
Accurate diagnosis of cracks in concrete structures requires integrating data-driven perception with mechanismbased knowledge. This paper presents a knowledge-infused framework integrating crack perception, knowledgeguided reasoning, and human-in-the-loop validation, implemented as an engineer-friendly interactive system for interpretable assessment. A domain-specific knowledge graph containing 530 nodes and 716 edges is constructed to represent crack characteristics, deterioration mechanisms, and maintenance strategies, and RetrievalAugmented Generation (RAG) is integrated to enhance standards-guided reasoning and improve code compliance. The crack detection and quantification module achieved high accuracy (R2 = 0.976, MAE = 0.505 mm) with an average processing time of 4.6 s per image. The framework is completed within 164.6 s per case. Under the representative bridge inspection scenario considered in this study, the proposed framework is estimated to reduce inspection labor cost and time by 78.3% and 68.7%, respectively, compared with manual inspection. The framework provides an efficient, transparent, and standards-guided solution for intelligent infrastructure maintenance.
We study differentiability properties of convex operators defined on a Banach space with values in an $\Lc_p$ space and of their compositions with monotonic convex functionals on this space. We develop new tools for operators enjoying an additional feature known as the local property. The new approach and results go beyond the classical theory of normal integrands and lattice-valued operators. We further describe the subdifferentials of compositions of such operators with convex monotonic functionals. The new results are applied to obtain novel optimality conditions in the subdifferential form for a broad class of risk-averse stochastic optimization problems with risk functionals as objectives, with partial information, and with stochastic dominance constraints. While our analysis is motivated by the theory and methods of risk-averse optimization, it addresses problems of a more general structure and has potential for further applications.
Reward modeling is essential for aligning large language models with human preferences through reinforcement learning. To provide accurate reward signals, a reward model (RM) should stimulate deep thinking and conduct interpretable reasoning before assigning a score or a judgment. Inspired by recent advances of long chain-of-thought on reasoning-intensive tasks, we hypothesize and validate that integrating reasoning into reward modeling significantly enhances RM's interpretability and performance. We introduce a new class of generative reward models, Reasoning Reward Models (ReasRMs), which formulate reward modeling as a reasoning task. We propose a reasoning-oriented training pipeline and train a family of ReasRMs, RM-R1. RM-R1 features a chain-of-rubrics (CoR) mechanism -- self-generating sample-level chat rubrics or math/code solutions, and evaluating candidate responses against them. The training of RM-R1 consists of two key stages: (1) distillation of high-quality reasoning chains and (2) reinforcement learning with verifiable rewards. Empirically, our models achieve superior performance across three reward model benchmarks on average, outperforming much larger open-weight models (e.g., INF-ORM-Llama3.1-70B) and proprietary ones (e.g., GPT-4o) by up to 4.9%. Beyond final performance, we perform thorough analyses to understand the key ingredients of successful ReasRM training.
We examine how lead venture capital (VC) investors adjust VC syndicate resource diversity to meet ventures’ evolving developmental needs. As ventures transition from early to later stages, lead VCs tend to assemble more resource-diverse syndicates to support expanding resource requirements. However, greater diversity also increases coordination costs of managing the syndicate. We argue that lead VCs balance these competing considerations, such that syndicate composition depends on attributes of both the lead VC and the syndicate that shape this trade-off. We test our conjectures using a sample of US ventures receiving VC funding between 1990 and 2019. Our findings offer an evolutionary perspective on VC syndicate composition across venture life stages. More diverse investors do not always mean better support for startups. Indeed, too much diversity in a VC syndicate can backfire. This study examines how lead venture capital (VC) investors adjust the diversity of their investment syndicates as startups grow. We find that as startups move from early to later stages, lead VCs tend to bring in more diverse partners to meet expanding needs, but they do so selectively because greater diversity also makes coordination more difficult. As a result, syndicate composition reflects a trade-off between accessing a broader range of resources and maintaining manageable collaboration. Our findings imply that both investors and entrepreneurs should be strategic in assembling their investor base, seeking broader expertise as their ventures grow while avoiding overly complex syndicates that may slow decision-making and execution.
LLM-powered agents execute tasks through a sequential loop of model generation and tool execution. Today's serving systems serialize this loop, leaving tool latency exposed on the task critical path. This paper presents PASTE, a tool-aware agent-serving system that predicts concrete future tool invocations from recurring agent patterns and executes them speculatively while the LLM is still generating. PASTE isolates speculative results until confirmed by the LLM and jointly schedules tool execution and returning LLM sessions to avoid shifting bottlenecks to the GPU. Across deep research, coding, and scientific-agent workloads, PASTE reduces average task completion time by 43.5