Search

Recent advances in generative language models, such as ChatGPT, have demonstrated an uncanny ability to produce texts that appear to be comparable to those produced by humans. Several key empirical results related to human processing of language, such as analogical reasoning, have been replicated using these models. Nevertheless, there are some important differences between the language generated by these models and language produced by humans. In this paper, I examine how LLMs performs on two pronoun disambiguation tasks reported on by Rhode, Levy, and Kehler (2011) and Sagi and Rips (2014). While LLMs performed reasonably in these tasks, their responses demonstrate stronger language-based biases while the influence of world knowledge, such as causal relationships, was lessened. Because LLMs replicate language produced by humans, these results can help shed light on which aspects of language use are directly encoded in language and which require additional reasoning faculties beyond language processing.

Edge computing promises to decentralize cloud applications while providing more bandwidth and reducing latency. Based on the discussion of our first Dagstuhl Seminar and the continuation work that took place after the seminar, we continued our work on identified challenges that need to be further addressed within the community. These challenges included 1) large-scale deployment of the edge-cloud continuum, 2) energy optimization and sustainability of such large-scale AI/ML learning and modelling, and 3) trustworthiness, security, and ethical questions related to the whole continuum. In this seminar, we discussed the current state of Edge Intelligence and shaped a holistic view of its challenges and applications. The main concerns were 1) the assessment and applicability of Edge Intelligence solutions, 2) energy consumption and sustainability, and 3) the new trend of Large-Language Models.

Living in dynamic environments such as the social domain, where interaction with others determines the reproductive success of individuals, requires the ability to recognize opportunities to obtain natural rewards and cope with challenges that are associated with achieving them. As such, actions that promote survival and reproduction are reinforced by the brain reward system, whereas coping with the challenges associated with obtaining these rewards is mediated by stress-response pathways, the activation of which can impair health and shorten lifespan. While much research has been devoted to understanding mechanisms underlying the way by which natural rewards are processed by the reward system, less attention has been given to the consequences of failure to obtain a desirable reward. As a model system to study the impact of failure to obtain a natural reward, we used the well-established courtship suppression paradigm in Drosophila melanogaster as means to induce repeated failures to obtain sexual reward in male flies. We discovered that beyond the known reduction in courtship actions caused by interaction with non-receptive females, repeated failures to mate induce a stress response characterized by persistent motivation to obtain the sexual reward, reduced male-male social interaction, and enhanced aggression. This frustrative-like state caused by the conflict between high motivation to obtain sexual reward and the inability to fulfill their mating drive impairs the capacity of rejected males to tolerate stressors such as starvation and oxidative stress. We further show that sensitivity to starvation and enhanced social arousal is mediated by the disinhibition of a small population of neurons that express receptors for the fly homologue of neuropeptide Y. Our findings demonstrate for the first time the existence of social stress in flies and offers a framework to study mechanisms underlying the crosstalk between reward, stress, and reproduction in a simple nervous

"The engineered perimeter around the occupied Gaza Strip is formed by a sophisticated system of fences, forts and surveillance technologies. With each Israeli incursion, a military no-go area, or a 'buffer zone', is established along Gaza's 'borders', extending deep into Palestinian residential areas and farmlands— further compounding the Gaza Strip's isolation from the rest of Palestine. Since 2014, the bulldozing of Palestinian lands by the Israeli occupation forces has been complemented by unannounced aerial spraying of military herbicides, extending the reach of Israeli violence into the realm of chemical warfare. Today, the spraying has destroyed entire swaths of arable land in Gaza, forcibly changing a once-lush Palestinian landscape, and providing the Israeli army with better visibility to fire at Palestinian targets with lethal force from a distance. This book is a vivid document of this latest stage of Israeli warfare, including original maps, images and visualisations which deepen our understanding of its environmental and human impact. It collects new documents, original archival materials, stills of drone footage, first-hand testimonies of farmers, organisers and protesters, and documents affected vegetation in Gaza as 'silent witnesses' to Israeli settler-colonial violence." --Provided by publisher., Foreword by Eyal Weizman -- Introduction: Colonial Legacies of Farm Warfare -- Chapter 1: Reconfiguration of Biopolitical Landscapes in Palestine -- Chapter 2: The Disappearing Trees of Gaza -- Chapter 3: The Al-Shawwa Citrus Export Company -- Chapter 4: Israeli Herbicidal Warfare as Colonial Practice -- Conclusion: Sowing New Landscapes of Resistance, "The engineered perimeter around the occupied Gaza Strip is formed by a sophisticated system of fences, forts and surveillance technologies. With each Israeli incursion, a military no-go area, or a 'buffer zone', is established along Gaza's 'borders', extending deep into Palestinian residential areas and farmlands— further compounding the Gaza Strip's isolation from the rest of Palestine. Since 2014, the bulldozing of Palestinian lands by the Israeli occupation forces has been complemented by unannounced aerial spraying of military herbicides, extending the reach of Israeli violence into the realm of chemical warfare. Today, the spraying has destroyed entire swaths of arable land in Gaza, forcibly changing a once-lush Palestinian landscape, and providing the Israeli army with better visibility to fire at Palestinian targets with lethal force from a distance. This book is a vivid document of this latest stage of Israeli warfare, including original maps, images and visualisations which deepen our understanding of its environmental and human impact. It collects new documents, original archival materials, stills of drone footage, first-hand testimonies of farmers, organisers and protesters, and documents affected vegetation in Gaza as 'silent witnesses' to Israeli settler-colonial violence." --Provided by publisher.

