Search

Recent innovations in language model training have demonstrated that it is possible to create highly performant models that are small enough to run on a smartphone. As these models are deployed in an increasing number of domains, it is critical to ensure that they are aligned with human preferences and safety considerations. In this report, we present our methodology for safety aligning the Phi-3 series of language models. We utilized a "break-fix" cycle, performing multiple rounds of dataset curation, safety post-training, benchmarking, red teaming, and vulnerability identification to cover a variety of harm areas in both single and multi-turn scenarios. Our results indicate that this approach iteratively improved the performance of the Phi-3 models across a wide range of responsible AI benchmarks. Finally, we include additional red teaming strategies and evaluations that were used to test the safety behavior of Phi-3.5-mini and Phi-3.5-MoE, which were optimized for multilingual capabilities.

One compelling vision of the future of materials discovery and design involves the use of machine learning (ML) models to predict materials properties and then rapidly find materials tailored for specific applications. However, realizing this vision requires both providing detailed uncertainty quantification (model prediction errors and domain of applicability) and making models readily usable. At present, it is common practice in the community to assess ML model performance only in terms of prediction accuracy (e.g., mean absolute error), while neglecting detailed uncertainty quantification and robust model accessibility and usability. Here, we demonstrate a practical method for realizing both uncertainty and accessibility features with a large set of models. We develop random forest ML models for 33 materials properties spanning an array of data sources (computational and experimental) and property types (electrical, mechanical, thermodynamic, etc.). All models have calibrated ensemble error bars to quantify prediction uncertainty and domain of applicability guidance enabled by kernel-density-estimate-based feature distance measures. All data and models are publicly hosted on the Garden-AI infrastructure, which provides an easy-to-use, persistent interface for model dissemination that permits models to be invoked with only a few lines of Python code. We demonstrate the power of this approach by using our models to conduct a fully ML-based materials discovery exercise to search for new stable, highly active perovskite oxide catalyst materials.

Despite considerable progress in developing artificial intelligence (AI) algorithms for prostate cancer detection from whole slide images, the clinical applicability of these models remains limited due to variability in pathological annotations and existing dataset limitations. This article proposes a novel approach to overcome these challenges by leveraging a Bayesian framework to seamlessly integrate new data, and present results as a panel of annotations. The framework is demonstrated by integrating a Bayesian prior with one trained AI model to generate a distribution of Gleason patterns for each pixel of an image. It is shown that using this distribution of Gleason patterns rather than a ground-truth label can improve model applicability, mitigate errors, and highlight areas of interest for pathologists. Additionally, we present a high-quality, hand-curated dataset of prostate histopathological images annotated at the gland level by trained pre-medical students and verified by an expert pathologist. We highlight the potential of this adaptive and uncertainty-aware framework for developing clinically deployable AI tools that can support pathologists in accurate prostate cancer grading, improve diagnostic accuracy, and create positive patient outcomes., Comment: 5 pages, 3 figures, one table. Accepted for presentation at the 22nd International Conference on Artificial Intelligence in Medicine (AIME 2024)

We carried out very deep VLT/SPHERE imaging polarimetry of the nearby system Eps Eri based on 38.5 hours of integration time with a 600 - 900 nm broadband filter to search for polarized scattered light from a planet or from circumstellar dust using AO, coronagraphy, high precision differential polarimetry, and angular differential imaging. We have improved several data reduction and post-processing techniques and also developed new ones to further increase the sensitivity of SPHERE/ZIMPOL. The data provide unprecedented contrast limits, but no significant detection of a point source or an extended signal from circumstellar dust. For each observing epoch, we obtained a point source contrast for the polarized intensity between $2\cdot 10^{-8}$ and $4\cdot 10^{-8}$ at the expected separation of the planet Eps Eri b of 1'' near quadrature phase. The polarimetric contrast limits are about six to 50 times better than the intensity limits because polarimetric imaging is much more efficient in speckle suppression. Combining the entire 14-month data set to the search for a planet moving on a Keplerian orbit with the K-Stacker software further improves the contrast limits by a factor of about two, to about $8 \cdot 10^{-9}$ at 1''. This would allow the detection of a planet with a radius of about 2.5 Jupiter radii. The surface brightness contrast limits achieved for the polarized intensity from an extended scattering region are about 15 mag arcsec$^{-2}$ at 1'', or up to 3 mag arcsec$^{-2}$ deeper than previous limits. For Eps Eri, these limits exclude the presence of a narrow dust ring and they constrain the dust properties. This study shows that the polarimetric contrast limits for reflecting planets with SPHERE/ZIMPOL can be improved to a level $<10^{-8}$ simply by collecting more data over many nights and using the K-Stacker software., Comment: 26 pages, 20 figures, 5 tables. Accepted for publication in A&A

