Search

Benchmarks are important tools for tracking the rapid advancements in large language model (LLM) capabilities. However, benchmarks are not keeping pace in difficulty: LLMs now achieve over 90\% accuracy on popular benchmarks like MMLU, limiting informed measurement of state-of-the-art LLM capabilities. In response, we introduce Humanity's Last Exam (HLE), a multi-modal benchmark at the frontier of human knowledge, designed to be the final closed-ended academic benchmark of its kind with broad subject coverage. HLE consists of 3,000 questions across dozens of subjects, including mathematics, humanities, and the natural sciences. HLE is developed globally by subject-matter experts and consists of multiple-choice and short-answer questions suitable for automated grading. Each question has a known solution that is unambiguous and easily verifiable, but cannot be quickly answered via internet retrieval. State-of-the-art LLMs demonstrate low accuracy and calibration on HLE, highlighting a significant gap between current LLM capabilities and the expert human frontier on closed-ended academic questions. To inform research and policymaking upon a clear understanding of model capabilities, we publicly release HLE at https://lastexam.ai., Comment: 26 pages, 6 figures

This paper presents an analysis of the predicted optical-to-X-ray spectral index ($\alpha_{\rm ox}$) within the context of ultra-luminous X-ray sources (ULXs) associated with stellar mass black holes and neutron stars. We use the population synthesis code COSMIC to simulate the evolution of binary systems and investigate the relationship between UV and X-ray emission during the ULX phase, namely the $\alpha_{\rm ox}$ relation. The study investigates the impact of metallicity on $\alpha_{\rm ox}$ values. Notably, it predicts a significant anti-correlation between $\alpha_{\rm ox}$ and UV luminosity ($L_{\rm UV}$), consistent with observations, with the slope of this relationship varying with metallicity for BH-ULXs. The NS-ULX population shows a relatively consistent slope around $-0.33$ across metallicities, with minor variations. The number of ULXs decreases with increasing metallicity, consistent with observational data, and the X-ray luminosity function shows a slight variation in its slope with metallicity, exhibiting a relative excess of high-luminosity ULXs at lower metallicities. Inclusion of beaming effect in the analysis shows a significant impact on the XLF and $\alpha_{\rm ox}$, particularly at high accretion rates, where the emission is focused into narrower cones. Furthermore, the study finds that UV emission in ULXs is predominantly disk-dominated, which is the likely origin of the $\alpha_{\rm ox}$ relation, with the percentage of disk-dominated ULXs increasing as metallicity rises., Comment: 16 pages, 18 figures, accepted for publication in Astronomy & Astrophysics

Significant improvements in our understanding of nuclear $\gamma$-ray line production and instrument performance allow us to better characterize the continuum emission from electrons at energies $\gtrsim$ 300 keV during solar flares. We represent this emission by the sum of a power-law extension of hard X-rays (PL) and a power law times an exponential function (PLexp). We fit the $\gamma$-ray spectra in 25 large flares observed by SMM, RHESSI, and Fermi with this summed continuum along with calculated spectra of all known nuclear components. The PL, PLexp, and nuclear components are separated spectroscopically. A distinct origin of the PLexp is suggested by significant differences between its time histories and those of the PL and nuclear components. RHESSI imaging/spectroscopy of the 2005 January 20 flare, reveals that the PL and nuclear components come from the footpoints while the PLexp component comes from the corona. While the index and flux of the anisotropic PL component are strongly dependent on the flares' heliocentric angle, the PLexp parameters show no such dependency and are consistent with a component that is isotropic. The PLexp spectrum is flat at low energies and rolls over at a few MeV. Such a shape can be produced by inverse Compton scattering of soft X-rays by 10--20 MeV electrons and by thin-target bremsstrahlung from electrons with a spectrum that peaks between 3 -- 5 MeV, or by a combination of the two processes. These electrons can produce radiation detectable at other wavelengths., Comment: 33 pages, 16 figures. Accepted for publication in ApJ