Purpose: Astigmatism blurs the retinal image of a circular spot along a particular orientation rendering it an elliptical shape. Astigmatic patients demonstrate adaptation to residual astigmatic blur that may affect their discrimination between oval and circular targets. The Wilkins Egg and Ball Test (WEBT) was created to detect altered visual perception due to residual astigmatic blur by discriminating a circle within a row of oval elements. This prospective, cross-sectional study examined the utility of WEBT in detecting uncorrected residual astigmatism on the perception of form symmetry in astigmatic and keratoconic participants as well as normal participants with induced astigmatism at four primary meridians. Methods: The mean search time (sT) and number of errors (noE) of 33 non-astigmatic controls (mean age: 24±5, range: 18–43, 6 males), 23 astigmatic participants (mean age: 36±12, range: 18–43, 6 males) and 13 keratoconic participants (N=22 eyes, mean age: 36 ±12, range: 18–58, 6 males) were measured under baseline, and 2.00 DC induced cylinder at four primary meridians, and for uncorrected, spherical-correction only, and fully corrected conditions, respectively. Mean sT and noE were converted to Z-scores, combined for each condition, and compared using repeated measures ANOVA with post-hoc analysis. Results: Combined Z-scores for the controls were significantly worse (p<0.001) for all induced cylinder conditions. The induced 180° condition was significantly better than 45° and 90° conditions (p=0.04), but not the 135° condition. For both astigmatic and keratoconic cohorts, Z-scores of the uncorrected condition were significantly worse than the fully corrected condition (both p<0.01), but the fully corrected and spherical-only conditions did not differ significantly (p=0.06 and p=0.05, respectively). Conclusion: In accommodating young adults, WEBT detected altered visual perception due to overall blur, and moderate-high amounts of uncorrected induced astigmat, This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. Parts of this study were funded by the Hadassah Academic College Conference and Research Board grant awarded to Eyal Gal., Peer Reviewed, Postprint (published version)

Asaf Achiron,1 Tal Yahalomi,2 Amit Biran,1 Eliya Levinger,1 Eyal Cohen,1 Uri Elbaz,3 Asim Ali,4,5 Kamiar Mireskandari,4,5 Raimo Tuuminen,6,7 Oleksiy V Voytsekhivskyy8 1Tel Aviv Sourasky Medical Center, Tel Aviv, Israel and Sackler School of Medicine, Tel Aviv University, Tel Aviv, Israel; 2Department of Ophthalmology, Samson Assuta Ashdod Hospital and the Faculty of Health Sciences, Ben-Gurion University of the Negev, Be’er Sheva, Israel; 3Department of Ophthalmology, Rabin Medical Center, Petach-Tikva and Sackler School of Medicine, Tel Aviv University, Tel Aviv, Israel; 4Department of Ophthalmology and Vision Sciences, University of Toronto, Toronto, Ontario, Canada; 5Department of Ophthalmology and Vision Sciences, the Hospital for Sick Children, Toronto, Ontario, Canada; 6Helsinki Retina Research Group, Faculty of Medicine, University of Helsinki, Helsinki, Finland; 7Department of Ophthalmology, Kymenlaakso Central Hospital, Kotka, Finland; 8Kyiv Clinical Ophthalmology Hospital Eye Microsurgery Center, Kyiv, UkraineCorrespondence: Oleksiy V Voytsekhivskyy, Kyiv Clinical Ophthalmology Hospital Eye Microsurgery Center, Komarov Ave. 3, Medical City, Kyiv, 03680, Ukraine, Tel +38067-584-28-11, Email iolpower@ukr.netPurpose: To compare the accuracy of 16 intraocular lens (IOL) power calculation formulas in pediatric cataract eyes.Patients and Methods: The data records of pediatric patients who had been implanted with three IOL models (SA60AT, MA60AC, and enVista-MX60) between 2012 and 2018 were analyzed. The accuracy of 16 IOL power calculation methods was evaluated: Barrett Universal II (BUII), Castrop, EVO 2.0, Haigis, Hill-RBF 3.0, Hoffer Q, Hoffer QST, Holladay 1, Kane, LSF AI, Naeser 2, Pearl-DGS, SRK/T, T2, VRF, and VRF-G. The non-optimized (ULIB/IOLcon) and optimized constants were used for IOL power calculation. The mean prediction error (PE), Performance Index (FPI), and all descriptive statistics were calculated.Results: Ninety-seven eyes of 97 pediatric

Alkaloids play an essential role in protecting plants against herbivores. Humans can also benefit from the pharmacological effects of these compounds. Plants produce an immense variety of structurally different alkaloids, including quinolizidine alkaloids, a group of bi-, tri-, and tetracyclic compounds produced by Lupinus species. Various lupin species produce different alkaloid profiles. To study the composition of quinolizidine alkaloids in lupin seeds, we collected 31 populations of two wild species native to Israel, L. pilosus and L. palaestinus, and analyzed their quinolizidine alkaloid contents. Our goal was to study the alkaloid profiles of these two wild species to better understand the challenges and prospective uses of wild lupins. We compared their profiles with those of other commercial and wild lupin species. To this end, a straightforward method for extracting alkaloids from seeds and determining the quinolizidine alkaloid profile by LC–MS/MS was developed and validated in-house. For the quantification of quinolizidine alkaloids, 15 analytical reference standards were used. We used GC–MS to verify and cross-reference the identity of certain alkaloids for which no analytical standards were available. The results enabled further exploration of quinolizidine alkaloid biosynthesis. We reviewed and re-analyzed the suggested quinolizidine alkaloid biosynthesis pathway, including the relationship between the amino acid precursor l-lysine and the different quinolizidine alkaloids occurring in seeds of lupin species. Revealing alkaloid compositions and highlighting some aspects of their formation pathway are important steps in evaluating the use of wild lupins as a novel legume crop.

A principal incentivizes a group of agents to work by choosing a monitoring structure and a scheme of performance-contingent rewards. The monitoring structure partitions the set of agents into monitoring teams, each delivering a signal of joint performance. We show that unlike under partial implementation, the principal always exhausts her monitoring capacity to optimally implement work as a unique outcome. Optimal monitoring teams are homogeneous between them: equally sized and with agents allocated in an anti-assortative fashion. Higher-effort-cost agents receive lower rents, and they tend to be monitored more closely than lower-effort-cost agents when the principal's allocation is constrained.

Edge artificial intelligence (AI) is an innovative computing paradigm that aims to shift the training and inference of machine learning models to the edge of the network. This paradigm offers the opportunity to significantly impact our everyday lives with new services such as autonomous driving and ubiquitous personalized health care. Nevertheless, bringing intelligence to the edge involves several major challenges, which include the need to constrain model architecture designs, the secure distribution and execution of the trained models, and the substantial network load required to distribute the models and data collected for training. In this article, we highlight key aspects in the development of edge AI in the past and connect them to current challenges. This article aims to identify research opportunities for edge AI, relevant to bring together the research in the fields of artificial intelligence and edge computing., QC 20240828

This paper examines the distinct effects of linguistics distance and language literacy on the labor market integration of migrant men and women. Using data from the Programme for International Assessment of Adult Competencies (PIAAC) 2018 in 16 countries of destination mainly from Europe and more than 110 languages of origin, we assess migrant labor force participation, employment, working hours, and occupational prestige. The study finds that linguistics distance of the first language studied has a significant negative association with labor force participation, employment, and working hours of migrant women, even after controlling for their abilities in their destination language, education, and cultural distance between the country of origin and destination. In contrast, linguistics distance is only negatively associated with migrant men's working hours. This suggests that linguistic distance serves as a proxy for cultural aspects, which are not captured by cultural distance and hence shape the labor market integration of migrant women due to cultural factors rather than human capital. We suggest that the gender aspect of the effect of language proximity is essential in understanding the intersectional position of migrant women in the labor force.

