arXiv subjects

Alberto Pepe

Publications and source records attributed to Alberto Pepe.

At least 19 recordsLinked to original sources

arXiv2026

AI Research Preference Models

AI research agents (AIRA) can now carry machine learning experiments from proposal through implementation and evaluation. Yet progress on frontier tasks is throttled by the cost of evaluations that can consume days of GPU time. When an agent can propose far more candidates than it can afford to run, progress depends on its research preference: how it allocates a fixed execution budget across many candidates. We introduce AI Research Preference Models (RPMs) that predict which candidate solution is most promising, without paying the cost of running them all. We build RPMs from frozen pretrained language models in two variants: an inference-only model that reasons over candidate plans, code, and previously executed solutions, and an agentic model that additionally runs small-scale pilot experiments. Integrated into the AIRA-dojo research agent and evaluated on the machine learning research benchmark AIRS-Bench, the two variants increase the average normalized score from 0.684 to 0.711 and 0.729, respectively. Both reach the unguided agent's 24-hour performance in roughly 15 hours, using less than two-thirds of its execution budget, and together yield new state-of-the-art results on two AIRS-Bench tasks.

cs.AI↗

arXiv2026

Agentic Discovery of Neural Architectures: AIRA-Compose and AIRA-Design

Toward recursive self-improvement, we investigate LLM agents autonomously designing foundation models beyond standard Transformers. We introduce a dual-framework approach: AIRA-Compose for high-level architecture search, and AIRA-Design for low-level mechanistic implementation. AIRA-Compose uses 11 agents to explore fundamental computational primitives under a 24-hour budget. Agents evaluate million-parameter candidates, extrapolating top designs to 350M, 1B, and 3B scales. This yields 14 architectures across two families: AIRAformers (Transformer-based) and AIRAhybrids (Transformer-Mamba). Pre-trained at 1B scale, these consistently outperform Llama 3.2 and Composer-found baselines. On downstream tasks, AIRAformer-D and AIRAhybrid-D improve accuracy by 2.4% and 3.8% over Llama 3.2. Furthermore, AIRA-Compose finds models with highly efficient scaling frontiers: AIRAformer-C scales 54% and 71% faster than Llama 3.2 and Composer's best Transformer, while AIRAhybrid-C outscales Nemotron-2 by 23% and Composer's best hybrid by 37%. AIRA-Design tasks 20 agents with writing novel attention mechanisms for long-range dependencies and high-performing training scripts. On the Long Range Arena benchmark, agent-designed architectures reach within 2.3% and 2.6% of human state-of-the-art on document matching and text classification. On the Autoresearch benchmark, Greedy Opus 4.5 achieves 0.968 validation bits-per-byte under a fixed time budget, surpassing the published minimum. Together, these frameworks show AI agents can autonomously discover architectures and algorithmic optimizations matching or surpassing hand-designed baselines. This establishes a powerful paradigm for discovering next-generation foundation models, marking a clear step toward recursive self-improvement.

cs.AI↗

arXiv2026

AIRS-Bench: a Suite of Tasks for Frontier AI Research Science Agents

LLM agents hold significant promise for advancing scientific research. To accelerate this progress, we introduce AIRS-Bench (the AI Research Science Benchmark), a suite of 20 tasks sourced from state-of-the-art machine learning papers. These tasks span diverse domains, including language modeling, mathematics, bioinformatics, and time series forecasting. AIRS-Bench tasks assess agentic capabilities over the full research lifecycle -- including idea generation, experiment analysis and iterative refinement -- without providing baseline code. The AIRS-Bench task format is versatile, enabling easy integration of new tasks and rigorous comparison across different agentic frameworks. We establish baselines using frontier models paired with both sequential and parallel scaffolds. Our results show that agents exceed human SOTA in four tasks but fail to match it in sixteen others. Even when agents surpass human benchmarks, they do not reach the theoretical performance ceiling for the underlying tasks. These findings indicate that AIRS-Bench is far from saturated and offers substantial room for improvement. We open-source the AIRS-Bench task definitions and evaluation code to catalyze further development in autonomous scientific research.

