Search

Using Parker Solar Probe data from orbits 8 through 17, we examine fluctuation amplitudes throughout the critical region where the solar wind flow speed approaches and then exceeds the Alfv\'en wave speed, taking account of various exigencies of the plasma data. In contrast to WKB theory for non-interacting Alfv\'en waves streaming away from the Sun, the magnetic and kinetic fluctuation energies per unit volume are not monotonically decreasing. Instead, there is clear violation of conservation of standard WKB wave action, which is consistent with previous indications of strong in-situ fluctuation energy input in the solar wind near the Alfv\'en critical region. This points to strong violations of WKB theory due to nonlinearity (turbulence) and major energy input near the critical region, which we interpret as likely due to driving by large-scale coronal shear flows.

The IceCube neutrino observatory detects the diffuse astrophysical neutrino background with high significance, but the contribution of different classes of sources is not established. Because of their non-thermal spectrum, gamma-ray bursts (GRBs) are prime particle acceleration sites and one of the candidate classes for significant neutrino production. Exhaustive searches, based on stacking analysis of GRBs however could not establish the link between neutrinos and GRBs. Gamma-ray burst GRB 221009A had the highest time integrated gamma-ray flux of any detected GRB so far. The total fluence exceeds the sum of all Fermi Gamma-ray Burst Monitor (GBM) detected GRBs by a factor of two. Because it happened relatively nearby, it is one of the most favorable events for neutrino production from GRBs yet no neutrinos were detected. We calculate neutrino fluxes for this GRB in the TeV-PeV range using the most accurate, time-resolved spectral data covering the brightest intervals. We place limits on the physical parameters (Lorentz factor, baryon loading or emission radius) of the burst that are better by a factor of 2 compared to previous limits. The neutrino non-detection indicates a bulk Lorentz factor greater than 500 and possibly even 1000, consistent with other observations., Comment: 10 pages, 3 figures, 1 table. Submitted to AAS journals

The study of transmission and progression of human papillomavirus (HPV) is crucial for understanding the incidence of cervical cancers, and has been identified as a priority worldwide. The complexity of the disease necessitates a detailed model of HPV transmission and its progression to cancer; to infer properties of the above we require a careful process that can match to imperfect or incomplete observational data. In this paper, we describe the HPVsim simulator to satisfy the former requirement; to satisfy the latter we couple this stochastic simulator to a process of emulation and history matching using the R package hmer. With these tools, we are able to obtain a comprehensive collection of parameter combinations that could give rise to observed cancer data, and explore the implications of the variability of these parameter sets as it relates to future health interventions., Comment: 21 pages, 15 figures; submitted to Epidemics

Suppose we are given two OCaml modules implementing the same signature. How do we check that they are observationally equivalent -- that is, that they behave the same on all inputs? One established technique is to use a property-based testing (PBT) tool such as QuickCheck. Currently, however, this can require significant amounts of boilerplate code and ad-hoc test harnesses. To address this issue, we present Mica, an automated tool for testing observational equivalence of OCaml modules. Mica is implemented as a PPX compiler extension, allowing users to supply minimal annotations to a module signature. These annotations guide Mica to automatically derive specialized PBT code that checks observational equivalence. We discuss the design of Mica and demonstrate its efficacy as a testing tool on various modules taken from real-world OCaml libraries., Comment: OCaml Workshop 2024

To achieve confident non-zero polarization measurements for gamma-ray bursts (GRBs) we need sensitive polarimeters and bright GRBs. Here we report on the polarimetric analysis of the bright GRB 180720B using the \Fermi Gamma-ray Burst Monitor (GBM). We rely on the detection of photons that scattered off Earth's atmosphere and into GBM from this burst. Polarized gamma-rays will exhibit a characteristic pattern when scattering off the atmosphere that differs from an unpolarized beam. We compare the measured photon counts in the GBM detectors with extensive simulations of polarized beams to derive the most probable polarization degree (PD) and angle (PA). For the entire GRB, we find PD$=72^{+24}_{-30}\% ~(1\sigma)$ and PA$=91^{+11}_{-9}$ deg ($1\sigma$, equatorial frame). Interestingly, the PA value is broadly consistent with an early optical PA measurement by the Kanata telescope, starting shortly after the end of the prompt emission. The consistency of PAs lends support for this method. The relatively high polarization degree (albeit with large uncertainties) agrees with similar past measurements suggesting that some GRBs might be highly polarized. This will be confirmed or refuted by the upcoming dedicated GRB polarimeters., Comment: 8 pages, 5 figures, submitted to AAS journals

Deeply virtual exclusive scattering processes (DVES) serve as precise probes of nucleon quark and gluon distributions in coordinate space. These distributions are derived from generalized parton distributions (GPDs) via Fourier transform relative to proton momentum transfer. QCD factorization theorems enable DVES to be parameterized by Compton form factors (CFFs), which are convolutions of GPDs with perturbatively calculable kernels. Accurate extraction of CFFs from DVCS, benefiting from interference with the Bethe-Heitler (BH) process and a simpler final state structure, is essential for inferring GPDs. This paper focuses on extracting CFFs from DVCS data using a variational autoencoder inverse mapper (VAIM) and its constrained variant (C-VAIM). VAIM is shown to be consistent with Markov Chain Monte Carlo (MCMC) methods in extracting multiple CFF solutions for given kinematics, while C-VAIM effectively captures correlations among CFFs across different kinematic values, providing more constrained solutions. This study represents a crucial first step towards a comprehensive analysis pipeline towards the extraction of GPDs., Comment: 12 pages, 9 figures