Emotion Focused Group Therapy has demonstrated preliminary efficacy for multiple syndromes. The groups involve individual chair-work in the context of the group, followed by interpersonal processing of the aroused emotions. So far there has been limited investigation of the mechanisms of change, including participants’ experiences of the change process. This study aimed to qualitatively examine participants’ experiences of the change process during Emotion Focused Group Therapy. Sixteen participants were interviewed about their experiences of the change process during the therapy group. The interview was structured to move from overarching reflections on the process of change, to specific experiences of their own chair-work and watching others’ chair-work. During the interview, participants were shown salient video segments of the group to prime their recall of moments of high arousal that were deemed likely to contain emotional change processes. Thematic Analysis extracted five main themes: Value of Vulnerability; Connectedness; Vicarious Emotional Processing; Understanding My Critic; and Barriers to Emotional Processing. These findings suggest that there may be both common and unique change processes that occur during Emotion Focused Group Therapy, compared with the individual Emotion Focused Therapy.

We study binary action network games with strategic complementarities. An agent acts if the aggregate social influence of her friends exceeds a transfer levied on the agent by a principal. The principal seeks to maximize her revenue while inducing everyone to act in a unique equilibrium. We characterize optimal transfers showing that agents who are more popular than their friends receive preferential treatment. Our main result is that under mild conditions complete core‐periphery networks deliver the highest revenue to the principal. Furthermore, we show that the revenue is higher in networks where links are allocated unequally across agents. Hence, the principal benefits from creating “influentials” by linking well‐connected hubs to less popular periphery.

The Mediterranean Sea has been sampled irregularly by research vessels in the past, mostly by national expeditions in regional waters. To monitor the hydrographic, biogeochemical and circulation changes in the Mediterranean Sea, a systematic repeat oceanographic survey programme called Med-SHIP was recommended by the Mediterranean Science Commission (CIESM) in 2011, as part of the Global Ocean Ship-based Hydrographic Investigations Program (GO-SHIP). Med-SHIP consists of zonal and meridional surveys with different frequencies, where comprehensive physical and biogeochemical properties are measured with the highest international standards. The first zonal survey was done in 2011 and repeated in 2018. In addition, a network of meridional (and other key) hydrographic sections were designed: the first cycle of these sections was completed in 2016, with three cruises funded by the EU project EUROFLEETS2. This paper presents the physical and chemical data of the meridional and key transects in the Western and Eastern Mediterranean Sea collected during those cruises.

The Earth surface Mineral dust source InvesTigation (EMIT) is a remote visible to shortwave infrared (VSWIR) imaging spectrometer that has been operating onboard the International Space Station since July 2022. This article describes EMIT's on-orbit spectroradiometric calibration and validation. Accurate spectroscopy is vital to achieve consistent mapping results with orbital imaging spectrometers. EMIT takes a unique approach to this challenge, with just six optical elements, no shutter, and no onboard calibration systems. Its simple design focuses on uniformity and stability to enable vicarious spectroradiometric calibration. Our experiments demonstrate that this approach is successful, approaching the fidelity of manual field spectroscopy in some cases, and enabling new and more accurate products across diverse Earth science disciplines. EMIT achieves several notable firsts for an instrument of its class. It demonstrates successful on-orbit adjustments of Focal Plane Array (FPA) alignment with sub-micron precision. It offers spectral uniformity better than 98%. Optical artifacts in the measurement channels are at least three orders of magnitude below the primary solar-reflected surface signals. Its noise performance enables percent-level discrimination in the depths of mineral absorption features. In these aspects, EMIT satisfies the stringent performance needs for the next generation of VSWIR imaging spectrometers to observe the Earth's ecosystems, geology, and water resources., EMIT is supported by the National Aeronautics and Space Administration Earth Venture Instrument program, under the Earth Science Division of the Science Mission Directorate. K. Dana Chadwick is supported by the NASA Applied Sciences Program. Carlos P´ erez García- Pando and María Gonçalves Ageitos acknowledge support from the European Research Council (ERC) Consolidator Grant FRAGMENT (grant agreement No. 773051), and the AXA Chair on Sand and Dust Storms at the Barcelona Supercomputing Center funded by the AXA Research Fund. Martina Klose has received funding through the Helmholtz Association’s Initiative and Networking Fund (grant agreement No. VH-NG-1533). We thank Jeffrey Czapla-Myers and the University of Arizona team for their maintenance and operation of the Railroad Valley automated calibration facility. This research was performed at the Jet Propulsion Laboratory, California Institute of Technology, under a contract with the National Aeronautics and Space Administration. We acknowledge the support and assistance of NASA’s International Space Station Program. The USGS authors’ contribution to this published Work was prepared by U.S. federal government employees as part of their official duties and constitutes a “work of the United States government,” and is considered to be in the public domain and therefore domestic copyright does not apply. Any use of trade, firm, or product names is for descriptive purposes only and does not imply endorsement by the U.S. Government. Copyright 2024 California Institute of Technology. All rights reserved. US Government Support Acknowledged., Peer Reviewed, Article signat per 56 autors: David R. Thompson, Robert O. Green, Christine Bradley, Philip G. Brodrick, Natalie Mahowald, Eyal Ben Dor, Matthew Bennett, Michael Bernas, Nimrod Carmon, K. Dana Chadwick, Roger N. Clark, Red Willow Coleman, Evan Cox, Ernesto Diaz, Michael L. Eastwood, Regina Eckert, Bethany L. Ehlmann, Paul Ginoux, María Gonçalves Ageitos, Kathleen Grant, Luis Guanter, Daniela Heller Pearlshtien, Mark Helmlinger, Harrison Herzog, Todd Hoefen, Yue Huang, Abigail Keebler, Olga Kalashnikova, Didier Keymeulen, Raymond Kokaly, Martina Klose, Longlei Li, Sarah R. Lundeen, John Meyer, Elizabeth Middleton, Ron L. Miller, Pantazis Mouroulis, Bogdan Oaida, Vincenzo Obiso, Francisco Ochoa, Winston Olson-Duvall, Gregory S. Okin, Thomas H. Painter, Carlos Pérez García-Pando, Randy Pollock, Vincent Realmuto, Lucas Shaw, Peter Sullivan, Gregg Swayze, Erik Thingvold, Andrew K. Thorpe, Suresh Vannan, Catalina Villarreal, Charlene Ung, Daniel W. Wilson, Sander Zandbergen., Objectius de Desenvolupament Sostenible::13 - Acció per al Clima, Postprint (published version)