cs.AI↗

arXiv2024

STAResNet: a Network in Spacetime Algebra to solve Maxwell's PDEs

We introduce STAResNet, a ResNet architecture in Spacetime Algebra (STA) to solve Maxwell's partial differential equations (PDEs). Recently, networks in Geometric Algebra (GA) have been demonstrated to be an asset for truly geometric machine learning. In \cite{brandstetter2022clifford}, GA networks have been employed for the first time to solve partial differential equations (PDEs), demonstrating an increased accuracy over real-valued networks. In this work we solve Maxwell's PDEs both in GA and STA employing the same ResNet architecture and dataset, to discuss the impact that the choice of the right algebra has on the accuracy of GA networks. Our study on STAResNet shows how the correct geometric embedding in Clifford Networks gives a mean square error (MSE), between ground truth and estimated fields, up to 2.6 times lower than than obtained with a standard Clifford ResNet with 6 times fewer trainable parameters. STAREsNet demonstrates consistently lower MSE and higher correlation regardless of scenario. The scenarios tested are: sampling period of the dataset; presence of obstacles with either seen or unseen configurations; the number of channels in the ResNet architecture; the number of rollout steps; whether the field is in 2D or 3D space. This demonstrates how choosing the right algebra in Clifford networks is a crucial factor for more compact, accurate, descriptive and better generalising pipelines.

cs.LG↗

arXiv2023

CGA-PoseNet: Camera Pose Regression via a 1D-Up Approach to Conformal Geometric Algebra

We introduce CGA-PoseNet, which uses the 1D-Up approach to Conformal Geometric Algebra (CGA) to represent rotations and translations with a single mathematical object, the motor, for camera pose regression. We do so starting from PoseNet, which successfully predicts camera poses from small datasets of RGB frames. State-of-the-art methods, however, require expensive tuning to balance the orientational and translational components of the camera pose.This is usually done through complex, ad-hoc loss function to be minimized, and in some cases also requires 3D points as well as images. Our approach has the advantage of unifying the camera position and orientation through the motor. Consequently, the network searches for a single object which lives in a well-behaved 4D space with a Euclidean signature. This means that we can address the case of image-only datasets and work efficiently with a simple loss function, namely the mean squared error (MSE) between the predicted and ground truth motors. We show that it is possible to achieve high accuracy camera pose regression with a significantly simpler problem formulation. This 1D-Up approach to CGA can be employed to overcome the dichotomy between translational and orientational components in camera pose regression in a compact and elegant way.

cs.CV↗

arXiv2017

The arXiv of the future will not look like the arXiv

The arXiv is the most popular preprint repository in the world. Since its inception in 1991, the arXiv has allowed researchers to freely share publication-ready articles prior to formal peer review. The growth and the popularity of the arXiv emerged as a result of new technologies that made document creation and dissemination easy, and cultural practices where collaboration and data sharing were dominant. The arXiv represents a unique place in the history of research communication and the Web itself, however it has arguably changed very little since its creation. Here we look at the strengths and weaknesses of arXiv in an effort to identify what possible improvements can be made based on new technologies not previously available. Based on this, we argue that a modern arXiv might in fact not look at all like the arXiv of today.

cs.DL↗

arXiv2014

10 Simple Rules for the Care and Feeding of Scientific Data

This article offers a short guide to the steps scientists can take to ensure that their data and associated analyses continue to be of value and to be recognized. In just the past few years, hundreds of scholarly papers and reports have been written on questions of data sharing, data provenance, research reproducibility, licensing, attribution, privacy, and more, but our goal here is not to review that literature. Instead, we present a short guide intended for researchers who want to know why it is important to "care for and feed" data, with some practical advice on how to do that.

cs.DL↗

arXiv2012

How the Scientific Community Reacts to Newly Submitted Preprints: Article Downloads, Twitter Mentions, and Citations

We analyze the online response to the preprint publication of a cohort of 4,606 scientific articles submitted to the preprint database arXiv.org between October 2010 and May 2011. We study three forms of responses to these preprints: downloads on the arXiv.org site, mentions on the social media site Twitter, and early citations in the scholarly record. We perform two analyses. First, we analyze the delay and time span of article downloads and Twitter mentions following submission, to understand the temporal configuration of these reactions and whether one precedes or follows the other. Second, we run regression and correlation tests to investigate the relationship between Twitter mentions, arXiv downloads and article citations. We find that Twitter mentions and arXiv downloads of scholarly articles follow two distinct temporal patterns of activity, with Twitter mentions having shorter delays and narrower time spans than arXiv downloads. We also find that the volume of Twitter mentions is statistically correlated with arXiv downloads and early citations just months after the publication of a preprint, with a possible bias that favors highly mentioned articles.

