This handbook presents a systematic treatment of the foundations and architectures of artificial intelligence for motor insurance, grounded in large-scale real-world deployment. It formalizes a vertically integrated AI paradigm that unifies perception, multimodal reasoning, and production infrastructure into a cohesive intelligence stack for automotive risk assessment and claims processing. At its core, the handbook develops domain-adapted transformer architectures for structured visual understanding, relational vehicle representation learning, and multimodal document intelligence, enabling end-to-end automation of vehicle damage analysis, claims evaluation, and underwriting workflows. These components are composed into a scalable pipeline operating under practical constraints observed in nationwide motor insurance systems in Thailand. Beyond model design, the handbook emphasizes the co-evolution of learning algorithms and MLOps practices, establishing a principled framework for translating modern artificial intelligence into reliable, production-grade systems in high-stakes industrial environments.
<p>The ocean-surface downward longwave radiation (<span class="inline-formula"><i>R</i><sub>l</sub></span>) is one of the most fundamental components of the radiative energy balance, and it has a remarkable influence on air–sea interactions. Because of various shortcomings and limits, a lot of empirical models have been established for ocean-surface <span class="inline-formula"><i>R</i><sub>l</sub></span> estimation for practical applications. In this paper, based on comprehensive measurements collected from 65 moored buoys distributed across global seas from 1988 to 2019, a new model for estimating the all-sky ocean-surface <span class="inline-formula"><i>R</i><sub>l</sub></span> at both hourly and daily scales was built. The ocean-surface <span class="inline-formula"><i>R</i><sub>l</sub></span> was formulated as a nonlinear function of the screen-level air temperature, relative humidity, cloud fraction, total column cloud liquid, and ice water. A comprehensive evaluation of this new model relative to eight existing models was conducted under clear-sky and all-sky conditions at daytime/nighttime hourly and daily scales. The validation results showed that the accuracy of the newly constructed model is superior to that of other models, yielding overall root mean square error (RMSE) values of 13.44 and 8.34 W m<span class="inline-formula"><sup>−2</sup></span> under clear-sky conditions and 15.64 and 10.27 W m<span class="inline-formula"><sup>−2</sup></span> under all-sky conditions at hourly and daily scales, respectively. Our analysis indicates that the effects of the total column cloud liquid and ice water on the ocean-surface <span class="inline-formula"><i>R</i><sub>l</sub></span> also need to be considered in addition to cloud cover. Overall, the newly developed model has strong potential to be widely used.</p>
<p>Polarimetric radio occultation (PRO) extends the capability of standard radio occultation (RO) by providing not only the conventional thermodynamic profiles but also information on clouds and precipitation. The differential phase (<span class="inline-formula">ΔΦ</span>) is the cumulative phase shift between horizontal and vertical polarizations observed from PRO caused by aspherical hydrometeors along the propagation path, typically measured in millimeters. In early 2025, Yunyao Aerospace Technology Co., Ltd. successfully launched the first Chinese low Earth orbit satellite equipped with a PRO payload, generating over 500 measurements per day. Based on this mission, we established an end-to-end PRO data processing chain incorporating on-orbit calibration and tailored for operational applications. We analysed approximately 53 000 events collected between March and June 2025, in conjunction with the Integrated Multi-satellite Retrievals for Global Precipitation Measurement (GPM) precipitation product (IMERG). The results show that <span class="inline-formula">ΔΦ</span> remains close to zero under non-precipitating conditions but exhibits distinct peaks at 3–5 km altitude when traversing precipitation layers, with amplitudes strongly correlated with path-averaged rainfall rates. Thresholds of 1, 2, and 5 mm h<span class="inline-formula"><sup>−1</sup></span> are proposed as indicators of precipitation sensitivity, detection confidence, and heavy-rain events, respectively, and a <span class="inline-formula">ΔΦ</span>-to-rainfall intensity mapping table is derived to quantify this relationship. Yunyao PRO data preserve the thermodynamic retrieval quality of conventional RO while enabling effective precipitation detection, thereby providing important data support for the theoretical, technical and data research on the transition of meteorological observations from “temperature, humidity and pressure” observations to new types of observations such as precipitation.</p>