We used VLT/SPHERE to observe 20 systems in the Cha I cloud in polarized scattered light in the near-infrared. We combined the scattered light observations with existing literature data on stellar properties and with archival ALMA continuum data to study trends with system age and dust mass. We also connected resolved near-infrared observations with the spectral energy distributions of the systems. In 13 of the 20 systems included in this study we detected resolved scattered light signals from circumstellar dust. For the CR Cha, CT Cha, CV Cha, SY Cha, SZ Cha, and VZ Cha systems we present the first detailed descriptions of the disks in scattered light. The observations found typically smooth or faint disks, often with little substructure, with the notable exceptions of SZ Cha, which shows an extended multiple-ringed disk, and WW Cha, which shows interaction with the cloud environment. New high S/N K- band observations of the HD 97048 system in our survey reveal a significant brightness asymmetry that may point to disk misalignment and subsequent shadowing of outer disk regions, possibly related to the suggested planet candidate in the disk. We resolve for the first time the stellar binary in the CS Cha system. Multiple wavelength observations of the disk around CS Cha have revealed that the system contains small, compact dust grains that may be strongly settled, consistent with numerical studies of circumbinary disks. We find in our sample that there is a strong anti-correlation between the presence of a (close) stellar companion and the detection of circumstellar material with five of our seven nondetections located in binary systems., Comment: Accepted by A&A

Foundation models are powerful technologies: how they are released publicly directly shapes their societal impact. In this position paper, we focus on open foundation models, defined here as those with broadly available model weights (e.g. Llama 2, Stable Diffusion XL). We identify five distinctive properties (e.g. greater customizability, poor monitoring) of open foundation models that lead to both their benefits and risks. Open foundation models present significant benefits, with some caveats, that span innovation, competition, the distribution of decision-making power, and transparency. To understand their risks of misuse, we design a risk assessment framework for analyzing their marginal risk. Across several misuse vectors (e.g. cyberattacks, bioweapons), we find that current research is insufficient to effectively characterize the marginal risk of open foundation models relative to pre-existing technologies. The framework helps explain why the marginal risk is low in some cases, clarifies disagreements about misuse risks by revealing that past work has focused on different subsets of the framework with different assumptions, and articulates a way forward for more constructive debate. Overall, our work helps support a more grounded assessment of the societal impact of open foundation models by outlining what research is needed to empirically validate their theoretical benefits and risks.

Deep learning methods are transforming research, enabling new techniques, and ultimately leading to new discoveries. As the demand for more capable AI models continues to grow, we are now entering an era of Trillion Parameter Models (TPM), or models with more than a trillion parameters -- such as Huawei's PanGu-$\Sigma$. We describe a vision for the ecosystem of TPM users and providers that caters to the specific needs of the scientific community. We then outline the significant technical challenges and open problems in system design for serving TPMs to enable scientific research and discovery. Specifically, we describe the requirements of a comprehensive software stack and interfaces to support the diverse and flexible requirements of researchers., Comment: 10 pages, 3 figures, accepted for publication in the proceedings of the 10th IEEE/ACM International Conference on Big Data Computing, Applications and Technologies (BDCAT2023)