In overview of the recent activity of the newly funded EXCLusives with AI and Machine learning (EXCLAIM) collaboration is presented. The main goal of the collaboration is to develop a framework to implement AI and machine learning techniques in problems emerging from the phenomenology of high energy exclusive scattering processes from nucleons and nuclei, maximizing the information that can be extracted from various sets of experimental data, while implementing theoretical constraints from lattice QCD. A specific perspective embraced by EXCLAIM is to use the methods of theoretical physics to understand the working of ML, beyond its standardized applications to physics analyses which most often rely on industrially provided tools, in an automated way., Comment: 9 pages, 3 figures

Large Language Models (LLMs) have demonstrated impressive capabilities in generating diverse and contextually rich text. However, concerns regarding copyright infringement arise as LLMs may inadvertently produce copyrighted material. In this paper, we first investigate the effectiveness of watermarking LLMs as a deterrent against the generation of copyrighted texts. Through theoretical analysis and empirical evaluation, we demonstrate that incorporating watermarks into LLMs significantly reduces the likelihood of generating copyrighted content, thereby addressing a critical concern in the deployment of LLMs. Additionally, we explore the impact of watermarking on Membership Inference Attacks (MIAs), which aim to discern whether a sample was part of the pretraining dataset and may be used to detect copyright violations. Surprisingly, we find that watermarking adversely affects the success rate of MIAs, complicating the task of detecting copyrighted text in the pretraining dataset. Finally, we propose an adaptive technique to improve the success rate of a recent MIA under watermarking. Our findings underscore the importance of developing adaptive methods to study critical problems in LLMs with potential legal implications., Comment: 21 pages, 6 figures

In this work we propose a new efficient basis for the electronic structure problem. The basis is based on the Muffin Tin Orbital (MTO) idea that the eigenstates of the Khon Sham (KS) Hamiltonian may we be expanded in terms of eigenstates of the spherically averaged KS Hamiltonian inside the so called Muffin Tin (MT) spheres and Bessel functions in the interstitial multiplied by appropriate spherical Harmonics. Here we use the fact that the solution to problem of finding the ground state electron density is most often done through an iterative process, where generically on the order of over 20 iterations are taken till the ground state electron density and energy converges to the lowest values allowed by the correlation and exchange functional for the fixed form of the external potential. We use eigenstate information from the previous convergence iteration to choose the energies of the eigenstates of the spherically averaged KS Hamiltonian. Furthermore within the Atomic Sphere Approximation (ASA) the energies of the Bessel functions do not matter as they are cancelled out. This is an efficient method aimed at studying the electronic structure of materials with large unit cells especially if they are of close packed form where ASA is particularly accurate., Comment: Comments welcome

A Kondo model with symplectic symmetry was recently put forward as the effective low-energy theory of a superconducting-island device coupled to multiple leads. This model, which possesses non-Fermi liquid physics and effective anyons, was argued to belong to the class of topological Kondo effects. Here, we clarify the extent of stability of its exotic fixed point using perturbative and numerical renormalization group in conjunction with bosonization and conformal field theory. In contrast to previous claims, we show that asymmetry in the coupling to the leads destabilizes the non-Fermi liquid. Other destabilizing perturbations include asymmetry in the superconducting pairing or internal energy of the individual quantum dots in the island. Nevertheless, these perturbations all generate the same relevant operators. Thus, only a small number of couplings need to be tuned individually, and these can be selected according to experimental convenience. Our results highlight a common misconception that anisotropy in the Kondo coupling is always irrelevant. As demonstrated, relevant terms will emerge whenever the group generators do not span the full space of impurity operators. This calls for a more detailed inspection of models that exhibit this property, such as large-spin impurities and SO(M) Kondo models, Comment: 26 pages, 7 figures

We introduce GoldFinch, a hybrid Linear Attention/Transformer sequence model that uses a new technique to efficiently generate a highly compressed and reusable KV-Cache in linear time and space with respect to sequence length. GoldFinch stacks our new GOLD transformer on top of an enhanced version of the Finch (RWKV-6) architecture. We train up to 1.5B parameter class models of the Finch, Llama, and GoldFinch architectures, and find dramatically improved modeling performance relative to both Finch and Llama. Our cache size savings increase linearly with model layer count, ranging from 756-2550 times smaller than the traditional transformer cache for common sizes, enabling inference of extremely large context lengths even on limited hardware. Although autoregressive generation has O(n) time complexity per token because of attention, pre-fill computation of the entire initial cache state for a submitted context costs only O(1) time per token due to the use of a recurrent neural network (RNN) to generate this cache. We release our trained weights and training code under the Apache 2.0 license for community use.