The origins of economic behavior remain unresolved-not only in the social sciences but also in AI, where dominant theories often rely on predefined incentives or institutional assumptions. Contrary to the longstanding myth of barter as the foundation of exchange, converging evidence from early human societies suggests that reciprocity-not barter-was the foundational economic logic, enabling communities to sustain exchange and social cohesion long before formal markets emerged. Yet despite its centrality, reciprocity lacks a simulateable and cognitively grounded account. Here, we introduce a minimal behavioral framework based on three empirically supported cognitive primitives-individual recognition, reciprocal credence, and cost--return sensitivity-that enable agents to participate in and sustain reciprocal exchange, laying the foundation for scalable economic behavior. These mechanisms scaffold the emergence of cooperation, proto-economic exchange, and institutional structure from the bottom up. By bridging insights from primatology, developmental psychology, and economic anthropology, this framework offers a unified substrate for modeling trust, coordination, and economic behavior in both human and artificial systems. For an interactive visualization of the framework, see: https://egil158.github.io/cogfoundations-econ/
Luka Kovačević, Thomas Gaudelet, James Opzoomer
et al.
Despite substantial efforts, deep learning has not yet delivered a transformative impact on elucidating regulatory biology, particularly in the realm of predicting gene expression profiles. Here, we argue that genuine "foundation models" of regulatory biology will remain out of reach unless guided by frameworks that integrate mechanistic insight with principled experimental design. We present one such ground-up, semi-mechanistic framework that unifies perturbation-based experimental designs across both in vitro and in vivo CRISPR screens, accounting for differentiating and non-differentiating cellular systems. By revealing previously unrecognised assumptions in published machine learning methods, our approach clarifies links with popular techniques such as variational autoencoders and structural causal models. In practice, this framework suggests a modified loss function that we demonstrate can improve predictive performance, and further suggests an error analysis that informs batching strategies. Ultimately, since cellular regulation emerges from innumerable interactions amongst largely uncharted molecular components, we contend that systems-level understanding cannot be achieved through structural biology alone. Instead, we argue that real progress will require a first-principles perspective on how experiments capture biological phenomena, how data are generated, and how these processes can be reflected in more faithful modelling architectures.
Approximate Nearest Neighbour (ANN) search is a fundamental problem in information retrieval, underpinning large-scale applications in computer vision, natural language processing, and cross-modal search. Hashing-based methods provide an efficient solution by mapping high-dimensional data into compact binary codes that enable fast similarity computations in Hamming space. Over the past two decades, a substantial body of work has explored learning to hash, where projection and quantisation functions are optimised from data rather than chosen at random. This article offers a foundational survey of early learning-based hashing methods, with an emphasis on the core ideas that shaped the field. We review supervised, unsupervised, and semi-supervised approaches, highlighting how projection functions are designed to generate meaningful embeddings and how quantisation strategies convert these embeddings into binary codes. We also examine extensions to multi-bit and multi-threshold models, as well as early advances in cross-modal retrieval. Rather than providing an exhaustive account of the most recent methods, our goal is to introduce the conceptual foundations of learning-based hashing for ANN search. By situating these early models in their historical context, we aim to equip readers with a structured understanding of the principles, trade-offs, and open challenges that continue to inform current research in this area.
