise, unobstructed rural buildings of northern China provide ideal conditions for photovoltaic (PV) installation compared to shaded, high-density urban areas. Yet, progress in assessing rural solar potential is limited by the absence of accurate 3D building data. This study proposes a rapid estimation approach integrating deep learning, parametric modeling, and GPU-accelerated simulation. Convolutional neural net- works (CNNs) extract building footprints from satellite imagery, which are then processed in Grasshopper to generate refined vector outlines. Combined with digital surface model (DSM) data, these outlines produce precise 3D village models. Using Vitality 2.0 for GPU-based solar simulation, the method was applied to 31 villages in Tianjin, generating parametric 3D models and estimating their solar potential. Results show that low building heights and minimal mutual shading make photovoltaic capacity scale with roof area—larger villages have greater generation potential. Moreover, villages with metal roofs exhibit higher conversion efficiency and shorter cost-recovery periods than those with concrete or ceramic-tile roofs, due to better heat dissipation. Overall, the workflow offers a practical and efficient solution for estimating rural solar potential in data-scarce regions to guide renewable energy planning and investment.

Large Language Models (LLMs) have advanced Verilog code generation but still suffer from data quality, limited reasoning, and inefficiency. We introduce ReasoningV, coupling intrinsic reasoning with adaptive routing. Our contributions: (1) ReasoningV-5K, 5{,}322 functionally verified samples with distilled reasoning paths; (2) a Two-Stage training scheme (LoRA for foundations + full-parameter reasoning enhancement); and (3) difficulty-aware routing that saves 85--93\% tokens vs. a strong commercial model and 32--75\% vs. fixed-depth variants. On VerilogEval-human, RV-14B attains 73.9\% pass@1; RV-7B reaches 57.8\% with superior efficiency. Models, data, and code: \url{https://github.com/BUAA-CLab/ReasoningV}.

Transcriptional regulation, critical for cellular differentiation and adaptation to environmental changes, involves coordinated interactions among DNA sequences, regulatory proteins, and chromatin architecture. Despite extensive data from consortia like ENCODE, understanding the dynamics of cis-regulatory elements (CREs) in gene expression remains challenging. Deep learning is a powerful tool for learning gene expression and epigenomic signals from DNA sequences, exhibiting superior performance compared to conventional machine learning approaches. However, even the most advanced deep learning-based methods may fall short in capturing the regulatory effects of distal elements such as enhancers, limiting their predictive accuracy. In addition, these methods may require significant resources to train or to adapt to newly generated data. To address these challenges, we present EPInformer, a scalable deep-learning framework for predicting gene expression by integrating promoter-enhancer interactions with their sequences, epigenomic signals, and chromatin contacts. Our model outperforms existing gene expression prediction models in rigorous cross-chromosome validation, accurately recapitulates enhancer-gene interactions validated by CRISPR perturbation experiments, and identifies crucial transcription factor motifs within regulatory sequences.

The function of constructing the hierarchy of objects is important to the visual process of the human brain. Previous studies have successfully adopted capsule networks to decompose the digits and faces into parts in an unsupervised manner to investigate the similar perception mechanism of neural networks. However, their descriptions are restricted to the 2D space, limiting their capacities to imitate the intrinsic 3D perception ability of humans. In this paper, we propose an Inverse Graphics Capsule Network (IGC-Net) to learn the hierarchical 3D face representations from large-scale unlabeled images. The core of IGC-Net is a new type of capsule, named graphics capsule, which represents 3D primitives with interpretable parameters in computer graphics (CG), including depth, albedo, and 3D pose. Specifically, IGC-Net first decomposes the objects into a set of semantic-consistent part-level descriptions and then assembles them into object-level descriptions to build the hierarchy. The learned graphics capsules reveal how the neural networks, oriented at visual perception, understand faces as a hierarchy of 3D models.