The Mediterranean Sea has been sampled irregularly by research vessels in the past, mostly by national expeditions in regional waters. To monitor the hydrographic, biogeochemical and circulation changes in the Mediterranean Sea, a systematic repeat oceanographic survey programme called Med-SHIP was recommended by the Mediterranean Science Commission (CIESM) in 2011, as part of the Global Ocean Ship-based Hydrographic Investigations Program (GO-SHIP). Med-SHIP consists of zonal and meridional surveys with different frequencies, where comprehensive physical and biogeochemical properties are measured with the highest international standards. The first zonal survey was done in 2011 and repeated in 2018. In addition, a network of meridional (and other key) hydrographic sections were designed: the first cycle of these sections was completed in 2016, with three cruises funded by the EU project EUROFLEETS2. This paper presents the physical and chemical data of the meridional and key transects in the Western and Eastern Mediterranean Sea collected during those cruises.

Inherited macular dystrophies (iMDs) are a group of genetic disorders, which affect the central region of the retina. To investigate the genetic basis of iMDs, we used single-molecule Molecular Inversion Probes to sequence 105 maculopathy-associated genes in 1352 patients diagnosed with iMDs. Within this cohort, 39.8% of patients were considered genetically explained by 460 different variants in 49 distinct genes of which 73 were novel variants, with some affecting splicing. The top five most frequent causative genes were ABCA4 (37.2%), PRPH2 (6.7%), CDHR1 (6.1%), PROM1 (4.3%) and RP1L1 (3.1%). Interestingly, variants with incomplete penetrance were revealed in almost one-third of patients considered solved (28.1%), and therefore, a proportion of patients may not be explained solely by the variants reported. This includes eight previously reported variants with incomplete penetrance in addition to CDHR1:c.783G>A and CNGB3:c.1208G>A. Notably, segregation analysis was not routinely performed for variant phasing-a limitation, which may also impact the overall diagnostic yield. The relatively high proportion of probands without any putative causal variant (60.2%) highlights the need to explore variants with incomplete penetrance, the potential modifiers of disease and the genetic overlap between iMDs and age-related macular degeneration. Our results provide valuable insights into the genetic landscape of iMDs and warrant future exploration to determine the involvement of other maculopathy genes.

The Mediterranean Sea has been sampled irregularly by research vessels in the past, mostly by national expeditions in regional waters. To monitor the hydrographic, biogeochemical and circulation changes in the Mediterranean Sea, a systematic repeat oceanographic survey programme called Med-SHIP was recommended by the Mediterranean Science Commission (CIESM) in 2011, as part of the Global Ocean Ship-based Hydrographic Investigations Program (GO-SHIP). Med-SHIP consists of zonal and meridional surveys with different frequencies, where comprehensive physical and biogeochemical properties are measured with the highest international standards. The first zonal survey was done in 2011 and repeated in 2018. In addition, a network of meridional (and other key) hydrographic sections were designed: the first cycle of these sections was completed in 2016, with three cruises funded by the EU project EUROFLEETS2. This paper presents the physical and chemical data of the meridional and key transects in the Western and Eastern Mediterranean Sea collected during those cruises.

Liat Gantz,1 Joan Gispets,2 Arnold Wilkins,3 Eyal Gal1,2 1Department of Optometry and Vision Science, Hadassah Academic College, Jerusalem, Israel; 2University Vision Centre, Universitat Politècnica de Catalunya, Terrassa, Barcelona, Spain; 3Department of Psychology, University of Essex, Colchester, UKCorrespondence: Liat Gantz, Email liatg@hac.ac.il

Eyal Gal,1,2 Joan Gispets,2 Arnold Wilkins,3 Meira Zyroff,1 Efrat Netanya‌,1 Liat Gantz1 1Department of Optometry and Vision Science, Hadassah Academic College, Jerusalem, Israel; 2University Vision Centre, Universitat Politècnica de Catalunya, Terrassa, Barcelona, Spain; 3Department of Psychology, University of Essex, Colchester, UKCorrespondence: Liat Gantz; Eyal Gal, Department of Optometry and Vision Science, Hadassah Academic College, 37 Hanevi’im St, Jerusalem, Jerusalem, 9101001, Israel, Email liatg@hac.ac.il; gale@hac.ac.ilPurpose: Astigmatism blurs the retinal image of a circular spot along a particular orientation rendering it an elliptical shape. Astigmatic patients demonstrate adaptation to residual astigmatic blur that may affect their discrimination between oval and circular targets. The Wilkins Egg and Ball Test (WEBT) was created to detect altered visual perception due to residual astigmatic blur by discriminating a circle within a row of oval elements. This prospective, cross-sectional study examined the utility of WEBT in detecting uncorrected residual astigmatism on the perception of form symmetry in astigmatic and keratoconic participants as well as normal participants with induced astigmatism at four primary meridians.Methods: The mean search time (sT) and number of errors (noE) of 33 non-astigmatic controls (mean age: 24± 5, range: 18– 43, 6 males), 23 astigmatic participants (mean age: 36± 12, range: 18– 43, 6 males) and 13 keratoconic participants (N=22 eyes, mean age: 36± 12, range: 18– 58, 6 males) were measured under baseline, and 2.00 DC induced cylinder at four primary meridians, and for uncorrected, spherical-correction only, and fully corrected conditions, respectively. Mean sT and noE were converted to Z-scores, combined for each condition, and compared using repeated measures ANOVA with post-hoc analysis.Results: Combined Z-scores for the controls were significantly worse (p< 0.001) for all induced cylinder condition

The neuropeptide hormone oxytocin and the steroid hormone testosterone have received attention as modulators of behavior in the context of intergroup conflict. However, to date, their interactive effect has yet to be tested. Here, in a double-blind placebo-control design, 204 participants (102 female participants) self-administrated oxytocin or placebo and completed an experimental economic game modeling intergroup conflict. Salivary testosterone (n = 192) was measured throughout the task to assess endogenous reactivity. As a caveat, even at this sample size, our derived power to detect small effects for 2- and 3-way interactions was relatively low. For male participants, changes in testosterone predicted willingness to sacrifice investments for the betterment of the group. Intranasal administration of oxytocin strongly diminished this effect. In female participants, we found no credible evidence for association between changes in testosterone and investments, rather, oxytocin effects were independent of testosterone. This 3-way interaction was of medium to large effect size (Odds Ratio 5.11). Behavior was also affected by social cues such as signaling of ingroup and outgroup members. Our findings provide insights as to the biological processes underpinning parochial altruism and suggest an additional path for the dual influence of oxytocin and testosterone on human social behavior.