We present the Emotion-Gradient Metacognitive Recursive Self-Improvement (EG-MRSI) framework, a novel architecture that integrates introspective metacognition, emotion-based intrinsic motivation, and recursive self-modification into a unified theoretical system. The framework is explicitly capable of overwriting its own learning algorithm under formally bounded risk. Building upon the Noise-to-Meaning RSI (N2M-RSI) foundation, EG-MRSI introduces a differentiable intrinsic reward function driven by confidence, error, novelty, and cumulative success. This signal regulates both a metacognitive mapping and a self-modification operator constrained by provable safety mechanisms. We formally define the initial agent configuration, emotion-gradient dynamics, and RSI trigger conditions, and derive a reinforcement-compatible optimization objective that guides the agent's development trajectory. Meaning Density and Meaning Conversion Efficiency are introduced as quantifiable metrics of semantic learning, closing the gap between internal structure and predictive informativeness. This Part I paper establishes the single-agent theoretical foundations of EG-MRSI. Future parts will extend this framework to include safety certificates and rollback protocols (Part II), collective intelligence mechanisms (Part III), and feasibility constraints including thermodynamic and computational limits (Part IV). Together, the EG-MRSI series provides a rigorous, extensible foundation for open-ended and safe AGI.
Foundation models have garnered increasing attention for representation learning in remote sensing. Many such foundation models adopt approaches that have demonstrated success in computer vision with minimal domain-specific modification. However, the development and application of foundation models in this field are still burgeoning, as there are a variety of competing approaches for how to most effectively leverage remotely sensed data. This paper examines these approaches, along with their roots in the computer vision field. This is done to characterize potential advantages and pitfalls, while outlining future directions to further improve remote sensing-specific foundation models. We discuss the quality of the learned representations and methods to alleviate the need for massive compute resources. We first examine single-sensor remote foundation models to introduce concepts and provide context, and then place emphasis on incorporating the multi-sensor aspect of Earth observations into foundation models. In particular, we explore the extent to which existing approaches leverage multiple sensors in training foundation models in relation to multi-modal foundation models. Finally, we identify opportunities for further harnessing the vast amounts of unlabeled, seasonal, and multi-sensor remote sensing observations.
<p>The study evaluated a new model of a Plair SA airflow cytometer, Rapid-E<span class="inline-formula">+</span>, and assessed its suitability for airborne pollen monitoring within operational networks. Key features of the new model are compared with the previous one, Rapid-E. A machine learning algorithm is constructed and evaluated for (i) classification of reference pollen types in laboratory conditions and (ii) monitoring in real-life field campaigns. The second goal of the study was to evaluate the device usability in forthcoming monitoring networks, which would require similarity and reproducibility of the measurement signal across devices. We employed three devices and analysed (dis-)similarities of their measurements in laboratory conditions. The lab evaluation showed similar recognition performance to that of Rapid-E, but field measurements in conditions when several pollen types were present in the air simultaneously showed notably lower agreement of Rapid-E<span class="inline-formula">+</span> with manual Hirst-type observations than those of the older model. An exception was the total-pollen measurements. Comparison across the Rapid-E<span class="inline-formula">+</span> devices revealed noticeable differences in fluorescence measurements between the three devices tested. As a result, application of the recognition algorithm trained on the data from one device to another led to large errors. The study confirmed the potential of the fluorescence measurements for discrimination between different pollen classes, but each instrument needed to be trained individually to achieve acceptable skills. The large uncertainty of fluorescence measurements and their variability between different devices need to be addressed to improve the device usability.</p>
