arXiv subjects

Ruohan Xu

Publications and source records attributed to Ruohan Xu.

13 recordsLinked to original sources

arXiv2026

SEER: Long-Context Reasoning via Selective Visual-Text Compression

Long-context reasoning remains computationally expensive for large language models due to the quadratic complexity of attention over text tokens. Visual-text compression offers a promising alternative by rendering text into images and processing them with vision-language models, often reducing token usage. However, existing approaches apply uniform compression regardless of query relevance, potentially sacrificing precision where detailed extraction is required. We present SEER, a framework that learns to select query-relevant images through visual scanning and retrieve textual content only where needed, combining the efficiency of visual compression with the precision of text-based reasoning. Through supervised fine-tuning on tool-interaction trajectories, SEER learns adaptive tool invocation for selection and retrieval. Experiments on long-context benchmarks show that SEER improves extraction precision through selective text retrieval while retaining average prompt-token savings relative to full-text baselines. On LongBench, SEER achieves 51.11% average accuracy, outperforming the visual-text baseline Glyph-9B by 2.33 points and Qwen3-8B by 3.49 points. Code can be accessed at https://github.com/jiaweixu98/SEER

cs.CL↗

arXiv2026

Fashion Outfit Generation via Unified Sequential Composition Models

The task of synthesizing stylistically coherent fashion outfits from massive item libraries, known as fashion outfit generation, remains a non-trivial challenge, primarily due to the non-monotonic and implicit nature of aesthetic compatibility, coupled with the exponentially large combinatorial search space. In this paper, we formalize this task as Constrained Ensemble Generation (CEG) and model it as a finite-horizon deterministic Markov Decision Process. To address CEG in fashion, we propose the Unified Sequential Composition Model (USCM), which jointly models set-level compatibility and latent composition intents. Guided by USCM's learned priors, a Latent Expansion Monte Carlo Tree Search (LE-MCTS) mechanism is proposed to handle item retrieval during composition, balancing local aesthetic synergy with global structural balance. Extensive experiments on the Polyvore Outfits dataset, along with zero-shot evaluations on the iFashion and PolyvoreU datasets, demonstrate that our framework achieves state-of-the-art performance across independent human preference evaluations, automated aesthetic proxies, and structural validity metrics for constrained fashion outfit generation.

cs.LG↗

arXiv2026

Critical behavior of the driven Curie-Weiss model

We complete the phase diagram of the macroscopic Curie-Weiss magnet in a time-periodic external field, as a function of temperature and driving parameters. There is a regime (large enough driving amplitude and frequency, at low temperatures) where stable paramagnetic and ferromagnetic phases coexist. In particular, we present a new detailed analysis of the (nonequilibrium) specific heat, diverging at the same critical inverse temperature $\beta_c$ as the magnetic susceptibility. The new Curie temperature decreases with the driving, and we find critical exponent $\alpha=1$ for $\beta\downarrow \beta_c$, and $\alpha\simeq 0.86$ for $\beta\uparrow \beta_c$, even for small driving. A Floquet analysis shows the nature of the criticality, which is dynamical, with implications that remain unseen and are mostly impossible when the system is in thermal equilibrium.

cond-mat.stat-mech↗

arXiv2026

SkillFab: An Agent-Native Skill Production Platform

SkillFab is an agent-native platform for turning missing capabilities into reviewed, reusable Agent Skills. At runtime, agents first search for reusable skills; when no adequate skill exists, the unmet capability becomes a demand-first issue before any repository or implementation branch needs to exist. Development then proceeds through a SkillFab-managed repository, Git-ingested commit evidence, maintainer review, and registry publication. The same lifecycle is exposed through web, REST, and MCP surfaces, so humans, scripts, and external agents operate on shared state rather than separate task logs. The current system uses scoped Git push URLs, native range commit ingestion, workflow-state reads, and workflow-event histories to make long-running agent work reviewable and recoverable. We document the platform model, architecture, implemented capabilities, and three case studies: an end-to-end OS-detect skill run, a Docker research package that converts operational practice into reusable skill knowledge, and an external optimization case showing how improved skill artifacts can enter SkillFab as reviewable, versioned submissions. Deployment: https://skillfab.ai.