Standard temporal planning assumes that planning takes place offline and then execution starts at time 0. Recently, situated temporal planning was introduced, where planning starts at time 0 and execution occurs after planning terminates. Situated temporal planning reflects a more realistic scenario where time passes during planning. However, in situated temporal planning a complete plan must be generated before any action is executed. In some problems with time pressure, timing is too tight to complete planning before the first action must be executed. For example, an autonomous car that has a truck backing towards it should probably move out of the way now and plan how to get to its destination later. In this paper, we propose a new problem setting: concurrent planning and execution, in which actions can be dispatched (executed) before planning terminates. Unlike previous work on planning and execution, we must handle wall clock deadlines that affect action applicability and goal achievement (as in situated planning) while also supporting dispatching actions before a complete plan has been found. We extend previous work on metareasoning for situated temporal planning to develop an algorithm for this new setting. Our empirical evaluation shows that when there is strong time pressure, our approach outperforms situated temporal planning.

Despite it being the cornerstone of BPE, the most common tokenization algorithm, the importance of compression in the tokenization process is still unclear. In this paper, we argue for the theoretical importance of compression, that can be viewed as 0-gram language modeling where equal probability is assigned to all tokens. We also demonstrate the empirical importance of compression for downstream success of pre-trained language models. We control the compression ability of several BPE tokenizers by varying the amount of documents available during their training: from 1 million documents to a character-based tokenizer equivalent to no training data at all. We then pre-train English language models based on those tokenizers and fine-tune them over several tasks. We show that there is a correlation between tokenizers' compression and models' downstream performance, suggesting that compression is a reliable intrinsic indicator of tokenization quality. These correlations are more pronounced for generation tasks (over classification) or for smaller models (over large ones). We replicated a representative part of our experiments on Turkish and found similar results, confirming that our results hold for languages with typological characteristics dissimilar to English. We conclude that building better compressing tokenizers is a fruitful avenue for further research and for improving overall model performance.

Large language models hold significant promise in multilingual applications. However, inherent biases stemming from predominantly English-centric pre-training have led to the widespread practice of pre-translation, i.e., translating non-English inputs to English before inference, leading to complexity and information loss. This study re-evaluates the need for pre-translation in the context of PaLM2 models (Anil et al., 2023), which have been established as highly performant in multilingual tasks. We offer a comprehensive investigation across 108 languages and 6 diverse benchmarks, including open-end generative tasks, which were excluded from previous similar studies. Our findings challenge the pre-translation paradigm established in prior research, highlighting the advantages of direct inference in PaLM2. Specifically, PaLM2-L consistently outperforms pre-translation in 94 out of 108 languages. These findings pave the way for more efficient and effective multilingual applications, alleviating the limitations associated with pre-translation and unlocking linguistic authenticity.

We analyze the operation of transformer language adapters, which are small modules trained on top of a frozen language model to adapt its predictions to new target languages. We show that adapted predictions mostly evolve in the source language the model was trained on, while the target language becomes pronounced only in the very last layers of the model. Moreover, the adaptation process is gradual and distributed across layers, where it is possible to skip small groups of adapters without decreasing adaptation performance. Last, we show that adapters operate on top of the model's frozen representation space while largely preserving its structure, rather than on an 'isolated' subspace. Our findings provide a deeper view into the adaptation process of language models to new languages, showcasing the constraints imposed on it by the underlying model and introduces practical implications to enhance its efficiency., Comment: 18 pages

Queer youth face increased mental health risks, such as depression, anxiety, and suicidal ideation. Hindered by negative stigma, they often avoid seeking help and rely on online resources, which may provide incompatible information. Although access to a supportive environment and reliable information is invaluable, many queer youth worldwide have no access to such support. However, this could soon change due to the rapid adoption of Large Language Models (LLMs) such as ChatGPT. This paper aims to comprehensively explore the potential of LLMs to revolutionize emotional support for queers. To this end, we conduct a qualitative and quantitative analysis of LLM's interactions with queer-related content. To evaluate response quality, we develop a novel ten-question scale that is inspired by psychological standards and expert input. We apply this scale to score several LLMs and human comments to posts where queer youth seek advice and share experiences. We find that LLM responses are supportive and inclusive, outscoring humans. However, they tend to be generic, not empathetic enough, and lack personalization, resulting in nonreliable and potentially harmful advice. We discuss these challenges, demonstrate that a dedicated prompt can improve the performance, and propose a blueprint of an LLM-supporter that actively (but sensitively) seeks user context to provide personalized, empathetic, and reliable responses. Our annotated dataset is available for further research.

Model selection for a given target task can be costly, as it may entail extensive annotation of the quality of outputs of different models. We introduce DiffUse, an efficient method to make an informed decision between candidate text generation models based on preference annotations. DiffUse reduces the required amount of annotations, thus saving valuable time and resources in performing evaluation. DiffUse intelligently selects instances by clustering embeddings that represent the semantic differences between model outputs. Thus, it is able to identify a subset of examples that are more informative for preference decisions. Our method is model-agnostic, and can be applied to any text generation model for selecting between models, prompts and configurations. Moreover, we propose a practical iterative approach for dynamically determining how many instances to annotate. In a series of experiments over hundreds of model pairs, we demonstrate that DiffUse can dramatically reduce the required number of annotations -- by up to 75% -- while maintaining high evaluation reliability., Comment: Accepted to ACL (main conference)

Graph generation is a fundamental problem in various domains, including chemistry and social networks. Recent work has shown that molecular graph generation using recurrent neural networks (RNNs) is advantageous compared to traditional generative approaches which require converting continuous latent representations into graphs. One issue which arises when treating graph generation as sequential generation is the arbitrary order of the sequence which results from a particular choice of graph flattening method. In this work we propose using RNNs, taking into account the non-sequential nature of graphs by adding an Orderless Regularization (OLR) term that encourages the hidden state of the recurrent model to be invariant to different valid orderings present under the training distribution. We demonstrate that sequential graph generation models benefit from our proposed regularization scheme, especially when data is scarce. Our findings contribute to the growing body of research on graph generation and provide a valuable tool for various applications requiring the synthesis of realistic and diverse graph structures., Comment: 16 pages, 3 figures