<p>Globally available and highly vertically resolved wind fields are crucial for the analysis of atmospheric dynamics for the benefit of climate studies. Most observation techniques have problems to fulfill these requirements. Especially in the tropics and in the Southern Hemisphere more wind data are required. In this study, we investigate the potential of radio-occultation (RO) data for climate-oriented wind field monitoring in the tropics, with a specific focus on the equatorial band within <span class="inline-formula">±</span> 5° latitude. In this region, the geostrophic balance breaks down, due to the Coriolis force term approaching zero, and the equatorial-balance equation becomes relevant. One aim is to understand how the individual wind components of the geostrophic-balance and equatorial-balance approximations bridge across the Equator and where each component breaks down. Our central aim focuses on the equatorial-balance approximation, testing its quality by comparison with ERA5 reanalysis data. The analysis of the zonal and meridional wind components showed that while the zonal wind was well reconstructed, it was difficult to estimate the meridional wind from the approximation. However, we still found a somewhat better agreement from including both components in the zonal-mean total wind speed in the troposphere. In the stratosphere, the meridional wind component is close to zero for physical reasons and has no relevant impact on the total wind speed. In general, the equatorial-balance approximation works best in the stratosphere. As a second aim, we investigated the systematic data bias between using the RO and ERA5 data and find it smaller than the bias resulting from the approximations. We also inspected the monthly-mean RO wind data over the full example year of 2009. The bias in the core region of highest quality of RO data, which is the upper troposphere and lower stratosphere, was generally smaller than <span class="inline-formula">±</span> 2 <span class="inline-formula">m s<sup>−1</sup></span>. This is in line with the wind field requirements of the World Meteorological Organization. Overall, the study encourages the use of RO wind fields for regional-scale climate monitoring over the entire globe, including the equatorial region, and also showed a small improvement in the troposphere when including the meridional wind component in the zonal-mean total wind speed.</p>
<p>NASA's Vapor In-cloud Profiling Radar (VIPR) is a tunable G-band radar designed for in-cloud and precipitation humidity remote sensing. VIPR estimates humidity using the differential absorption radar (DAR) technique. This technique exploits the difference between atmospheric attenuation at different frequencies (“on” and “off” an absorption line) and combines it with the ranging capabilities of the radar to estimate the absorbing gas concentration along the radar path.</p>
<p>We analyze the VIPR humidity measurements during two NASA field campaigns: (1) the Investigation of Microphysics and Precipitation for Atlantic Coast-Threatening Snowstorms (IMPACTS) campaign, with the objective of studying wintertime snowstorms focusing on east coast cyclones; and (2) the Synergies Of Active optical and Active microwave Remote Sensing Experiment (SOA<span class="inline-formula"><sup>2</sup></span>RSE) campaign, which studied the synergy between DAR (VIPR) and differential absorption lidar (DIAL, the High altitude Lidar Observatory – HALO) measurements. We discuss a comparison with dropsondes launched during these campaigns as well as an intercomparison against the ERA5 reanalysis fields. Thus, this study serves as an additional evaluation of ERA5 lower tropospheric humidity fields. Overall, in-cloud and in-snowstorm comparisons suggest that ERA5 and VIPR agree within 20 % or better against the dropsondes. The exception is during SOA<span class="inline-formula"><sup>2</sup></span>RSE (i.e., in fair weather), where ERA5 exhibits up to a 50 % underestimation above 4 km. We also show a smooth transition in water vapor profiles between the in-cloud and clear-sky measurements obtained from VIPR and HALO respectively, which highlights the complementary nature of these two measurement techniques for future airborne and space-based missions.</p>
Daniel Gratzer, Håkon Gylterud, Anders Mörtberg
et al.
When working in Homotopy Type Theory and Univalent Foundations, the traditional role of the category of sets, Set, is replaced by the category hSet of homotopy sets (h-sets); types with h-propositional identity types. Many of the properties of Set hold for hSet ((co)completeness, exactness, local cartesian closure, etc.). Notably, however, the univalence axiom implies that Ob(hSet) is not itself an h-set, but an h-groupoid. This is expected in univalent foundations, but it is sometimes useful to also have a stricter universe of sets, for example when constructing internal models of type theory. In this work, we equip the type of iterative sets V0, due to Gylterud (2018) as a refinement of the pioneering work of Aczel (1978) on universes of sets in type theory, with the structure of a Tarski universe and show that it satisfies many of the good properties of h-sets. In particular, we organize V0 into a (non-univalent strict) category and prove that it is locally cartesian closed. This enables us to organize it into a category with families with the structure necessary to model extensional type theory internally in HoTT/UF. We do this in a rather minimal univalent type theory with W-types, in particular we do not rely on any HITs, or other complex extensions of type theory. Furthermore, the construction of V0 and the model is fully constructive and predicative, while still being very convenient to work with as the decoding from V0 into h-sets commutes definitionally for all type constructors. Almost all of the paper has been formalized in Agda using the agda-unimath library of univalent mathematics.