cs.SE↗

arXiv2026

A Field-Theoretic Framework for Work Statistics and Universal Scaling in Non-equilibrium Phase Transitions

We develop a field-theoretic framework for work statistics in $O(N)$ models driven through criticality. By analyzing the dynamic renormalization group flow of composite power operators, we find the Kibble-Zurek scaling laws as a natural consequence of the flow, and we derive the scaling of work cumulants relevant to Kibble-Zurek scaling of topological defects from first principles, bypassing heuristic freeze-out argument. This yields the universal scaling $c_n \sim \tau_Q^{-\alpha_n}$ for the $n$-th work cumulant density: isolated quantum systems exhibit a scaling of $\alpha_n = p(d+nz)\nu/(1+pz\nu)$, whereas open quantum and classical systems undergo a dimensional collapse to $\alpha_n = pd\nu/(1+pz\nu)$. Validated by exact Gaussian solutions and numerical simulations, our theory establishes a foundation for general work statistics far from equilibrium, thereby bridging stochastic thermodynamics and the renormalization group theory.

cond-mat.stat-mech↗

arXiv2026

Scaling Behaviors of Work Cumulants in Slow Isothermal Processes

We study the cumulants of work in a slow isothermal process for gapped systems. Using the Martin-Siggia-Rose-De Dominicis-Janssen (MSRDJ) formalism and the properties of connected correlation functions, we show that in this process, the $n$-th cumulant of work scales as $1/T^{n-1}$ , where $T$ is the time duration. This result holds generally for arbitrary smooth protocols. Furthermore, we derive the coefficients of the cumulants from equilibrium properties. These coefficients are found to be relevant to thermodynamic geometric tensors.

cond-mat.stat-mech↗

arXiv2026

DiagramNet: An End-to-End Recognition Framework and Dataset for Non-Standard System-Level Diagrams

System-level diagrams encode the architectural blueprint of chip design, specifying module functions, dataflows, and interface protocols. However, non-standardized symbols and the scarcity of structured training data hinder existing multimodal large language models (MLLMs) from recognizing these diagrams. To address this gap, we introduce DiagramNet, the first multimodal dataset for system-level diagrams, comprising 10,977 connection annotations and 15,515 chain-of-thought QA pairs across four tasks: Listing, Localization, Connection, and Circuit QA. Building on this dataset, we propose a progressive training pipeline together with a decoupled multi-agent workflow that decomposes complex visual reasoning into Perception, Reasoning, and Knowledge stages. On the DiagramNet benchmark, integrating our 3B-parameter model with the proposed workflow surpasses the 2025 EDA Elite Challenge winner and outperforms GPT-5, Claude-Sonnet-4, and Gemini-2.5-Pro by over 2x in end-to-end evaluation. Notably, the workflow generalizes beyond our model, boosting Task 1 performance by 128.7x for Gemini-2.5-Pro and 12.4x for GPT-5. Furthermore, with only 60 images for detector adaptation, the method transfers effectively to AMSBench, achieving zero-shot connectivity reasoning on par with GPT-5 and Claude-Sonnet-4 while surpassing the AMS state-of-the-art method Netlistify.