A junction of two 2/3 fractional quantum Hall (FQH) edges, with no charge tunneling between them, may exhibit Anderson localization of neutral modes. Manifestations of such localization in transport properties of the junction are explored. There are two competing localization channels, ``neutral-mode superconductivity'' and ``neutral-mode backscattering''. Localization in any of these channels leads to an effective theory of the junction that is characteristic for FQH effect of bosons, with a minimal integer excitation charge equal to two, and with elementary quasiparticle charge equal to 2/3. These values can be measured by studying shot noise in tunneling experiments. Under the assumption of ballistic transport in the arms connecting the junction to contacts, the two-terminal conductance of the junction is found to be 4/3 for the former localization channel and 1/3 for the latter. The four-terminal conductance matrix reveals in this regime a strong quantized drag between the edges induced by neutral-mode localization. The two localization channels lead to opposite signs of the drag conductance, equal to $\pm 1/4$, which can also be interpreted as a special type of Andreev scattering. Coherent random tunneling in arms of the device (which are segments of 2/3 edges) leads to strong mesoscopic fluctuations of the conductance matrix. In the case of fully equilibrated arms, transport via the junction is insensitive to neutral-mode localization: The two-terminal conductance is quantized to 2/3 and the drag is absent., Comment: 12 pages, 3 figures

Massive particles produced during inflation impact soft limits of primordial correlators. Searching for these signatures presents an exciting opportunity to uncover the particle spectrum in the inflationary epoch. We present non-perturbative methods to constrain intermediate-mass scalars ($0\leq m/H<3/2$, where $H$ is the inflationary Hubble scale) produced during inflation, which give rise to a power-law scaling in the squeezed primordial bispectrum. Exploiting the large-scale structure consistency relations and the separate universe approach, we derive models for the late-time squeezed matter bispectrum and collapsed matter trispectrum sourced by these fields. To validate our models, we run $N$-body simulations with the "Cosmological Collider" squeezed bispectrum for two different particle masses. Our models yield unbiased constraints on the amplitude of non-Gaussianity, $f_{\rm NL}^{\Delta}$, from the squeezed bispectrum and collapsed trispectrum deep into the non-linear regime ($k_{\rm max}\approx 2~h/{\rm Mpc}$ at $z=0$). We assess the information content of these summary statistics, emphasizing the importance of sample variance cancellation in the matter sector. We also study the scale-dependent halo bias in our simulations. For mass-selected halos, the non-Gaussian bias estimated from our simulations agrees with predictions based on (i) separate universe simulations and (ii) universal mass functions. With further work, these results can be used to search for inflationary massive particle production with upcoming galaxy surveys., Comment: 25 pages, 11 figures; comments welcome

Reversing a diffusion process by learning its score forms the heart of diffusion-based generative modeling and for estimating properties of scientific systems. The diffusion processes that are tractable center on linear processes with a Gaussian stationary distribution. This limits the kinds of models that can be built to those that target a Gaussian prior or more generally limits the kinds of problems that can be generically solved to those that have conditionally linear score functions. In this work, we introduce a family of tractable denoising score matching objectives, called local-DSM, built using local increments of the diffusion process. We show how local-DSM melded with Taylor expansions enables automated training and score estimation with nonlinear diffusion processes. To demonstrate these ideas, we use automated-DSM to train generative models using non-Gaussian priors on challenging low dimensional distributions and the CIFAR10 image dataset. Additionally, we use the automated-DSM to learn the scores for nonlinear processes studied in statistical physics.

For the one-dimensional Facilitated Exclusion Process with initial state a product measure of density $\rho=1/2-\delta$, $\delta\ge0$, there exists an infinite-time limiting state $\nu_\rho$ in which all particles are isolated and hence cannot move. We study the variance $V(L)$, under $\nu_\rho$, of the number of particles in an interval of $L$ sites. Under $\nu_{1/2}$ either all odd or all even sites are occupied, so that $V(L)=0$ for $L$ even and $V(L)=1/4$ for $L$ odd: the state is hyperuniform, since $V(L)$ grows more slowly than $L$. We prove that for densities approaching 1/2 from below there exist three regimes in $L$, in which the variance grows at different rates: for $L\gg\delta^{-2}$, $V(L)\simeq\rho(1-\rho)L$, just as in the initial state; for $A(\delta)\ll L\ll\delta^{-2}$, with $A(\delta)=\delta^{-2/3}$ for $L$ odd and $A(\delta)=1$ for $L$ even, $V(L)\simeq CL^{3/2}$ with $C=2\sqrt{2/\pi}/3$; and for $L\ll\delta^{-2/3}$ with $L$ odd, $V(L)\simeq1/4$. The analysis is based on a careful study of a renewal process with a long tail. Our study is motivated by simulation results showing similar behavior in higher dimensions; we discuss this background briefly., Comment: 21 pages, no figures

Within biological fluid dynamics, it is conventional to distinguish between "puller" and "pusher" microswimmers on the basis of the forward or aft location of the flagella relative to the cell body: typically, bacteria are pushers and algae are pullers. Here we note that since many pullers have "outboard" cilia or flagella displaced laterally from the cell centerline on both sides of the organism, there are two important subclasses whose far-field is that of a stresslet, but whose near field is qualitatively more complex. The ciliary beat creates not only a propulsive force but also swirling flows that can be represented by paired rotlets with two possible senses of rotation, either "feeders" that sweep fluid toward the cell apex, or "expellers" that push fluid away. Experimental studies of the rotifer $Brachionus~plicatilis$ in combination with earlier work on the green algae $Chlamydomonas~reinhardtii$ show that the two classes have markedly different interactions with surfaces. When swimming near a surface, expellers such as $C.~reinhardtii$ scatter from the wall, whereas a feeder like $B.~plicatilis$ stably attaches. This results in a stochastic "run-and-stick" locomotion, with periods of ballistic motion parallel to the surface interrupted by trapping at the surface., Comment: 16 pages, 6 figures, supplementary videos available at website of REG