Growing interest and investment in the capabilities of foundation models has positioned such systems to impact a wide array of public services. Alongside these opportunities is the risk that these systems reify existing power imbalances and cause disproportionate harm to marginalized communities. Participatory approaches hold promise to instead lend agency and decision-making power to marginalized stakeholders. But existing approaches in participatory AI/ML are typically deeply grounded in context - how do we apply these approaches to foundation models, which are, by design, disconnected from context? Our paper interrogates this question. First, we examine existing attempts at incorporating participation into foundation models. We highlight the tension between participation and scale, demonstrating that it is intractable for impacted communities to meaningfully shape a foundation model that is intended to be universally applicable. In response, we develop a blueprint for participatory foundation models that identifies more local, application-oriented opportunities for meaningful participation. In addition to the "foundation" layer, our framework proposes the "subfloor'' layer, in which stakeholders develop shared technical infrastructure, norms and governance for a grounded domain, and the "surface'' layer, in which affected communities shape the use of a foundation model for a specific downstream task. The intermediate "subfloor'' layer scopes the range of potential harms to consider, and affords communities more concrete avenues for deliberation and intervention. At the same time, it avoids duplicative effort by scaling input across relevant use cases. Through three case studies in clinical care, financial services, and journalism, we illustrate how this multi-layer model can create more meaningful opportunities for participation than solely intervening at the foundation layer.
Recently, large models, or foundation models, have exhibited remarkable performance, profoundly impacting research paradigms in diverse domains. Foundation models, trained on extensive and diverse datasets, provide exceptional generalization abilities, allowing for their straightforward application across various use cases and domains. Exploration geophysics is the study of the Earth's subsurface to find natural resources and help with environmental and engineering projects. It uses methods like analyzing seismic, magnetic, and electromagnetic data, which presents unique challenges and opportunities for the development of geophysical foundation models (GeoFMs). This perspective explores the potential applications and future research directions of GeoFMs in exploration geophysics. We also review the development of foundation models, including large language models, large vision models, and large multimodal models, as well as their advancement in the field of geophysics. Furthermore, we discuss the hierarchy of GeoFMs for exploration geophysics and the critical techniques employed, providing a foundational research workflow for their development. Lastly, we summarize the challenges faced in developing GeoFMs, along with future trends and their potential impact on the field. In conclusion, this perspective provides a comprehensive overview of the development, hierarchy, applications, development workflow, and challenges of foundation models, highlighting their transformative potential in advancing exploration geophysics.
Rishi Bommasani, Kevin Klyman, Shayne Longpre
et al.
Foundation models are critical digital technologies with sweeping societal impact that necessitates transparency. To codify how foundation model developers should provide transparency about the development and deployment of their models, we propose Foundation Model Transparency Reports, drawing upon the transparency reporting practices in social media. While external documentation of societal harms prompted social media transparency reports, our objective is to institutionalize transparency reporting for foundation models while the industry is still nascent. To design our reports, we identify 6 design principles given the successes and shortcomings of social media transparency reporting. To further schematize our reports, we draw upon the 100 transparency indicators from the Foundation Model Transparency Index. Given these indicators, we measure the extent to which they overlap with the transparency requirements included in six prominent government policies (e.g., the EU AI Act, the US Executive Order on Safe, Secure, and Trustworthy AI). Well-designed transparency reports could reduce compliance costs, in part due to overlapping regulatory requirements across different jurisdictions. We encourage foundation model developers to regularly publish transparency reports, building upon recommendations from the G7 and the White House.
Juan Luis Gastaldi, John Terilla, Luca Malagutti
et al.