SIRI-2 is a collection of Strontium Iodide gamma-ray detectors sensitive at approximately 400 keV to 10 MeV, launched on the Department of Defense's STPSat-6 to geosynchronous orbit. SIRI-2 detected the gamma-ray burst (GRB) 221009A and, unlike most GRB detectors, was not saturated and did not require any pulse pile-up corrections. The energetics of this burst as measured by SIRI-2 are consistent with those found by other instruments, and the Band function fits to the spectra are consistent with that from the unsaturated Insight and GECAM instruments, and similar to corrected spectra from the Fermi Gamma-ray Burst Monitor, but softer than those found by Konus-Wind when that instrument was saturated. The total fluence measured with SIRI-2 was measured to be 0.140 +/- 0.002 erg cm-2, lower than other instruments, likely due to the increasing background of SIRI-2 forcing the calculation to use a smaller time interval. An extrapolation of the distributions of fluences from GRBs to the fluence of 221009A measured with SIRI-2 indicates bursts brighter than this one should occur about once every 4,000 years., Comment: 9 pages, 8 figures. Accepted by ApJ

Machine unlearning in neural information retrieval (IR) systems requires removing specific data whilst maintaining model performance. Applying existing machine unlearning methods to IR may compromise retrieval effectiveness or inadvertently expose unlearning actions due to the removal of particular items from the retrieved results presented to users. We formalise corrective unranking, which extends machine unlearning in (neural) IR context by integrating substitute documents to preserve ranking integrity, and propose a novel teacher-student framework, Corrective unRanking Distillation (CuRD), for this task. CuRD (1) facilitates forgetting by adjusting the (trained) neural IR model such that its output relevance scores of to-be-forgotten samples mimic those of low-ranking, non-retrievable samples; (2) enables correction by fine-tuning the relevance scores for the substitute samples to match those of corresponding to-be-forgotten samples closely; (3) seeks to preserve performance on samples that are not targeted for forgetting. We evaluate CuRD on four neural IR models (BERTcat, BERTdot, ColBERT, PARADE) using MS MARCO and TREC CAR datasets. Experiments with forget set sizes from 1 % and 20 % of the training dataset demonstrate that CuRD outperforms seven state-of-the-art baselines in terms of forgetting and correction while maintaining model retention and generalisation capabilities., Comment: submitted to Information Sciences

Resonant anomaly detection methods have great potential for enhancing the sensitivity of traditional bump hunt searches. A key component of these methods is a high quality background template used to produce an anomaly score. Using the LHC Olympics R&D dataset, we demonstrate that this background template can also be repurposed to directly estimate the background expectation in a simple cut and count setup. In contrast to a traditional bump hunt, no fit to the invariant mass distribution is needed, thereby avoiding the potential problem of background sculpting. Furthermore, direct background estimation allows working with large background rejection rates, where resonant anomaly detection methods typically show their greatest improvement in significance., Comment: 26 pages, 9 figures, 2 tables

Gamma rays selectively interact with nuclei, induce and mediate nuclear reactions and elementary particle interactions, and exceed x-rays in penetrating power and thus are indispensable for analysis and modification of dense objects. Yet, the available gamma sources lack sufficient power and brightness. The predicted and highly desirable laser-driven gamma flash, from here on termed "Gamma Flash", based on inverse Compton scattering from solid targets at extreme irradiances (>$10^{23}W/cm^2$), would be the highest-power and the brightest terrestrial gamma source with a 30-40% laser-to-gamma energy conversion. However, Gamma Flash remains inaccessible experimentally due to the Bremsstrahlung background. Here we experimentally demonstrate a new interaction regime at the highest effective irradiance where Gamma Flash scaled quickly with the laser power and produced several times the number of Bremsstrahlung photons. Simulations revealed an attosecond, Terawatt Gamma Flash with a nanometre source size achieving a record brightness exceeding $~10^{23}photons/mm^2mrad^2s$ per 0.1% bandwidth at tens of MeV photon energies, surpassing astrophysical Gamma Ray Bursts. These findings could revolutionize inertial fusion energy by enabling unprecedented sub-micrometre/femtosecond resolution radiography of fuel mixing instabilities in extremely-compressed targets. The new gamma source could facilitate significant advances in time-resolved nuclear physics, homeland security, nuclear waste management and non-proliferation, while opening possibilities for spatially-coherent gamma rays., Comment: 39 pages, 12 figures