Test set contamination, wherein test data from a benchmark ends up in a newer model's training set, is a well-documented obstacle for fair LLM evaluation and can quickly render benchmarks obsolete. To mitigate this, many recent benchmarks crowdsource new prompts and evaluations from human or LLM judges; however, these can introduce significant biases, and break down when scoring hard questions. In this work, we introduce a new benchmark for LLMs designed to be immune to both test set contamination and the pitfalls of LLM judging and human crowdsourcing. We release LiveBench, the first benchmark that (1) contains frequently-updated questions from recent information sources, (2) scores answers automatically according to objective ground-truth values, and (3) contains a wide variety of challenging tasks, spanning math, coding, reasoning, language, instruction following, and data analysis. To achieve this, LiveBench contains questions that are based on recently-released math competitions, arXiv papers, news articles, and datasets, and it contains harder, contamination-free versions of tasks from previous benchmarks such as Big-Bench Hard, AMPS, and IFEval. We evaluate many prominent closed-source models, as well as dozens of open-source models ranging from 0.5B to 110B in size. LiveBench is difficult, with top models achieving below 65% accuracy. We release all questions, code, and model answers. Questions will be added and updated on a monthly basis, and we will release new tasks and harder versions of tasks over time so that LiveBench can distinguish between the capabilities of LLMs as they improve in the future. We welcome community engagement and collaboration for expanding the benchmark tasks and models.

This paper examines the question of whether Large Language Models (LLMs) like ChatGPT possess minds, focusing specifically on whether they have a genuine folk psychology encompassing beliefs, desires, and intentions. We approach this question by investigating two key aspects: internal representations and dispositions to act. First, we survey various philosophical theories of representation, including informational, causal, structural, and teleosemantic accounts, arguing that LLMs satisfy key conditions proposed by each. We draw on recent interpretability research in machine learning to support these claims. Second, we explore whether LLMs exhibit robust dispositions to perform actions, a necessary component of folk psychology. We consider two prominent philosophical traditions, interpretationism and representationalism, to assess LLM action dispositions. While we find evidence suggesting LLMs may satisfy some criteria for having a mind, particularly in game-theoretic environments, we conclude that the data remains inconclusive. Additionally, we reply to several skeptical challenges to LLM folk psychology, including issues of sensory grounding, the "stochastic parrots" argument, and concerns about memorization. Our paper has three main upshots. First, LLMs do have robust internal representations. Second, there is an open question to answer about whether LLMs have robust action dispositions. Third, existing skeptical challenges to LLM representation do not survive philosophical scrutiny.

Large language models (LLMs) have shown an impressive ability to perform tasks believed to require thought processes. When the model does not document an explicit thought process, it becomes difficult to understand the processes occurring within its hidden layers and to determine if these processes can be referred to as reasoning. We introduce a novel and interpretable analysis of internal multi-hop reasoning processes in LLMs. We demonstrate that the prediction process for compositional reasoning questions can be modeled using a simple linear transformation between two semantic category spaces. We show that during inference, the middle layers of the network generate highly interpretable embeddings that represent a set of potential intermediate answers for the multi-hop question. We use statistical analyses to show that a corresponding subset of tokens is activated in the model's output, implying the existence of parallel reasoning paths. These observations hold true even when the model lacks the necessary knowledge to solve the task. Our findings can help uncover the strategies that LLMs use to solve reasoning tasks, offering insights into the types of thought processes that can emerge from artificial intelligence. Finally, we also discuss the implication of cognitive modeling of these results.

This study empirically tests the $\textit{Narrative Economics}$ hypothesis, which posits that narratives (ideas that are spread virally and affect public beliefs) can influence economic fluctuations. We introduce two curated datasets containing posts from X (formerly Twitter) which capture economy-related narratives (Data will be shared upon paper acceptance). Employing Natural Language Processing (NLP) methods, we extract and summarize narratives from the tweets. We test their predictive power for $\textit{macroeconomic}$ forecasting by incorporating the tweets' or the extracted narratives' representations in downstream financial prediction tasks. Our work highlights the challenges in improving macroeconomic models with narrative data, paving the way for the research community to realistically address this important challenge. From a scientific perspective, our investigation offers valuable insights and NLP tools for narrative extraction and summarization using Large Language Models (LLMs), contributing to future research on the role of narratives in economics.

Training large vision-language models requires extensive, high-quality image-text pairs. Existing web-scraped datasets, however, are noisy and lack detailed image descriptions. To bridge this gap, we introduce PixelProse, a comprehensive dataset of over 16M (million) synthetically generated captions, leveraging cutting-edge vision-language models for detailed and accurate descriptions. To ensure data integrity, we rigorously analyze our dataset for problematic content, including child sexual abuse material (CSAM), personally identifiable information (PII), and toxicity. We also provide valuable metadata such as watermark presence and aesthetic scores, aiding in further dataset filtering. We hope PixelProse will be a valuable resource for future vision-language research. PixelProse is available at https://huggingface.co/datasets/tomg-group-umd/pixelprose, Comment: pixelprose 16M dataset