Allergic diseases are a growing health concern in industrialized countries. Despite a number of effective therapeutic options for the prevention and treatment of the pathophysiologic responses which characterize allergic diseases, the induction of true allergen desensitization remains an elusive therapeutic goal. Only immunotherapy (IT) has been shown to have any effect on the underlying hypersensitivities which mediate allergic reactions, and traditional protein-based allergen IT has a limited scope of efficacy. However, a number of reagents collectively termed DNA‐based immunotherapeutics have proven highly effective in both the prevention and reversal of Th2-mediated hypersensitivity states in mouse models of allergic disease. Four basic DNA-based immunotherapeutic modalities have been used for these studies. These include immunization with gene vaccines, allergen mixed with immunostimulatory oligodeoxynucleotide (ISS-ODN), and physical allergen-ISS-ODN conjugates (AIC), as well as immunomodulation with ISS-ODN alone. Results from many laboratories have generated guarded optimism that DNA-based immunotherapeutics may be effective for the reversal of allergic hypersensitivity states in humans, and several clinical trials have already been initiated. This review will focus on our present understanding of the biological activities of DNA-based immunotherapeutics and their application to the treatment of allergic diseases., Institutos Nacionales de Salud, Dynavax Thecnologies Corporation, Depto. de Medicina, Fac. de Medicina, TRUE, pub

This study investigated whether immunostimulatory DNA sequences (ISS) induce a transient or sustained inhibition of Th2 responses to inhaled antigen. We sensitized mice with subcutaneous injections to develop a Th2 response to ovalbumin (ova) and then administered a dose of ISS prior to ova inhalation challenge. Mice were then rechallenged with ova by inhalation a second time at varying time points after the first ova inhalation (1 to 8 weeks later) to determine whether the ISS dose administered prior to the first ova inhalation protected against a subsequent second ova inhalation challenge. A single dose of ISS inhibited the Th2 response to the first inhalation of ova antigen, as well as 4 weeks later to the second inhalation of ova. However, ISS did not inhibit a Th2 response to the second inhalation of ova 8 weeks later. The reversible inhibition of Th2 responses at 8 weeks suggests the need for repeated ISS administration at monthly intervals., Depto. de Medicina, Fac. de Medicina, TRUE, pub

The lack of high-quality data for content-grounded generation tasks has been identified as a major obstacle to advancing these tasks. To address this gap, we propose Genie, a novel method for automatically generating high-quality content-grounded data. It consists of three stages: (a) Content Preparation, (b) Generation: creating task-specific examples from the content (e.g., question-answer pairs or summaries). (c) Filtering mechanism aiming to ensure the quality and faithfulness of the generated data. We showcase this methodology by generating three large-scale synthetic data, making wishes, for Long-Form Question-Answering (LFQA), summarization, and information extraction. In a human evaluation, our generated data was found to be natural and of high quality. Furthermore, we compare models trained on our data with models trained on human-written data -- ELI5 and ASQA for LFQA and CNN-DailyMail for Summarization. We show that our models are on par with or outperforming models trained on human-generated data and consistently outperforming them in faithfulness. Finally, we applied our method to create LFQA data within the medical domain and compared a model trained on it with models trained on other domains., Comment: Accepted to ICLR24

We introduce the first Natural Office Talkers in Settings of Far-field Audio Recordings (``NOTSOFAR-1'') Challenge alongside datasets and baseline system. The challenge focuses on distant speaker diarization and automatic speech recognition (DASR) in far-field meeting scenarios, with single-channel and known-geometry multi-channel tracks, and serves as a launch platform for two new datasets: First, a benchmarking dataset of 315 meetings, averaging 6 minutes each, capturing a broad spectrum of real-world acoustic conditions and conversational dynamics. It is recorded across 30 conference rooms, featuring 4-8 attendees and a total of 35 unique speakers. Second, a 1000-hour simulated training dataset, synthesized with enhanced authenticity for real-world generalization, incorporating 15,000 real acoustic transfer functions. The tasks focus on single-device DASR, where multi-channel devices always share the same known geometry. This is aligned with common setups in actual conference rooms, and avoids technical complexities associated with multi-device tasks. It also allows for the development of geometry-specific solutions. The NOTSOFAR-1 Challenge aims to advance research in the field of distant conversational speech recognition, providing key resources to unlock the potential of data-driven methods, which we believe are currently constrained by the absence of comprehensive high-quality training and benchmarking datasets., Comment: preprint

As instruction-tuned large language models (LLMs) gain global adoption, their ability to follow instructions in multiple languages becomes increasingly crucial. In this work, we investigate how multilinguality during instruction tuning of a multilingual LLM affects instruction-following across languages from the pre-training corpus. We first show that many languages transfer some instruction-following capabilities to other languages from even monolingual tuning. Furthermore, we find that only 40 multilingual examples integrated in an English tuning set substantially improve multilingual instruction-following, both in seen and unseen languages during tuning. In general, we observe that models tuned on multilingual mixtures exhibit comparable or superior performance in multiple languages compared to monolingually tuned models, despite training on 10x fewer examples in those languages. Finally, we find that diversifying the instruction tuning set with even just 2-4 languages significantly improves cross-lingual generalization. Our results suggest that building massively multilingual instruction-tuned models can be done with only a very small set of multilingual instruction-responses., Comment: Findings of ACL 2024

Large Language Models (LLMs) often do not perform well on queries that require the aggregation of information across texts. To better evaluate this setting and facilitate modeling efforts, we introduce TACT - Text And Calculations through Tables, a dataset crafted to evaluate LLMs' reasoning and computational abilities using complex instructions. TACT contains challenging instructions that demand stitching information scattered across one or more texts, and performing complex integration on this information to generate the answer. We construct this dataset by leveraging an existing dataset of texts and their associated tables. For each such tables, we formulate new queries, and gather their respective answers. We demonstrate that all contemporary LLMs perform poorly on this dataset, achieving an accuracy below 38\%. To pinpoint the difficulties and thoroughly dissect the problem, we analyze model performance across three components: table-generation, Pandas command-generation, and execution. Unexpectedly, we discover that each component presents substantial challenges for current LLMs. These insights lead us to propose a focused modeling framework, which we refer to as IE as a tool. Specifically, we propose to add "tools" for each of the above steps, and implement each such tool with few-shot prompting. This approach shows an improvement over existing prompting techniques, offering a promising direction for enhancing model capabilities in these tasks., Comment: Website (https://tact-benchmark.github.io), Huggingface (https://huggingface.co/datasets/google/TACT)