Product lifecycle tracing is increasingly in the focus of regulators and producers, as shown with the initiative of the Digital Product Pass. Likewise, new methods of counterfeit detection are developed that are, e.g., based on Physical Unclonable Functions (PUFs). In order to ensure trust and integrity of product lifecycle data, multiple existing supply chain tracing systems are built on blockchain technology. However, only few solutions employ secure identifiers such as PUFs. Furthermore, existing systems that publish the data of individual products, in part fully transparently, have a detrimental impact on scalability and the privacy of users. This work proposes SPOQchain, a novel blockchain-based platform that provides comprehensive lifecycle traceability and originality verification while ensuring high efficiency and user privacy. The improved efficiency is achieved by a sophisticated batching mechanism that removes lifecycle redundancies. In addition to the successful evaluation of SPOQchain's scalability, this work provides a comprehensive analysis of privacy and security aspects, demonstrating the need and qualification of SPOQchain for the future of supply chain tracing.

Facial recognition systems have become an integral part of the modern world. These methods accomplish the task of human identification in an automatic, fast, and non-interfering way. Past research has uncovered high vulnerability to simple imitation attacks that could lead to erroneous identification and subsequent authentication of attackers. Similar to face recognition, imitation attacks can also be detected with Machine Learning. Attack detection systems use a variety of facial features and advanced machine learning models for uncovering the presence of attacks. In this work, we assess existing work on liveness detection and propose a novel approach that promises high classification accuracy by combining previously unused features with time-aware deep learning strategies.

We tackle the problem of machine unlearning within neural information retrieval, termed Neural Machine UnRanking (NuMuR) for short. Many of the mainstream task- or model-agnostic approaches for machine unlearning were designed for classification tasks. First, we demonstrate that these methods perform poorly on NuMuR tasks due to the unique challenges posed by neural information retrieval. Then, we develop a methodology for NuMuR named Contrastive and Consistent Loss (CoCoL), which effectively balances the objectives of data forgetting and model performance retention. Experimental results demonstrate that CoCoL facilitates more effective and controllable data removal than existing techniques.

Gamma-ray bursts (GRBs) brighter than the GRB 221009A, the brightest yet observed, have previously been estimated to occur at a rate of 1 per 10,000 years, based on the extrapolation of the distribution of fluences of the Long GRB population. We show that bursts this bright could instead have a rate as high as approximately one per 200 years if they are from a separate population of narrow-jet GRBs. This population must have a maximum redshift of about $z\approx 0.38$ in order to avoid over-producing the observed rate of fainter GRBs. We show that it will take $> 100$ years to confirm this new population based on observing another GRB from it with a $\gamma$-ray detector; observing an orphan optical afterglow from this population with Vera Rubin Observatory or an orphan radio afterglow with the Square Kilometer Array will also take similarly long times to observe, and it is unclear if they could be distinguished from the standard GRB population. We show that the nearby narrow-jet population has more favorable energetics for producing ultra-high energy cosmic rays than standard GRBs. The rate of bursts in the Milky Way bright enough to cause mass extinctions of life on Earth from the narrow jet population is estimated to be approximately 1 per 500 Myr. This GRB population could make life in the Milky Way less likely, with implications for future searches for life on exoplanets., Comment: 15 pages, 6 Figures. Submitted to AAS Journals