Most public instruction finetuning datasets are relatively small compared to the closed source datasets used to train industry models. To study questions about finetuning at scale, such as curricula and learning rate cooldown schedules, there is a need for industrial-scale datasets. However, this scale necessitates a data generation process that is almost entirely automated. In this work, we study methods for generating large instruction datasets from a single prompt. With little human oversight, we get LLMs to write diverse sets of instruction examples ranging from simple completion tasks to complex multi-turn dialogs across a variety of subject areas. When finetuning a Llama-3 8B base model, our dataset meets or exceeds both WizardLM and Ultrachat on both knowledge-intensive leaderboard tasks as well as conversational evaluations. We release our dataset, the "generator" prompts that created it, and our finetuned model checkpoints., Comment: 9.5 pages, 6 Figures, and 3 tables in the main body. Dataset available at https://huggingface.co/datasets/tomg-group-umd/GenQA

Recent advancements in novel view synthesis have enabled real-time rendering speeds and high reconstruction accuracy. 3D Gaussian Splatting (3D-GS), a foundational point-based parametric 3D scene representation, models scenes as large sets of 3D Gaussians. Complex scenes can comprise of millions of Gaussians, amounting to large storage and memory requirements that limit the viability of 3D-GS on devices with limited resources. Current techniques for compressing these pretrained models by pruning Gaussians rely on combining heuristics to determine which ones to remove. In this paper, we propose a principled spatial sensitivity pruning score that outperforms these approaches. It is computed as a second-order approximation of the reconstruction error on the training views with respect to the spatial parameters of each Gaussian. Additionally, we propose a multi-round prune-refine pipeline that can be applied to any pretrained 3D-GS model without changing the training pipeline. After pruning 88.44% of the Gaussians, we observe that our PUP 3D-GS pipeline increases the average rendering speed of 3D-GS by 2.65$\times$ while retaining more salient foreground information and achieving higher image quality metrics than previous pruning techniques on scenes from the Mip-NeRF 360, Tanks & Temples, and Deep Blending datasets.

Large language models can memorize and repeat their training data, causing privacy and copyright risks. To mitigate memorization, we introduce a subtle modification to the next-token training objective that we call the goldfish loss. During training, a randomly sampled subset of tokens are excluded from the loss computation. These dropped tokens are not memorized by the model, which prevents verbatim reproduction of a complete chain of tokens from the training set. We run extensive experiments training billion-scale Llama-2 models, both pre-trained and trained from scratch, and demonstrate significant reductions in extractable memorization with little to no impact on downstream benchmarks., Comment: 9.5 pages, 8 figures, and 1 table in the main body. Code available at https://github.com/ahans30/goldfish-loss

Reinforcement learning with human feedback~(RLHF) is critical for aligning Large Language Models (LLMs) with human preference. Compared to the widely studied offline version of RLHF, \emph{e.g.} direct preference optimization (DPO), recent works have shown that the online variants achieve even better alignment. However, online alignment requires on-the-fly generation of new training data, which is costly, hard to parallelize, and suffers from varying quality and utility. In this paper, we propose a more efficient data exploration strategy for online preference tuning (OPTune), which does not rely on human-curated or pre-collected teacher responses but dynamically samples informative responses for on-policy preference alignment. During data generation, OPTune only selects prompts whose (re)generated responses can potentially provide more informative and higher-quality training signals than the existing responses. In the training objective, OPTune reweights each generated response (pair) by its utility in improving the alignment so that learning can be focused on the most helpful samples. Throughout our evaluations, OPTune'd LLMs maintain the instruction-following benefits provided by standard preference tuning whilst enjoying 1.27-1.56x faster training speed due to the efficient data exploration strategy., Comment: 16 pages, 7 figures

Eliciting reasoning capabilities from language models (LMs) is a critical direction on the path towards building intelligent systems. Most recent studies dedicated to reasoning focus on out-of-distribution performance on procedurally-generated synthetic benchmarks, bespoke-built to evaluate specific skills only. This trend makes results hard to transfer across publications, slowing down progress. Three years ago, a similar issue was identified and rectified in the field of neural algorithmic reasoning, with the advent of the CLRS benchmark. CLRS is a dataset generator comprising graph execution traces of classical algorithms from the Introduction to Algorithms textbook. Inspired by this, we propose CLRS-Text -- a textual version of these algorithmic traces. Out of the box, CLRS-Text is capable of procedurally generating trace data for thirty diverse, challenging algorithmic tasks across any desirable input distribution, while offering a standard pipeline in which any additional algorithmic tasks may be created in the benchmark. We fine-tune and evaluate various LMs as generalist executors on this benchmark, validating prior work and revealing a novel, interesting challenge for the LM reasoning community. Our code is available at https://github.com/google-deepmind/clrs/tree/master/clrs/_src/clrs_text., Comment: Preprint, under review. Comments welcome

Gamma-ray Bursts (GRBs) are one of the most energetic phenomena in the cosmos, whose study probes physics extremes beyond the reach of laboratories on Earth. Our quest to unravel the origin of these events and understand their underlying physics is far from complete. Central to this pursuit is the rapid classification of GRBs to guide follow-up observations and analysis across the electromagnetic spectrum and beyond. Here, we introduce a compelling approach for a new and robust GRB prompt classification. Leveraging self-supervised deep learning, we pioneer a previously unexplored data product to approach this task: the GRB waterfalls., Comment: Submitted for review to PRL