In this paper, we propose an efficient, fast, and versatile distillation method to accelerate the generation of pre-trained diffusion models: Flash Diffusion. The method reaches state-of-the-art performances in terms of FID and CLIP-Score for few steps image generation on the COCO2014 and COCO2017 datasets, while requiring only several GPU hours of training and fewer trainable parameters than existing methods. In addition to its efficiency, the versatility of the method is also exposed across several tasks such as text-to-image, inpainting, face-swapping, super-resolution and using different backbones such as UNet-based denoisers (SD1.5, SDXL) or DiT (Pixart-$\alpha$), as well as adapters. In all cases, the method allowed to reduce drastically the number of sampling steps while maintaining very high-quality image generation. The official implementation is available at https://github.com/gojasper/flash-diffusion., Comment: 16 pages + 16 pages appendices

A long-standing open conjecture of Branko Gr\"unbaum from 1972 states that any simple arrangement of $n$ pairwise intersecting pseudocircles in the plane can have at most $2n-2$ digons. Agarwal et al. proved this conjecture for arrangements of pairwise intersecting pseudocircles in which there is a common point surrounded by all pseudocircles. Recently, Felsner, Roch and Scheucher showed that Gr\"unbaum's conjecture is true for arrangements of pairwise intersecting pseudocircles in which there are three pseudocircles every pair of which create a digon. In this paper we prove this over 50-year-old conjecture of Gr\"unbaum for any simple arrangement of pairwise intersecting circles in the plane.

A central problem in quantum mechanics involves solving the Electronic Schrodinger Equation for a molecule or material. The Variational Monte Carlo approach to this problem approximates a particular variational objective via sampling, and then optimizes this approximated objective over a chosen parameterized family of wavefunctions, known as the ansatz. Recently neural networks have been used as the ansatz, with accompanying success. However, sampling from such wavefunctions has required the use of a Markov Chain Monte Carlo approach, which is inherently inefficient. In this work, we propose a solution to this problem via an ansatz which is cheap to sample from, yet satisfies the requisite quantum mechanical properties. We prove that a normalizing flow using the following two essential ingredients satisfies our requirements: (a) a base distribution which is constructed from Determinantal Point Processes; (b) flow layers which are equivariant to a particular subgroup of the permutation group. We then show how to construct both continuous and discrete normalizing flows which satisfy the requisite equivariance. We further demonstrate the manner in which the non-smooth nature ("cusps") of the wavefunction may be captured, and how the framework may be generalized to provide induction across multiple molecules. The resulting theoretical framework entails an efficient approach to solving the Electronic Schrodinger Equation.

Linearizable datastores are desirable because they provide users with the illusion that the datastore is run on a single machine that performs client operations one at a time. To reduce the performance cost of providing this illusion, many specialized algorithms for linearizable reads have been proposed which significantly improve read performance compared to write performance. The main difference between these specialized algorithms is their performance under different workloads. Unfortunately, since a datastore's workload is often unknown or changes over time and system designers must decide on a single read algorithm to implement ahead of time, a datastore's performance is often suboptimal as it cannot adapt to workload changes. In this paper, we lay the groundwork for addressing this problem by proposing Chameleon, an algorithm for linearizable reads that provides a principled approach for datastores to switch between existing read algorithms at runtime. The key observation that enables this generalization is that all existing algorithms are specific read-write quorum systems. Chameleon constructs a generic read-write quorum system, by using tokens that are included to complete write and read operations. This token quorum system enables Chameleon to mimic existing read algorithms and switch between them by transferring these tokens between processes., Comment: 7 pages, 4 figures

The risen complexity of automotive systems requires new development strategies and methods to master the upcoming challenges. Traditional methods need thus to be changed by an increased level of automation, and a faster continuous improvement cycle. In this context, current vehicle performance tests represent a very time-consuming and expensive task due to the need to perform the tests in real driving conditions. As a consequence, agile/iterative processes like DevOps are largely hindered by the necessity of triggering frequent tests. This paper reports on a practical experience of developing an AI-augmented solution based on Machine Learning and Model-based Engineering to support continuous vehicle development and testing. In particular, historical data collected in real driving conditions is leveraged to synthesize a high-fidelity driving simulator and hence enable performance tests in virtual environments. Based on this practical experience, this paper also proposes a conceptual framework to support predictions based on real driving behavior.

Dense vector representations for sentences made significant progress in recent years as can be seen on sentence similarity tasks. Real-world phrase retrieval applications, on the other hand, still encounter challenges for effective use of dense representations. We show that when target phrases reside inside noisy context, representing the full sentence with a single dense vector, is not sufficient for effective phrase retrieval. We therefore look into the notion of representing multiple, sub-sentence, consecutive word spans, each with its own dense vector. We show that this technique is much more effective for phrase mining, yet requires considerable compute to obtain useful span representations. Accordingly, we make an argument for contextualized word/token embeddings that can be aggregated for arbitrary word spans while maintaining the span's semantic meaning. We introduce a modification to the common contrastive loss used for sentence embeddings that encourages word embeddings to have this property. To demonstrate the effect of this method we present a dataset based on the STS-B dataset with additional generated text, that requires finding the best matching paraphrase residing in a larger context and report the degree of similarity to the origin phrase. We demonstrate on this dataset, how our proposed method can achieve better results without significant increase to compute.

When large language models are aligned via supervised fine-tuning, they may encounter new factual information that was not acquired through pre-training. It is often conjectured that this can teach the model the behavior of hallucinating factually incorrect responses, as the model is trained to generate facts that are not grounded in its pre-existing knowledge. In this work, we study the impact of such exposure to new knowledge on the capability of the fine-tuned model to utilize its pre-existing knowledge. To this end, we design a controlled setup, focused on closed-book QA, where we vary the proportion of the fine-tuning examples that introduce new knowledge. We demonstrate that large language models struggle to acquire new factual knowledge through fine-tuning, as fine-tuning examples that introduce new knowledge are learned significantly slower than those consistent with the model's knowledge. However, we also find that as the examples with new knowledge are eventually learned, they linearly increase the model's tendency to hallucinate. Taken together, our results highlight the risk in introducing new factual knowledge through fine-tuning, and support the view that large language models mostly acquire factual knowledge through pre-training, whereas fine-tuning teaches them to use it more efficiently.