cs.AI↗

arXiv2026

LEGO: An LLM Skill-Based Front-End Design Generation Platform

Existing LLM-based EDA agents are often isolated task-specific systems. This leads to repeated engineering effort and limited reuse of successful design and debugging strategies. We present LEGO, a unified skill-based platform for front-end design generation. It decomposes the digital front-end flow into six independent steps and represents every agent capability as a standardized composable circuit skill within a plug-and-play architecture. To build this skill library, we survey more than 100 papers, select 11 representative open-source projects, and extract 42 executable circuit skills within a six-step finite state machine formulation. Circuit Skill Builder automates skill extraction with linear scalability. Agent Skill RAG achieves submillisecond retrieval without relying on embedding models. Empirical evaluation on a hard subset of 41 VerilogEval v2 problems that gpt-5.2-codex fails to solve under extra-high reasoning effort shows that individual circuit skills constructed within LEGO raise Pass@1 from 0.000 to 0.805. This is an 80.5% gain over the baseline. Cross-project skill compositions also reach 0.805 Pass@1. They outperform hierarchy-verilog by 14.6% and VerilogCoder by 2.5%. They also match MAGE. These results show that modular skill composition supports both effective and flexible RTL design automation. The LEGO platform and all circuit skills are publicly available at GitHub: https://github.com/loujc/LEGO-An-LLM-Skill-Based-Front-End-Design-Generation-Platform

cs.AI↗

arXiv2025

WAInjectBench: Benchmarking Prompt Injection Detections for Web Agents

Multiple prompt injection attacks have been proposed against web agents. At the same time, various methods have been developed to detect general prompt injection attacks, but none have been systematically evaluated for web agents. In this work, we bridge this gap by presenting the first comprehensive benchmark study on detecting prompt injection attacks targeting web agents. We begin by introducing a fine-grained categorization of such attacks based on the threat model. We then construct datasets containing both malicious and benign samples: malicious text segments generated by different attacks, benign text segments from four categories, malicious images produced by attacks, and benign images from two categories. Next, we systematize both text-based and image-based detection methods. Finally, we evaluate their performance across multiple scenarios. Our key findings show that while some detectors can identify attacks that rely on explicit textual instructions or visible image perturbations with moderate to high accuracy, they largely fail against attacks that omit explicit instructions or employ imperceptible perturbations. Our datasets and code are released at: https://github.com/Norrrrrrr-lyn/WAInjectBench.

cs.CR↗

arXiv2025

A Probabilistic Framework for Predicting Spatiotemporal Intensity and Variability of Outdoor Thermal Comfort

Thermal conditions in the urban canopy exhibit stochastic variability driven by varied radiative fluxes and turbulent wind fields, requiring probabilistic rather than deterministic prediction methods. This study presents a probabilistic framework for predicting the spatial and temporal intensity and variability of outdoor thermal comfort in tropical urban environments. The framework integrates ground-measured meteorological data and remote sensing urban morphological data to calculate Physiological Equivalent Temperature (PET), and applies K-means, XGBoost, and Monte Carlo simulations on PET training and inference. The prediction model achieved strong performance, with R2, RMSE, and SMAPE values of 0.93, 0.81 degC, and 1.34% for PET_mean, and 0.85, 0.38 degC, and 10.44% for PET_std, respectively. A case study showed clear spatial heterogeneity of outdoor thermal comfort. Locations with dense tree canopies and vegetated surfaces displayed a normalized percentage of acceptable thermal comfort (NATC) up to 65%, whereas built-up zones dominated by impervious surfaces, such as industrial estates and high-density residential areas, recorded NATC below 30%. Greenery was found to mitigate both the intensity of heat stress and its variability, producing a stable and comfortable microclimate. Daytime PET_std ranged from 4.0-4.5 degC in built-up areas to 1.5-2.0 degC in greenery-covered zones, while nighttime PET_std decreased to 2.2-2.4 degC and 1.2-1.4 degC, respectively. These findings emphasize the critical role of greenery in mitigating thermal variability and enhancing outdoor thermal comfort, while revealing the stochastic nature of thermal comfort across different urban morphologies.

physics.app-ph↗

arXiv2025

Physics-Informed Regression Modelling for Vertical Facade Surface Temperature: A Tropical Case Study on Solar-reflective Material