Direct imaging of Earth-like exoplanets is one of the most prominent scientific drivers of the next generation of ground-based telescopes. Typically, Earth-like exoplanets are located at small angular separations from their host stars, making their detection difficult. Consequently, the adaptive optics (AO) system's control algorithm must be carefully designed to distinguish the exoplanet from the residual light produced by the host star. A new promising avenue of research to improve AO control builds on data-driven control methods such as Reinforcement Learning (RL). RL is an active branch of the machine learning research field, where control of a system is learned through interaction with the environment. Thus, RL can be seen as an automated approach to AO control, where its usage is entirely a turnkey operation. In particular, model-based reinforcement learning (MBRL) has been shown to cope with both temporal and misregistration errors. Similarly, it has been demonstrated to adapt to non-linear wavefront sensing while being efficient in training and execution. In this work, we implement and adapt an RL method called Policy Optimization for AO (PO4AO) to the GHOST test bench at ESO headquarters, where we demonstrate a strong performance of the method in a laboratory environment. Our implementation allows the training to be performed parallel to inference, which is crucial for on-sky operation. In particular, we study the predictive and self-calibrating aspects of the method. The new implementation on GHOST running PyTorch introduces only around 700 microseconds in addition to hardware, pipeline, and Python interface latency. We open-source well-documented code for the implementation and specify the requirements for the RTC pipeline. We also discuss the important hyperparameters of the method, the source of the latency, and the possible paths for a lower latency implementation., Comment: Accepted for publication in JATIS

When a damaging earthquake occurs, immediate information about casualties is critical for time-sensitive decision-making by emergency response and aid agencies in the first hours and days. Systems such as Prompt Assessment of Global Earthquakes for Response (PAGER) by the U.S. Geological Survey (USGS) were developed to provide a forecast within about 30 minutes of any significant earthquake globally. Traditional systems for estimating human loss in disasters often depend on manually collected early casualty reports from global media, a process that's labor-intensive and slow with notable time delays. Recently, some systems have employed keyword matching and topic modeling to extract relevant information from social media. However, these methods struggle with the complex semantics in multilingual texts and the challenge of interpreting ever-changing, often conflicting reports of death and injury numbers from various unverified sources on social media platforms. In this work, we introduce an end-to-end framework to significantly improve the timeliness and accuracy of global earthquake-induced human loss forecasting using multi-lingual, crowdsourced social media. Our framework integrates (1) a hierarchical casualty extraction model built upon large language models, prompt design, and few-shot learning to retrieve quantitative human loss claims from social media, (2) a physical constraint-aware, dynamic-truth discovery model that discovers the truthful human loss from massive noisy and potentially conflicting human loss claims, and (3) a Bayesian updating loss projection model that dynamically updates the final loss estimation using discovered truths. We test the framework in real-time on a series of global earthquake events in 2021 and 2022 and show that our framework streamlines casualty data retrieval, achieving speed and accuracy comparable to manual methods by USGS., Comment: 10 pages, 8 figures

Polarisation is a powerful remote-sensing tool to study the nature of particles scattering the starlight. It is widely used to characterise interplanetary dust particles in the Solar System and increasingly employed to investigate extrasolar dust in debris discs' systems. We aim to measure the scattering properties of the dust from the debris ring around HD 181327 at near-infrared wavelengths. We obtained high-contrast polarimetric images of HD 181327 in the H band with the SPHERE / IRDIS instrument on the Very Large Telescope (ESO). We complemented them with archival data from HST / NICMOS in the F110W filter reprocessed in the context of the Archival Legacy Investigations of Circumstellar Environments (ALICE) project. We developed a combined forward-modelling framework to simultaneously retrieve the scattering phase function in polarisation and intensity. We detected the debris disc around HD 181327 in polarised light and total intensity. We measured the scattering phase function and the degree of linear polarisation of the dust at 1.6 micron in the birth ring. The maximum polarisation is 23.6% +/- 2.6% and occurs between a scattering angle of 70 deg and 82 deg. We show that compact spherical particles made of a highly refractive and relatively absorbing material in a differential power-law size distribution of exponent $-3.5$ can simultaneously reproduce the polarimetric and total intensity scattering properties of the dust. This type of material cannot be obtained with a mixture of silicates, amorphous carbon, water ice, and porosity, and requires a more refracting component such as iron-bearing minerals. We reveal a striking analogy between the near-infrared polarisation of comets and that of HD 181327. The methodology developed here combining VLT/SPHERE and HST/NICMOS may be applicable in the future to combine the polarimetric capabilities of SPHERE with the sensitivity of JWST., Comment: 21 pages, 22 figures, accepted in Astronomy & Astrophysics