The recent detection of gamma-ray burst GRB~221009A has attracted attention due to its record brightness and first-ever detection of $\gtrsim 10$ TeV $\gamma$ rays from a GRB. Despite being the second-nearest GRB ever detected, at a redshift of $z=0.151$, the distance is large enough for severe attenuation of $\gamma$-ray flux at these energies due to $\gamma\gamma\to e^\pm$ pair production with the extragalactic background light (EBL). Here, we investigate whether the presence of cosmic voids along the line of sight can significantly impact the detectability of very-high energy (VHE, $>$ 100 GeV) gamma rays from distant sources. Notably, we find that the gamma-gamma opacity for VHE gamma rays can be reduced by approximately 10\% and up to 30\% at around 13 TeV, the highest-energy photon detected from GRB~221009A, for intervening cosmic voids along the line-of-sight with a combined radius of 110 Mpc, typically found from voids catalogs, and 250 Mpc, respectively. This reduction is substantially higher for TeV photons compared to GeV photons, attributable to the broader target photon spectrum that TeV photons interact with. This finding implies that VHE photons are more susceptible to variations in the EBL spectrum, especially in regions dominated by cosmic voids. Our study sheds light on the detection of $\gtrsim 10$ TeV gamma rays from GRB 221009A in particular, and on the detection of extragalactic VHE sources in general., Comment: 11 pages, 10 figures, Accepted for publication in MNRAS

I present a model for light travel time effects for emission from a plasma blob in a blazar jet. This calculation could be incorporated into more complex models with particle acceleration and radiation mechanisms, but as presented here it is a agnostic as to these mechanisms. This model includes light travel time effects for an expanding or contracting blob. As an example, this model is applied to a flare observed by VERITAS and MAGIC from Mrk 421 in 2013; and to a flare observed by the Fermi Large Area Telescope from 3C 454.3 in 2010., Comment: 12 pages, 5 figures, 2 tables. Accepted for special issue of Frontiers in Astronomy and Space Science titled "The Dynamic Universe: Realizing the Potential of Time Domain and Multimessenger Astrophysics"

Peer review is a laborious, yet essential, part of academic publishing with crucial impact on the scientific endeavor. The current lack of incentives and transparency harms the credibility of this process. Researchers are neither rewarded for superior nor penalized for bad reviews. Additionally, confidential reports cause a loss of insights and make the review process vulnerable to scientific misconduct. We propose a community-owned and -governed system that 1) remunerates reviewers for their efforts, 2) publishes the (anonymized) reports for scrutiny by the community, 3) tracks reputation of reviewers and 4) provides digital certificates. Automated by transparent smart-contract blockchain technology, the system aims to increase quality and speed of peer review while lowering the chance and impact of erroneous judgements., Comment: 14 pages, 1 figure

We compute in detail the absorption optical depth for astrophysical $\gamma$-ray photons interacting with solar photons to produce electron positron pairs. This effect is greatest for $\gamma$-ray sources at small angular distances from the Sun, reaching optical depths as high as $\tau_{\gamma\gamma}\sim 10^{-2}$. We also calculate this effect including modifications to the absorption cross section threshold from subluminal Lorentz invariance violation (LIV). We show for the first time that subluminal LIV can lead to increases or decreases in $\tau_{\gamma\gamma}$ compared to the non-LIV case. We show that, at least in principle, LIV can be probed with this effect with observations of $\gamma$-ray sources near the Sun at $\gtrsim20$ TeV by HAWC or LHAASO, although a measurement will be extremely difficult due to the small size of the effect., Comment: 12 pages, 8 figures, 3 tables. Accepted by ApJ

High harmonic generation (HHG) has become a multipurpose source of coherent XUV radiation used in various applications. One of the notable aspects of HHG is its wide spectrum consisting of many harmonic orders. This might represent a bottleneck in HHG utility for applications requiring a single wavelength. We propose a method to generate radiation consisting of a single high-order harmonic frequency employing Bessel-Gauss driving beam and a periodically modulated gaseous medium. We validate it by analytical calculations and numerical simulations. Our method provides a way to generate monochromatic harmonic radiation directly from the source without the need for additional monochromatizing optics. Thus, it represents a substantial enhancement of the flux and simplification of the setup for numerous applications requiring monochromatic short-wavelength radiation.