cs.SI↗

arXiv2012

A measure of total research impact independent of time and discipline

Authorship and citation practices evolve with time and differ by academic discipline. As such, indicators of research productivity based on citation records are naturally subject to historical and disciplinary effects. We observe these effects on a corpus of astronomer career data constructed from a database of refereed publications. We employ a simple mechanism to measure research output using author and reference counts available in bibliographic databases to develop a citation-based indicator of research productivity. The total research impact (tori) quantifies, for an individual, the total amount of scholarly work that others have devoted to his/her work, measured in the volume of research papers. A derived measure, the research impact quotient (riq), is an age independent measure of an individual's research ability. We demonstrate that these measures are substantially less vulnerable to temporal debasement and cross-disciplinary bias than the most popular current measures. The proposed measures of research impact, tori and riq, have been implemented in the Smithsonian/NASA Astrophysics Data System.

astro-ph.IM↗

arXiv2012

Do Linguistic Style and Readability of Scientific Abstracts affect their Virality?

Reactions to textual content posted in an online social network show different dynamics depending on the linguistic style and readability of the submitted content. Do similar dynamics exist for responses to scientific articles? Our intuition, supported by previous research, suggests that the success of a scientific article depends on its content, rather than on its linguistic style. In this article, we examine a corpus of scientific abstracts and three forms of associated reactions: article downloads, citations, and bookmarks. Through a class-based psycholinguistic analysis and readability indices tests, we show that certain stylistic and readability features of abstracts clearly concur in determining the success and viral capability of a scientific article.

cs.SI↗

arXiv2012

WorldWide Telescope in Research and Education

The WorldWide Telescope computer program, released to researchers and the public as a free resource in 2008 by Microsoft Research, has changed the way the ever-growing Universe of online astronomical data is viewed and understood. The WWT program can be thought of as a scriptable, interactive, richly visual browser of the multi-wavelength Sky as we see it from Earth, and of the Universe as we would travel within it. In its web API format, WWT is being used as a service to display professional research data. In its desktop format, WWT works in concert (thanks to SAMP and other IVOA standards) with more traditional research applications such as ds9, Aladin and TOPCAT. The WWT Ambassadors Program (founded in 2009) recruits and trains astrophysically-literate volunteers (including retirees) who use WWT as a teaching tool in online, classroom, and informal educational settings. Early quantitative studies of WWTA indicate that student experiences with WWT enhance science learning dramatically. Thanks to the wealth of data it can access, and the growing number of services to which it connects, WWT is now a key linking technology in the Seamless Astronomy environment we seek to offer researchers, teachers, and students alike.

astro-ph.IM↗

arXiv2011

The ADS All-Sky Survey

The ADS All-Sky Survey (ADSASS) is an ongoing effort aimed at turning the NASA Astrophysics Data System (ADS), widely known for its unrivaled value as a literature resource for astronomers, into a data resource. The ADS is not a data repository per se, but it implicitly contains valuable holdings of astronomical data, in the form of images, tables and object references contained within articles. The objective of the ADSASS effort is to extract these data and make them discoverable and available through existing data viewers. The resulting ADSASS data layer promises to greatly enhance workflows and enable new research by tying astronomical literature and data assets into one resource.

astro-ph.IM↗

arXiv2011

One, None and One Hundred Thousand Profiles: Re-imagining the Pirandellian Identity Dilemma in the Era of Online Social Networks

Uno, Nessuno, Centomila ("One, No One and One Hundred Thousand") is a classic novel by Italian playwright Luigi Pirandello. Published in 1925, it recounts the tragedy of Vitangelo Moscarda, a man who struggles to reclaim a coherent and unitary identity for himself in the face of an inherently social and multi-faceted world. What would Moscarda identity tragedy look like today? In this article we transplant Moscarda's identity play from its offline setting to the contemporary arena of social media and online social networks. With reference to established theories on identity construction, performance, and self-presentation, we re-imagine how Moscarda would go about defending the integrity of his selfhood in the face of the discountenancing influences of the online world.

cs.SI↗

arXiv2011

The use of microblogging for field-based scientific research

Documenting the context in which data are collected is an integral part of the scientific research lifecycle. In field-based research, contextual information provides a detailed description of scientific practices and thus enables data interpretation and reuse. For field data, losing contextual information often means losing the data altogether. Yet, documenting the context of distributed, collaborative, field-based research can be a significant challenge due to the unpredictable nature of real-world settings and to the high degree of variability in data collection methods and scientific practices of different researchers. In this article, we propose the use of microblogging as a mechanism to support collection, ingestion, and publication of contextual information about the variegated digital artifacts that are produced in field research. We perform interviews with scholars involved in field-based environmental and urban sensing research, to determine the extent of adoption of Twitter and similar microblogging platforms and their potential use for field-specific research applications. Based on the results of these interviews as well as participant observation of field activities, we present the design, development, and pilot evaluation of a microblogging application integrated with an existing data collection platform on a handheld device. We investigate whether microblogging accommodates the variable and unpredictable nature of highly mobile research and whether it represents a suitable mechanism to document the context of field research data early in the scientific information lifecycle.