The Projected Augmented Wave (PAW) Method is based on generating a transform between the pseudo wavefunction and all electron wavefunction. For the accuracy of the method, it is important that the local part of the transform (in each atomic sphere $\mathbf{R}$) be over a complete basis set (with deviations from completeness leading to corrections to the total energy not computed within current implementations of PAW). Here we show how to make this basis much closer to complete without significant additional computational work without modifying the transformation in any way by extending the resolution of unity used for the transform to include more wavefunctions and having them transform via the identity., Comment: Comments welcome, V2: Many corrections

The poor performance of transformers on arithmetic tasks seems to stem in large part from their inability to keep track of the exact position of each digit inside of a large span of digits. We mend this problem by adding an embedding to each digit that encodes its position relative to the start of the number. In addition to the boost these embeddings provide on their own, we show that this fix enables architectural modifications such as input injection and recurrent layers to improve performance even further. With positions resolved, we can study the logical extrapolation ability of transformers. Can they solve arithmetic problems that are larger and more complex than those in their training data? We find that training on only 20 digit numbers with a single GPU for one day, we can reach state-of-the-art performance, achieving up to 99% accuracy on 100 digit addition problems. Finally, we show that these gains in numeracy also unlock improvements on other multi-step reasoning tasks including sorting and multiplication.

Large vision-language models (LVLMs) have achieved impressive results in various visual question-answering and reasoning tasks through vision instruction tuning on specific datasets. However, there is still significant room for improvement in the alignment between visual and language modalities. Previous methods to enhance this alignment typically require external models or data, heavily depending on their capabilities and quality, which inevitably sets an upper bound on performance. In this paper, we propose SIMA, a framework that enhances visual and language modality alignment through self-improvement, eliminating the needs for external models or data. SIMA leverages prompts from existing vision instruction tuning datasets to self-generate responses and employs an in-context self-critic mechanism to select response pairs for preference tuning. The key innovation is the introduction of three vision metrics during the in-context self-critic process, which can guide the LVLM in selecting responses that enhance image comprehension. Through experiments across 14 hallucination and comprehensive benchmarks, we demonstrate that SIMA not only improves model performance across all benchmarks but also achieves superior modality alignment, outperforming previous approaches., Comment: 15 pages, 8 figures

Cross-domain alignment refers to the task of mapping a concept from one domain to another. For example, ``If a \textit{doctor} were a \textit{color}, what color would it be?''. This seemingly peculiar task is designed to investigate how people represent concrete and abstract concepts through their mappings between categories and their reasoning processes over those mappings. In this paper, we adapt this task from cognitive science to evaluate the conceptualization and reasoning abilities of large language models (LLMs) through a behavioral study. We examine several LLMs by prompting them with a cross-domain mapping task and analyzing their responses at both the population and individual levels. Additionally, we assess the models' ability to reason about their predictions by analyzing and categorizing their explanations for these mappings. The results reveal several similarities between humans' and models' mappings and explanations, suggesting that models represent concepts similarly to humans. This similarity is evident not only in the model representation but also in their behavior. Furthermore, the models mostly provide valid explanations and deploy reasoning paths that are similar to those of humans., Comment: CogSci

In this work we present a new basis set for electronic structure calculations of crystalline solids using Density Functional theory (DFT) methods. In this construction we take advantage of the fact that most DFT calculations use a convergence loop in order to obtain the eigenstates of a final Khon Sham (KS) Hamiltonian matrix whose eigenstates also give the appropriate electron density needed to obtain the KS potential needed for that KS Hamiltonian matrix. Here we propose that for the basis of each step of the iteration we use the previous eigenstate basis in the interstitial region but augmented inside the MT sphere with the solution to the spherically averaged KS Hamiltonian for the energy window of that eigenstate. To reduce the number of times the KS potential needs to be solved inside the MT spheres it is advantageous to use energy windows and solve the KS Hamiltonian inside the MT region only once per window (at some energy inside the window) so that the KS Hamiltonian needs only be solved a small number of times per iteration, for practical applications on the order of 10 to 100 windows. This method combines the energy dependence of methods such as Projected Augmented Wave functions (PAW) with the ability of the basis set to adjust to the solid state (rather then atomic) environment of basis sets such as Linearized Augmented Plane Waves (LAPW)., Comment: Comments welcome

The transition from subAlfv\'enic to superAlfv\'enic flow in the solar atmosphere is examined by means of Parker Solar Probe (PSP) measurements during solar encounters 8 to 14. Around 220 subAlfv\'enic periods with a duration $\ge$ 10 minutes are identified. The distribution of their durations, heliocentric distances, and Alfv\'en Mach number are analyzed and compared with a global magnetohydrodynamic model of the solar corona and wind, which includes turbulence effects. The results are consistent with a patchy and fragmented morphology, and suggestive of a turbulent Alfv\'en zone within which the transition from subAlfv\'enic to superAlfv\'enic flow occurs over an extended range of helioradii. These results inform and establish context for detailed analyses of subAlfv\'enic coronal plasma that are expected to emerge from PSP's final mission phase, as well as for NASA's planned PUNCH mission.