The confinement of ultracold neutrons (UCNs) in a three dimensional magnetic field gradient trap allows for a measurement of the free neutron lifetime with superior control over spurious loss channels and can provide a large kinetic energy acceptance to enhance statistical sensitivity. In this paper, we present the first successful implementation of a pulsed spin-flip based loading scheme for a three-dimensional magnetic UCN trap. The measurements with the $\tau$SPECT experiment were performed at the pulsed UCN source of the research reactor TRIGA Mainz. We report on detailed investigations of major systematic effects influencing the neutron storage time, statistically limited by the size of the recorded data set. The extracted neutron storage time constant of $\tau = 859(16)\mathrm{s}$ is compatible with, but not to be interpreted as, a measurement of the free neutron lifetime., Comment: 15 pages, 19 figures, version accepted for publication in Journal of Physics G

Context. The study of protoplanetary disks is fundamental to understand their evolution and interaction with the surrounding environment, and to constrain planet formation mechanisms. Aims. We aim at characterising the young binary system HD 34700 A, which shows a wealth of structures. Methods. Taking advantage of the high-contrast imaging instruments SPHERE at the VLT, LMIRCam at the LBT, and of ALMA observations, we analyse this system at multiple wavelengths. We study the rings and spiral arms morphology and the scattering properties of the dust. We discuss the possible causes of all the observed features. Results. We detect for the first time, in the H${\alpha}$ band, a ring extending from $\sim$65 au to ${\sim}$120 au, inside the ring already known from recent studies. These two have different physical and geometrical properties. Based on the scattering properties, the outer ring may consist of grains of typical size $a_{out} > 4 {\mu}m$, while the inner ring of smaller grains ($a_{in} <= 0.4 {\mu m}$). Two extended logarithmic spiral arms stem from opposite sides of the disk. The outer ring appears as a spiral arm itself, with a variable radial distance from the centre and extended substructures. ALMA data confirm the presence of a millimetric dust substructure centred just outside the outer ring, and detect misaligned gas rotation patterns for HD 34700 A and B. Conclusions. The complexity of HD 34700 A, revealed by the variety of observed features, suggests the existence of one or more disk-shaping physical mechanisms. Possible scenarios, compatible with our findings, involve the presence inside the disk of a yet undetected planet of several Jupiter masses and the system interaction with the surroundings by means of gas cloudlet capture or flybys. Further observations with JWST/MIRI or ALMA (gas kinematics) could shed more light on these., Comment: Accepted for publication on A&A. 14 + 5 pages, 9 + 7 figures (text + appendix)

The distribution-free chain ladder of Mack justified the use of the chain ladder predictor and enabled Mack to derive an estimator of conditional mean squared error of prediction for the chain ladder predictor. Classical insurance loss models, i.e. of compound Poisson type, are not consistent with Mack's distribution-free chain ladder. However, for a sequence of compound Poisson loss models indexed by exposure (e.g. number of contracts), we show that the chain ladder predictor and Mack's estimator of conditional mean squared error of prediction can be derived by considering large exposure asymptotics. Hence, quantifying chain ladder prediction uncertainty can be done with Mack's estimator without relying on the validity of the model assumptions of the distribution-free chain ladder.

Warps have often been used to explain disc properties, but well characterised examples are important due to their role in disc evolution. Scattered light images of discs with central gaps have revealed sharp warps, such that the outer rings are shadowed by tilted inner discs. The near-IR intensity drops along the ring around TTauri star DoAr44 have been interpreted in terms of a central warp. We report new ALMA observations of DoAr44 in the continuum at 230 GHz and 350 GHz (at ~10 au), along with a new epoch of SPHERE/IRDIS differential polarised imaging taken during excellent weather conditions. The ALMA observations resolve the ring and confirm the decrements proposed from deconvolution of coarse 336 GHz data. The scattered light image constrains the dips, which correspond to a misaligned inner disc with a relative inclination $\xi$ = 21.4 $^{+6.7}_{-8.3}$ deg. The SPHERE intensity profile shows a morphological change compared to a previous epoch that may be interpreted as a variable orientation of the inner disc, from $\xi$ ~30 deg to $\xi$ ~20 deg. The intensity dips probably correspond to temperature decrements, as their mm-spectral index, $\alpha^{230 GHz}_{350 GHz}$ ~2.0 $\pm$ 0.1, is indicative of optically thick emission. The azimuth of the two temperature decrements are leading clockwise relative to the IR-dips, by $\eta$ = 14.95 deg and $\eta$ = 7.92 deg. For a retrograde disc, such shifts are expected from a thermal lag and imply gas surface densities of $\Sigma_g$ = 117 $\pm$ 10 g/cm$^2$ and $\Sigma_g$ = 48 $\pm$ 10 g/cm$^2$. A lopsided disc, with contrast ratio $f_r$=2.4 $\pm$ 0.5, is also consistent with the large continuum crescent., Comment: accepted in MNRAS