cs.SI↗

arXiv2011

The relationship between acquaintanceship and coauthorship in scientific collaboration networks

This article examines the relationship between acquaintanceship and coauthorship patterns in a multi-disciplinary, multi-institutional, geographically distributed research center. Two social networks are constructed and compared: a network of coauthorship, representing how researchers write articles with one another, and a network of acquaintanceship, representing how those researchers know each other on a personal level, based on their responses to an online survey. Statistical analyses of the topology and community structure of these networks point to the importance of small-scale, local, personal networks predicated upon acquaintanceship for accomplishing collaborative work in scientific communities.

cs.CY↗

arXiv2011

Political protest Italian-style: The dissonance between the blogosphere and mainstream media in the promotion and coverage of Beppe Grillo's V-day

We analyze the organization, promotion and public perception of V-day, a political rally that took place on September 8, 2007, to protest against corruption in the Italian Parliament. Launched by blogger Beppe Grillo, and promoted via a word of mouth mobilization on the Italian blogosphere, V-day brought close to one million Italians in the streets on a single day, but was mostly ignored by mainstream media. This article is divided into two parts. In the first part, we analyze the volume and content of online articles published by both bloggers and mainstream news sources from June 14 (the day V-day was announced) until September 15, 2007 (one week after it took place) . We find that the success of V-day can be attributed to the coverage of bloggers and small-scale local news outlets only, suggesting a strong grassroots component in the organization of the rally. We also find a dissonant thematic relationship between content published by blogs and mainstream media: while the majority of blogs analyzed promote V-day, major mainstream media sources critique the methods of information production and dissemination employed by Grillo. Based on this finding, in the second part of the study, we explore the role of Grillo in the organization of the rally from a network analysis perspective. We study the interlinking structure of the V-day blogosphere network, to determine its structure, its levels of heterogeneity, and resilience. Our analysis contradicts the hypothesis that Grillo served as a top-down, broadcast-like source of information. Rather, we find that information about V-day was transferred across heterogeneous nodes in a moderately robust and resilient core network of blogs. We speculate that the organization of V-day represents the very first case, in Italian history, of a political demonstration developed and promoted primarily via the use of social media on the web.

cs.SI↗

arXiv2010

The Dilated Triple

The basic unit of meaning on the Semantic Web is the RDF statement, or triple, which combines a distinct subject, predicate and object to make a definite assertion about the world. A set of triples constitutes a graph, to which they give a collective meaning. It is upon this simple foundation that the rich, complex knowledge structures of the Semantic Web are built. Yet the very expressiveness of RDF, by inviting comparison with real-world knowledge, highlights a fundamental shortcoming, in that RDF is limited to statements of absolute fact, independent of the context in which a statement is asserted. This is in stark contrast with the thoroughly context-sensitive nature of human thought. The model presented here provides a particularly simple means of contextualizing an RDF triple by associating it with related statements in the same graph. This approach, in combination with a notion of graph similarity, is sufficient to select only those statements from an RDF graph which are subjectively most relevant to the context of the requesting process.

cs.AI↗

arXiv2009

Collaboration in sensor network research: an in-depth longitudinal analysis of assortative mixing patterns

Many investigations of scientific collaboration are based on statistical analyses of large networks constructed from bibliographic repositories. These investigations often rely on a wealth of bibliographic data, but very little or no other information about the individuals in the network, and thus, fail to illustrate the broader social and academic landscape in which collaboration takes place. In this article, we perform an in-depth longitudinal analysis of a relatively small network of scientific collaboration (N = 291) constructed from the bibliographic record of a research center involved in the development and application of sensor network and wireless technologies. We perform a preliminary analysis of selected structural properties of the network, computing its range, configuration and topology. We then support our preliminary statistical analysis with an in-depth temporal investigation of the assortative mixing of selected node characteristics, unveiling the researchers' propensity to collaborate preferentially with others with a similar academic profile. Our qualitative analysis of mixing patterns offers clues as to the nature of the scientific community being modeled in relation to its organizational, disciplinary, institutional, and international arrangements of collaboration.

physics.data-an↗