Current datasets for long-form video understanding often fall short of providing genuine long-form comprehension challenges, as many tasks derived from these datasets can be successfully tackled by analyzing just one or a few random frames from a video. To address this issue, we present a novel dataset and benchmark, CinePile, specifically designed for authentic long-form video understanding. This paper details our innovative approach for creating a question-answer dataset, utilizing advanced LLMs with human-in-the-loop and building upon human-generated raw data. Our comprehensive dataset comprises 305,000 multiple-choice questions (MCQs), covering various visual and multimodal aspects, including temporal comprehension, understanding human-object interactions, and reasoning about events or actions within a scene. Additionally, we evaluate recent video-centric LLMs, both open-source and proprietary, on the test split of our dataset. The findings reveal that even state-of-the-art video-centric LLMs significantly lag behind human performance in these tasks, highlighting the complexity and challenge inherent in video understanding. The dataset is available at https://hf.co/datasets/tomg-group-umd/cinepile, Comment: Project page with all the artifacts - https://ruchitrawal.github.io/cinepile/. Updated version with results on Gemini Flash model and additional related work

We develop a methodology for analyzing language model task performance at the individual example level based on training data density estimation. Experiments with paraphrasing as a controlled intervention on finetuning data demonstrate that increasing the support in the training distribution for specific test queries results in a measurable increase in density, which is also a significant predictor of the performance increase caused by the intervention. Experiments with pretraining data demonstrate that we can explain a significant fraction of the variance in model perplexity via density measurements. We conclude that our framework can provide statistical evidence of the dependence of a target model's predictions on subsets of its training data, and can more generally be used to characterize the support (or lack thereof) in the training data for a given test task., Comment: 10 pages in the main body

How to compute the probability distribution of a detection time, i.e., of the time which a detector registers as the arrival time of a quantum particle, is a long-debated problem. In this regard, Bohmian mechanics provides in a straightforward way the distribution of the time at which the particle actually does arrive at a given surface in 3-space in the absence of detectors. However, as we discuss here, since the presence of detectors can change the evolution of the wave function and thus the particle trajectories, it cannot be taken for granted that the arrival time of the Bohmian trajectories in the absence of detectors agrees with the one in the presence of detectors, and even less with the detection time. In particular, we explain why certain distributions that Das and D\"urr [arXiv:1802.07141] presented as the distribution of the detection time in a case with spin, based on assuming that all three times mentioned coincide, is actually not what Bohmian mechanics predicts., Comment: 24 pages LaTeX, 2 figures; v2 appendices added, two paragraphs on page 2 added

This paper focuses on the opportunities and the ethical and societal risks posed by advanced AI assistants. We define advanced AI assistants as artificial agents with natural language interfaces, whose function is to plan and execute sequences of actions on behalf of a user, across one or more domains, in line with the user's expectations. The paper starts by considering the technology itself, providing an overview of AI assistants, their technical foundations and potential range of applications. It then explores questions around AI value alignment, well-being, safety and malicious uses. Extending the circle of inquiry further, we next consider the relationship between advanced AI assistants and individual users in more detail, exploring topics such as manipulation and persuasion, anthropomorphism, appropriate relationships, trust and privacy. With this analysis in place, we consider the deployment of advanced assistants at a societal scale, focusing on cooperation, equity and access, misinformation, economic impact, the environment and how best to evaluate advanced AI assistants. Finally, we conclude by providing a range of recommendations for researchers, developers, policymakers and public stakeholders.

In the paper, we address the following problem, often encountered in geometric constructions: having defined an orientation preserving diffeomorphism in a number of `patches' on a manifold $\mathcal{M}$, can we extend it to a diffeomorphism of the whole manifold $\mathcal{M}$? We give a positive answer, provided the `patches' are sufficiently regular. Then, we extend the same result to the case of bi-Lipschitz homeomorphisms and homeomorphisms.

Since their first discovery in the late 1960s, Gamma-ray bursts have attracted an exponentially growing interest from the international community due to their central role in the most highly debated open questions of the modern research of astronomy, astrophysics, cosmology, and fundamental physics. These range from the intimate nuclear composition of high density material within the core of ultra-dense neuron stars, to stellar evolution via the collapse of massive stars, the production and propagation of gravitational waves, as well as the exploration of the early Universe by unveiling first stars and galaxies (assessing also their evolution and cosmic re-ionization). GRBs have stimulated in the past $\sim$50 years the development of cutting-edge technological instruments for observations of high energy celestial sources from space, leading to the launch and successful operations of many different scientific missions (several of them still in data taking mode nowadays). In this review, we provide a brief description of the GRB-dedicated missions from space being designed and developed for the future. The list of these projects, not meant to be exhaustive, shall serve as a reference to interested readers to understand what is likely to come next to lead the further development of GRB research and associated phenomenology., Comment: Accepted for publication on Universe. Invited review, contribution to the Universe Special Issue "Recent Advances in Gamma Ray Astrophysics and Future Perspectives", P. Romano eds. (https://www.mdpi.com/journal/universe/special_issues/7299902Z97)