Whisper is a state-of-the-art automatic speech recognition (ASR) model (Radford et al., 2022). Although Swiss German dialects are allegedly not part of Whisper's training data, preliminary experiments showed that Whisper can transcribe Swiss German quite well, with the output being a speech translation into Standard German. To gain a better understanding of Whisper's performance on Swiss German, we systematically evaluate it using automatic, qualitative, and human evaluation. We test its performance on three existing test sets: SwissDial (Dogan-Sch\"onberger et al., 2021), STT4SG-350 (Pl\"uss et al., 2023), and Swiss Parliaments Corpus (Pl\"uss et al., 2021). In addition, we create a new test set for this work, based on short mock clinical interviews. For automatic evaluation, we used word error rate (WER) and BLEU. In the qualitative analysis, we discuss Whisper's strengths and weaknesses and anylyze some output examples. For the human evaluation, we conducted a survey with 28 participants who were asked to evaluate Whisper's performance. All of our evaluations suggest that Whisper is a viable ASR system for Swiss German, so long as the Standard German output is desired., Comment: Accepted at VarDial 2024 (the eleventh Workshop on NLP for Similar Languages, Varieties and Dialects 2024), Mexico City

Background: Recent advancements in Artificial Intelligence (AI) contributed significantly to suicide assessment, however, our theoretical understanding of this complex behavior is still limited. Objective: This study aimed to harness AI methodologies to uncover hidden risk factors that trigger or aggravate suicide behaviors. Method: The primary dataset included 228,052 Facebook postings by 1,006 users who completed the gold-standard Columbia Suicide Severity Rating Scale. This dataset was analyzed using a bottom-up research pipeline without a-priory hypotheses and its findings were validated using a top-down analysis of a new dataset. This secondary dataset included responses by 1,062 participants to the same suicide scale as well as to well-validated scales measuring depression and boredom. Results: An almost fully automated, AI-guided research pipeline resulted in four Facebook topics that predicted the risk of suicide, of which the strongest predictor was boredom. A comprehensive literature review using APA PsycInfo revealed that boredom is rarely perceived as a unique risk factor of suicide. A complementing top-down path analysis of the secondary dataset uncovered an indirect relationship between boredom and suicide, which was mediated by depression. An equivalent mediated relationship was observed in the primary Facebook dataset as well. However, here, a direct relationship between boredom and suicide risk was also observed. Conclusions: Integrating AI methods allowed the discovery of an under-researched risk factor of suicide. The study signals boredom as a maladaptive 'ingredient' that might trigger suicide behaviors, regardless of depression. Further studies are recommended to direct clinicians' attention to this burdening, and sometimes existential experience.

The legal system is concerned with interpreting and applying existing laws to real world circumstances. In the U.S., the final arbiters of these laws are the nine justices of the U.S. Supreme Court. The cases presented before the court generally involve disputes between two parties, the petitioner and the respondent. However, in some cases third parties provide additional information to the court as amici curiae (‘friends of the court’). The present study uses corpus statistics and moral foundation theory to explore and compare the types of arguments brought up by the different parties during oral arguments over the past 6 decades, with a focus on amicus curiae arguments and how they affect the results of the case. Overall, the results show that both amici and respondents rely on moral arguments more frequently than petitioners do.

In permissionless consensus, the ordering of transactions or inputs in each block is freely determined by an anonymously elected block leader. A rational block leader will choose an ordering of inputs that maximizes financial gain; the emergence of automatic market makers in decentralized finance enables the block leader to front-run honest trade orders by injecting its own inputs prior to and after honest trades. Front-running is rampant in decentralized finance and reduces the utility of the system by extracting financial value from honest trades and increasing demand for block-space. Current proposals to prevent input order attacks by encrypting user inputs are not permissionless, as they rely on small static committees to perform distributed key generation and threshold decryption. Such committees require party authentication, knowledge of the number of participating parties or do not permit player replaceability and are therefore not permissionless. Moreover, alternative solutions based on sequencing inputs in order of their arrival cannot prevent front-running in an unauthenticated peer-2-peer network where message arrival is adversarially controlled. We present FairPoS, the first consensus protocol to achieve input fairness in the permissionless setting with security against adaptive adversaries in semi-synchronous networks. In FairPoS, the adversary cannot learn the plaintext of any client input before it is included in a block in the chain’s common-prefix. Thus, input ordering attacks that depend on observing pending client inputs in the clear are no longer possible. In FairPoS, this is achieved via Delay Encryption (DeFeo et al., EUROCRYPT 2021), a recent cryptographic primitive related to time-lock puzzles, allowing all client inputs in a given round to be encrypted under a key that can only be extracted after enough time has elapsed. In contrast to alternative approaches, the key extraction task in delay encryption can, in principle, be performed by any

Nakamoto’s consensus protocol works in a permissionless model and tolerates Byzantine failures, but only offers probabilistic agreement. Recently, the Sandglass protocol has shown such weaker guarantees are not a necessary consequence of a permissionless model; yet, Sandglass only tolerates benign failures, and operates in an unconventional partially synchronous model. We present Gorilla Sandglass, the first Byzantine tolerant consensus protocol to guarantee, in the same synchronous model adopted by Nakamoto, deterministic agreement and termination with probability 1 in a permissionless setting. We prove the correctness of Gorilla by mapping executions that would violate agreement or termination in Gorilla to executions in Sandglass, where we know such violations are impossible. Establishing termination proves particularly interesting, as the mapping requires reasoning about infinite executions and their probabilities.

This report introduces a new family of multimodal models, Gemini, that exhibit remarkable capabilities across image, audio, video, and text understanding. The Gemini family consists of Ultra, Pro, and Nano sizes, suitable for applications ranging from complex reasoning tasks to on-device memory-constrained use-cases. Evaluation on a broad range of benchmarks shows that our most-capable Gemini Ultra model advances the state of the art in 30 of 32 of these benchmarks - notably being the first model to achieve human-expert performance on the well-studied exam benchmark MMLU, and improving the state of the art in every one of the 20 multimodal benchmarks we examined. We believe that the new capabilities of the Gemini family in cross-modal reasoning and language understanding will enable a wide variety of use cases. We discuss our approach toward post-training and deploying Gemini models responsibly to users through services including Gemini, Gemini Advanced, Google AI Studio, and Cloud Vertex AI.