We propose to explore a cascade extreme Adaptive optics (ExAO) approach with a second stage based on a Zernike wavefront sensor (ZWFS) for exoplanet imaging and spectroscopy. Most exoplanet imagers currently use a single-stage ExAO to correct for the effects of atmospheric turbulence and produce high-Strehl images of observed stars in the near-infrared. While such systems enable the observation of warm gaseous companions around nearby stars, adding a second-stage AO enables to push the wavefront correction further and possibly observe colder or smaller planets. This approach is currently investigated in different exoplanet imagers (VLT/SPHERE, Mag-AOX, Subaru/SCExAO) by considering a Pyramid wavefront sensor (PWFS) in the second arm to measure the residual atmospheric turbulence left from the first stage. Since these aberrations are expected to be very small (a few tens of nm in the near-infrared domain), we propose to investigate an alternative approach based on the ZWFS. This sensor is a promising concept with a small capture range to estimate residual wavefront errors thanks to its large sensitivity, simple phase reconstruction and easiness of implementation. In this contribution, we perform preliminary tests on the GHOST testbed at ESO to validate this approach experimentally. Additional experiments with petalling effects are also showed, giving promising wavefront correction results. Finally, we briefly discuss a first comparison between PWFS-based and ZWFS-based second-stage AO to draw preliminary conclusions on the interests of both schemes for exoplanet imaging and spectroscopy with the upgrade of the current exoplanet imagers and the envisioned ExAO instruments for ELTs., Comment: 17 pages, 10 figures, pre-print of the proceeding of the AO4ELT7 conference held in June 2023 in Avignon, France

A variety of experiments investigating properties of neutrons can be performed at pulsed reactor neutron sources like the research reactor TRIGA Mainz. A typical problem faced by these experiments is the non-availability of a reliable facility-provided trigger signal in coincidence with the neutron production. Here we present the design and implementation of a neutron pulse detector that provides a coincident trigger signal for experimental timing with a relative precision of 4.5 ms., Comment: 18 pages, 10 figures

Since 2019, the direct imaging B-star Exoplanet Abundance Study (BEAST) at SPHERE@VLT has been scanning the surroundings of young B-type stars in order to ascertain the ultimate frontiers of giant planet formation. Recently, the $17^{+3}_{-4}$ Myr HIP 81208 was found to host a close-in (~50 au) brown dwarf and a wider (~230 au) late M star around the central 2.6Msun primary. Alongside the continuation of the survey, we are undertaking a complete reanalysis of archival data aimed at improving detection performances so as to uncover additional low-mass companions. We present here a new reduction of the observations of HIP 81208 using PACO ASDI, a recent and powerful algorithm dedicated to processing high-contrast imaging datasets, as well as more classical algorithms and a dedicated PSF-subtraction approach. The combination of different techniques allowed for a reliable extraction of astrometric and photometric parameters. A previously undetected source was recovered at a short separation from the C component of the system. Proper motion analysis provided robust evidence for the gravitational bond of the object to HIP 81208 C. Orbiting C at a distance of ~20 au, this 15Mjup brown dwarf becomes the fourth object of the hierarchical HIP 81208 system. Among the several BEAST stars which are being found to host substellar companions, HIP 81208 stands out as a particularly striking system. As the first stellar binary system with substellar companions around each component ever found by direct imaging, it yields exquisite opportunities for thorough formation and dynamical follow-up studies., Comment: 12 pages, 9 figures, 5 tables. Accepted for publication as a Letter in Astronomy and Astrophysics, section 1. Letters to the Editor