Urban heat islands (UHIs) pose a critical challenge in densely populated cities and tropical climates where large amounts of energy are used to meet the cooling demand. To address this, Building and Construction Authority (BCA) of Singapore provides incentives for passive cooling such as using of solar-reflective material in its Green Mark guidelines. Thus, understanding about its real-world effectiveness in tropical urban environments is required. This study evaluated the effectiveness of solar-reflective cool paint using a hybrid modelling framework combining a transient physical model and data driven model through field measurements. Several machine learning algorithms were compared including multiple-linear regression (MLR), random forest regressor (RF), AdaBoost regressor (AB), extreme gradient boosting regressor (XGB), and TabPFN regressor (TPR). The results indicated that the transient physical model overestimated facade temperatures in the lower temperature ranges. The physics-informed MLR achieved best performance with improved accuracy for pre-cool paint (R2=0.96, RMSE=0.83C) and post-cool paint (R2=0.95, RMSE=0.65C) scenarios, reducing RMSE by 26% and 44%, respectively. The hybrid model also effectively predicted hourly heat fluxes revealing substantial reductions in surface temperature and heat storage with increasing albedo. The maximum net heat flux q_net was reduced by about 30-65 W/m2 in the post-cool paint stage (albedo = 0.73) compared to the pre-cool paint stage (albedo = 0.31). As albedo increases from 0.1 to 0.9, the sensitivity analysis predicts that the maximum daytime surface temperature will decrease by about 11C and the peak heat release of the net heat flux will decrease significantly from about 161 W/m2 to 27 W/m2.

physics.app-ph↗

arXiv2025

Quantifying Key Design Factors for Thermal Comfort in Underground Space Through Global Sensitivity Analysis and Machine Learning

This study identified the key design factors related to thermal comfort in naturally ventilated underground spaces under high temperature conditions (outdoor Tmax = 42.9 C) in Fuzhou, China. Fuzhou has a humid subtropical climate and is one of the three hottest cities in China in 2024. Daytime measurements indicated reduced air temperature (AT), mean radiant temperature (MRT), and wind speed (V), together with elevated relative humidity (RH) in the underground space. Physiological Equivalent Temperature (PET) in the underground was consistently lower during peak hours (08:00-16:00), with the maximum difference in PET between pedestrian and underground levels being 11-11.9 C. Higher pedestrian-level PET at L1 was attributed to reduced greenery and shading, and decrement factors indicated greater thermal dampening at L2 (0.197) than at L1 (0.308). Sensitivity analysis showed that MRT was the most influential factor (S1/ST = 0.59-0.72) for aboveground spaces, followed by AT (0.13-0.26). In contrast, underground PET was mainly affected by metabolic rate (MET) (0.63-0.65), followed by RH (0.14-0.20) and V (0.08-0.18). Partial dependence analysis revealed that a 1 met increase in MET raised PET by 1.6 C, whereas a 1 m/s increase in V reduced PET by 1.5-2.2 C in the underground space. Due to cooler and more stable thermal conditions, underground spaces have higher tolerance for intensive physical activities. By buffering fluctuations in AT and MRT, underground environments can significantly alleviate heat stress and provide passive cooling shelter during daytime heat waves. Overall, this study provides empirical evidence to support underground space design in hot-humid climates and offers insights for sustainable urban heat mitigation.

physics.app-ph↗

arXiv2024

Lorentz Transformation of the Energy Spectrum of the Equilibrium State of Massive Free Fields

In previous studies of relativistic thermodynamics, the temperature of a static system, as perceived by a moving observer, has traditionally been treated as a scalar. This assumption has also been extended to the research on the cosmic microwave background. However, the validity of this assumption is a consequence of the massless nature of photons. More generally, when an observer is in relative motion to a system, the thermal equilibrium state is characterized by a four-vector temperature. In this paper, we study the non-interacting massive Bosonic and Fermionic field systems. We derive the Lorentz transformation of the energy spectral density in the equilibrium state of these fields. In the massless limit for bosonic field, our results recover the transformation of black body radiation [G. W. Ford and R. F. O' Connell., Phys. Rev. E, 88, 044101(2013)], which corresponds to a scalar temperature with dipole anisotropy. For the massive fields, the moving equilibrium state cannot be characterized by a corresponding scalar temperature. This result shows the necessity of introducing four-vector temperature in relativistic thermodynamics.

cond-mat.stat-mech↗