Thermal equilibrium states of local quantum many-body systems are notorious for their spatially decaying correlations, which place severe restrictions on the types of many-body entanglement structures that may be observed at finite temperatures. These restrictions may however be defied when an out-of-equilibrium steady state is considered instead. In this paper, we study the entanglement properties of free fermions on a one-dimensional lattice that contains a generic charge- and energy-conserving noninteracting impurity, and that is connected at its edges to two reservoirs with different equilibrium energy distributions. These distributions may differ in either temperature, chemical potential, or both, thereby inducing an external bias. We analytically derive exact asymptotic expressions for several quantum information measures -- the mutual information, its R\'enyi generalizations, and the fermionic negativity -- that quantify the correlation and entanglement between two subsystems located on opposite sides of the impurity. We show that all these measures scale (to a leading order) linearly with the overlap between one subsystem and the mirror image of the other (upon reflection of the latter about the impurity), independently of the distance between the subsystems. While a simple proportionality relation between the negativity and R\'enyi versions of the mutual information is observed to hold at zero temperature, it breaks down at finite temperatures, suggesting that these quantities represent strong long-range correlations of different origins. Our results generalize previous findings that were limited to the case of a chemical-potential bias at zero temperature, rigorously demonstrating that the effect of long-range volume-law entanglement is robust at finite temperatures., Comment: 16+4 pages, 4 figures. v2: Published version, 17+4 pages, 5 figures

In this work we make some progress on studying four center integrals for the Coulomb energy for both Hartree Fock (HF) and Density Functional Theory (DFT) calculations for small molecules. We consider basis wave functions of the form of an arbitrary radial wave function multiplied by a spherical harmonic and study four center Coulomb integrals for them. We reformulated these Coulomb four center integrals in terms of some derivatives of integrals of nearly factorable functions which then depend on the Bessel transform of the radial wave functions considered., Comment: Comments welcome

The success behind many pseudopotential methods such as Projected Augmented Waves (PAW) and Phillips-Kleinman pseudopotential methods is that these methods are all electron methods in disguise. For the Phillips Kleinman and PAW methods we show that there is an all electron formulation. In the all electron formulation there are regular low wavevector basis wave functions and several, specially chosen, high wavevector basis wave functions that are specialized to the atomic environment. Using this as inspiration here we propose a new, PAW method inspired, basis set for small molecules, where we use the recently introduced LO or lo (Localized orbitals) basis set is paired with a Gaussian basis, Gaussian type orbitals (GTO), for a hybrid basis for small molecules. Contracted Gaussian functions are also possible (CGFs). We also show how to extend this idea to Slater basis set, Slater type orbitals (STO), instead of GTO orbitals combined with LO or lo orbitals., Comment: Comments welcome, V2: Minor typos corrected; V3: Some extensions

We present Eagle (RWKV-5) and Finch (RWKV-6), sequence models improving upon the RWKV (RWKV-4) architecture. Our architectural design advancements include multi-headed matrix-valued states and a dynamic recurrence mechanism that improve expressivity while maintaining the inference efficiency characteristics of RNNs. We introduce a new multilingual corpus with 1.12 trillion tokens and a fast tokenizer based on greedy matching for enhanced multilinguality. We trained four Eagle models, ranging from 0.46 to 7.5 billion parameters, and two Finch models with 1.6 and 3.1 billion parameters and find that they achieve competitive performance across a wide variety of benchmarks. We release all our models on HuggingFace under the Apache 2.0 license. Models at: https://huggingface.co/RWKV Training code at: https://github.com/RWKV/RWKV-LM Inference code at: https://github.com/RWKV/ChatRWKV Time-parallel training code at: https://github.com/RWKV/RWKV-infctx-trainer

We evaluate ChatGPT's ability to solve algorithm problems from the CLRS benchmark suite that is designed for GNNs. The benchmark requires the use of a specified classical algorithm to solve a given problem. We find that ChatGPT outperforms specialist GNN models, using Python to successfully solve these problems. This raises new points in the discussion about learning algorithms with neural networks and how we think about what out of distribution testing looks like with web scale training data.

Generative models are now widely used by graphic designers and artists. Prior works have shown that these models remember and often replicate content from their training data during generation. Hence as their proliferation increases, it has become important to perform a database search to determine whether the properties of the image are attributable to specific training data, every time before a generated image is used for professional purposes. Existing tools for this purpose focus on retrieving images of similar semantic content. Meanwhile, many artists are concerned with style replication in text-to-image models. We present a framework for understanding and extracting style descriptors from images. Our framework comprises a new dataset curated using the insight that style is a subjective property of an image that captures complex yet meaningful interactions of factors including but not limited to colors, textures, shapes, etc. We also propose a method to extract style descriptors that can be used to attribute style of a generated image to the images used in the training dataset of a text-to-image model. We showcase promising results in various style retrieval tasks. We also quantitatively and qualitatively analyze style attribution and matching in the Stable Diffusion model. Code and artifacts are available at https://github.com/learn2phoenix/CSD.

It is commonplace to produce application-specific models by fine-tuning large pre-trained models using a small bespoke dataset. The widespread availability of foundation model checkpoints on the web poses considerable risks, including the vulnerability to backdoor attacks. In this paper, we unveil a new vulnerability: the privacy backdoor attack. This black-box privacy attack aims to amplify the privacy leakage that arises when fine-tuning a model: when a victim fine-tunes a backdoored model, their training data will be leaked at a significantly higher rate than if they had fine-tuned a typical model. We conduct extensive experiments on various datasets and models, including both vision-language models (CLIP) and large language models, demonstrating the broad applicability and effectiveness of such an attack. Additionally, we carry out multiple ablation studies with different fine-tuning methods and inference strategies to thoroughly analyze this new threat. Our findings highlight a critical privacy concern within the machine learning community and call for a reevaluation of safety protocols in the use of open-source pre-trained models.