Tokenization - the practice of converting strings of characters from an alphabet into sequences of tokens over a vocabulary - is a critical step in the NLP pipeline. The use of token representations is widely credited with increased model performance but is also the source of many undesirable behaviors, such as spurious ambiguity or inconsistency. Despite its recognized importance as a standard representation method in NLP, the theoretical underpinnings of tokenization are not yet fully understood. In particular, the impact of tokenization on language model estimation has been investigated primarily through empirical means. The present paper contributes to addressing this theoretical gap by proposing a unified formal framework for representing and analyzing tokenizer models. Based on the category of stochastic maps, this framework enables us to establish general conditions for a principled use of tokenizers and, most importantly, the necessary and sufficient conditions for a tokenizer model to preserve the consistency of statistical estimators. In addition, we discuss statistical and computational concerns crucial for designing and implementing tokenizer models, such as inconsistency, ambiguity, finiteness, and sequentiality. The framework and results advanced in this paper contribute to building robust theoretical foundations for representations in neural language modeling that can inform future theoretical and empirical research.
The issue of choosing a structural design and material for supporting structures is an important technical and economic task at the stage of developing a design solution. It depends on a number of factors: consequence class of the structure, reliability of the design solution, savings in basic building materials. The publication provides a classification of design schemes and types of load-bearing structures used in the corresponding solutions. The disadvantages and advantages of using prefabricated and monolithic structures for the installation and subsequent operation of the structural scheme of buildings and structures are considered. Modern design requirements include ensuring the ability of a damaged structure to adapt to new conditions while continuing to function while ensuring the integrity of human life, property and equipment. New conditions mean the consequences of the occurrence of a certain emergency situation, accompanied by weakening or overloading of the load-bearing structures of a structure or soil foundation: a change in the structural design, a combination of new existing loads and a redistribution of internal forces. The publication reflects the results of assessing the redistribution of the stress-strain state of the elements of the “base - foundations - load-bearing structures” system as a result of the implementation of a hypothetical emergency situation with the exclusion of the load-bearing structure from operation. The case of the collapse of one of the vertical load-bearing elements (local failure of the pylon) of an underground floor, which can be used as a dual-use structure, is considered. Calculations for the stability of the structure against progressive collapse were carried out by numerical modeling in the LIRA SAPR-2019 software using a quasi-static calculation and the method of direct integration of dynamics over time. It has been demonstrated that the method of numerical modeling the joint work of a building with a soil base affects the results of a calculation of the progressive collapse of the building frame. The influence of local collapse of a vertical load-bearing element on the redistribution of stresses and strains in the foundation structures of a building section is assessed. The load on the piles under the pylons around the element removed under the local failure scenario is expected to increase by 15...25%.
<p>Polarimetric microphysical retrievals reveal a great potential for the evaluation of numerical models and data assimilation. However, the accuracy of ice microphysical retrievals is still poorly explored. To evaluate these retrievals and assess their accuracy, polarimetric radar measurements are spatially and temporally collocated with in situ aircraft measurements obtained during the OLYMPEX campaign (Olympic Mountain Experiment). Retrievals for ice water content (IWC), total number concentration <span class="inline-formula"><i>N</i><sub>t</sub></span>, and mean volume diameter <span class="inline-formula"><i>D</i><sub>m</sub></span> of ice particles are assessed by comparing an in situ dataset obtained by the University of North Dakota (UND) Citation II aircraft with X-band Doppler on Wheels (DOW) measurements. Sector-averaged range height indicator (RHI) scans are used to derive vertical profiles of microphysical retrievals. The comparison of these estimates with in situ data provides insights into strengths, weaknesses, and the accuracy of the different retrievals and quantifies the improvements in polarimetry-informed retrievals compared to conventional, non-polarimetric ones. In particular, the recently introduced hybrid ice water content retrieval exploiting reflectivity <span class="inline-formula"><i>Z</i><sub>H</sub></span>, differential reflectivity <span class="inline-formula"><i>Z</i><sub>DR</sub></span>, and specific differential phase <span class="inline-formula"><i>K</i><sub>DP</sub></span> outperforms other retrievals based on either (<span class="inline-formula"><i>Z</i><sub>H</sub></span>, <span class="inline-formula"><i>Z</i><sub>DR</sub></span>) or (<span class="inline-formula"><i>Z</i><sub>H</sub></span>, <span class="inline-formula"><i>K</i><sub>DP</sub></span>) or non-polarimetric retrievals in terms of correlations with in situ measurements and the root mean square error.</p>