State-of-the-art methods for Bayesian inference in state-space models are (a) conditional sequential Monte Carlo (CSMC) algorithms; (b) sophisticated 'classical' MCMC algorithms like MALA, or mGRAD from Titsias and Papaspiliopoulos (2018, arXiv:1610.09641v3 [stat.ML]). The former propose $N$ particles at each time step to exploit the model's 'decorrelation-over-time' property and thus scale favourably with the time horizon, $T$ , but break down if the dimension of the latent states, $D$, is large. The latter leverage gradient-/prior-informed local proposals to scale favourably with $D$ but exhibit sub-optimal scalability with $T$ due to a lack of model-structure exploitation. We introduce methods which combine the strengths of both approaches. The first, Particle-MALA, spreads $N$ particles locally around the current state using gradient information, thus extending MALA to $T > 1$ time steps and $N > 1$ proposals. The second, Particle-mGRAD, additionally incorporates (conditionally) Gaussian prior dynamics into the proposal, thus extending the mGRAD algorithm to $T > 1$ time steps and $N > 1$ proposals. We prove that Particle-mGRAD interpolates between CSMC and Particle-MALA, resolving the 'tuning problem' of choosing between CSMC (superior for highly informative prior dynamics) and Particle-MALA (superior for weakly informative prior dynamics). We similarly extend other 'classical' MCMC approaches like auxiliary MALA, aGRAD, and preconditioned Crank-Nicolson-Langevin (PCNL) to $T > 1$ time steps and $N > 1$ proposals. In experiments, for both highly and weakly informative prior dynamics, our methods substantially improve upon both CSMC and sophisticated 'classical' MCMC approaches., Comment: 29 pages + 31 pages appendix. 6 figures and tables (+ 7 in appendix). Code available at https://github.com/AdrienCorenflos/particle_mala/

In practice, deep neural networks are often able to easily interpolate their training data. To understand this phenomenon, many works have aimed to quantify the memorization capacity of a neural network architecture: the largest number of points such that the architecture can interpolate any placement of these points with any assignment of labels. For real-world data, however, one intuitively expects the presence of a benign structure so that interpolation already occurs at a smaller network size than suggested by memorization capacity. In this paper, we investigate interpolation by adopting an instance-specific viewpoint. We introduce a simple randomized algorithm that, given a fixed finite data set with two classes, with high probability constructs an interpolating three-layer neural network in polynomial time. The required number of parameters is linked to geometric properties of the two classes and their mutual arrangement. As a result, we obtain guarantees that are independent of the number of samples and hence move beyond worst-case memorization capacity bounds. We verify our theoretical result with numerical experiments and additionally investigate the effectiveness of the algorithm on MNIST and CIFAR-10., Comment: The current version of the manuscript has been accepted to Journal of Machine Learning Research

Weakly supervised methods have emerged as a powerful tool for model-agnostic anomaly detection at the Large Hadron Collider (LHC). While these methods have shown remarkable performance on specific signatures such as di-jet resonances, their application in a more model-agnostic manner requires dealing with a larger number of potentially noisy input features. In this paper, we show that using boosted decision trees as classifiers in weakly supervised anomaly detection gives superior performance compared to deep neural networks. Boosted decision trees are well known for their effectiveness in tabular data analysis. Our results show that they not only offer significantly faster training and evaluation times, but they are also robust to a large number of noisy input features. By using advanced gradient boosted decision trees in combination with ensembling techniques and an extended set of features, we significantly improve the performance of weakly supervised methods for anomaly detection at the LHC. This advance is a crucial step towards a more model-agnostic search for new physics., Comment: 11 pages, 9 figures