Search

Text-to-image (T2I) diffusion models rely on encoded prompts to guide the image generation process. Typically, these prompts are extended to a fixed length by adding padding tokens before text encoding. Despite being a default practice, the influence of padding tokens on the image generation process has not been investigated. In this work, we conduct the first in-depth analysis of the role padding tokens play in T2I models. We develop two causal techniques to analyze how information is encoded in the representation of tokens across different components of the T2I pipeline. Using these techniques, we investigate when and how padding tokens impact the image generation process. Our findings reveal three distinct scenarios: padding tokens may affect the model's output during text encoding, during the diffusion process, or be effectively ignored. Moreover, we identify key relationships between these scenarios and the model's architecture (cross or self-attention) and its training process (frozen or trained text encoder). These insights contribute to a deeper understanding of the mechanisms of padding tokens, potentially informing future model design and training practices in T2I systems., Comment: Published in: NAACL 2025. Project webpage: https://padding-tone.github.io/

Automotive radar is a key component of sensing suites in autonomous driving (AD) and advanced driver-assist systems (ADAS). However, limited line-of-sight (LOS) significantly reduces radar efficiency in dense urban environments. Therefore, automotive radars need to extend their capabilities beyond LOS by localizing occluding and reflective surfaces and non-line-of-sight (NLOS) targets. This work addresses the NLOS target localization challenge by revisiting the NLOS radar signal propagation model and introducing a hybrid localization approach. The proposed approach first detects and localizes reflective surfaces, then identifies the LOS/NLOS propagation conditions, and finally localizes the target without prior scene knowledge, without using Doppler information, and without any auxiliary sensors. The proposed hybrid approach addresses the computational complexity challenge by integrating a physical radar electromagnetic wave propagation model with a deep neural network (DNN) to estimate occluding surface parameters. The efficiency of the proposed approach to localize the NLOS targets and to identify the NLOS/LOS propagation conditions is evaluated via simulations in a broad range of realistic automotive scenarios. Extending automotive radar sensing beyond LOS is expected to enhance the safety and reliability of autonomous and ADAS-equipped vehicles., Comment: This work has been submitted to the IEEE for possible publication

We present a study of the galaxy merger and interaction activity within the Hyperion Proto-supercluster at z~2.5 in an effort to assess the occurrence of galaxy mergers and interactions in contrast to the coeval field and their impact on the build up of stellar mass in high density environments at higher-z. For this work, we utilize data from the Charting Cluster Construction with VUDS and ORELSE Survey (C3VO) along with extensive spectroscopic and photometric datasets available for the COSMOS field, including the HST-Hyperion Survey. To evaluate potential merger and interaction activity, we measure the fraction of galaxies with close kinematic companions ($f_{ckc}$) both within Hyperion and the coeval field by means of a Monte Carlo (MC) methodology developed in this work that probabilistically employs our entire combined spectroscopic and photometric dataset. We validate our $f_{ckc}$ MC methodology on a simulated lightcone built from the GAlaxy Evolution and Assembly semi-analytic model, and we determine correction factors that account for the underlying spectroscopic sampling rate of our dataset. We find that galaxies in Hyperion have close kinematic companions $\gtrsim 2\times$ more than galaxies in the field and measure a corrected $f_{ckc}=49_{-7.8}^{+7.4}$% for Hyperion and a corrected $f_{ckc}=23_{-1.3}^{+1.2}$% for the surrounding field; a $>3\sigma$ difference. This increase in $f_{ckc}$ indicates an enhancement in the merger and interaction activity within Hyperion and matches the trend seen in other structures. The rate of merger and interactions within the field implied from our field $f_{ckc}$ measurement is well aligned with values measured from other observations in similar redshift ranges. The enhanced $f_{ckc}$ measured within Hyperion suggests that merger and interaction activity play an important role in the mass growth of galaxies in denser environments at higher z., Comment: 23 pages, 10 figures, 4 tables. Submitted to A&A. Comments welcome

We present first results and catalogs from the HST-Hyperion survey. This survey has collected 50 orbits of WFC3/F160W imaging and WFC3/G141 grism spectroscopy in the most overdense regions of the Hyperion proto-supercluster at $z\sim2.45$, which are analyzed in conjunction with the adjacent 56 orbits of WFC3/F140W imaging and WFC3/G141 grism spectroscopy from the 3D-HST survey. Sources were identified and spectra extracted using GRIZLI, which subsequently fit the combined grism data with object-matched photometric data from the COSMOS2020 catalog to obtain a redshift and best-fit spectral model. Each source was then visually inspected by multiple team members and quality flags were assigned. A total of 12814 objects with $m_{HST} \leq 25.0$ were inspected, of which 5629 (44%) have reliable redshifts from the grism data, which are sensitive to emission lines at a level of $\sim8.8 \times10^{-18}$ erg s$^{-1}$ cm$^{-2}$ ($1\sigma$). Comparison to high-quality ground-based spectroscopic redshifts yields a scatter of $\sigma_{\rm NMAD} = 0.0016$. The resulting catalogs contain 125 confirmed members of the Hyperion structure within $2.40

Classical shape descriptors such as Heat Kernel Signature (HKS), Wave Kernel Signature (WKS), and Signature of Histograms of OrienTations (SHOT), while widely used in shape analysis, exhibit sensitivity to mesh connectivity, sampling patterns, and topological noise. While differential geometry offers a promising alternative through its theory of differential invariants, which are theoretically guaranteed to be robust shape descriptors, the computation of these invariants on discrete meshes often leads to unstable numerical approximations, limiting their practical utility. We present a self-supervised learning approach for extracting geometric features from 3D surfaces. Our method combines synthetic data generation with a neural architecture designed to learn sampling-invariant features. By integrating our features into existing shape correspondence frameworks, we demonstrate improved performance on standard benchmarks including FAUST, SCAPE, TOPKIDS, and SHREC'16, showing particular robustness to topological noise and partial shapes.

Machine unlearning aims to remove the influence of problematic training data after a model has been trained. The primary challenge in machine unlearning is ensuring that the process effectively removes specified data without compromising the model's overall performance on the remaining dataset. Many existing machine unlearning methods address this challenge by carefully balancing gradient ascent on the unlearn data with the gradient descent on a retain set representing the training data. Here, we propose OrthoGrad, a novel approach that mitigates interference between the unlearn set and the retain set rather than competing ascent and descent processes. Our method projects the gradient of the unlearn set onto the subspace orthogonal to all gradients in the retain batch, effectively avoiding any gradient interference. We demonstrate the effectiveness of OrthoGrad on multiple machine unlearning benchmarks, including automatic speech recognition, outperforming competing methods., Comment: Under Review

Understanding and interpreting the internal representations of large language models (LLMs) remains an open challenge. Patchscopes introduced a method for probing internal activations by patching them into new prompts, prompting models to self-explain their hidden representations. We introduce Superscopes, a technique that systematically amplifies superposed features in MLP outputs (multilayer perceptron) and hidden states before patching them into new contexts. Inspired by the "features as directions" perspective and the Classifier-Free Guidance (CFG) approach from diffusion models, Superscopes amplifies weak but meaningful features, enabling the interpretation of internal representations that previous methods failed to explain-all without requiring additional training. This approach provides new insights into how LLMs build context and represent complex concepts, further advancing mechanistic interpretability.

This study addresses the challenge of self-supervised learning for 3D mesh analysis. It presents an new approach that uses random walks as a form of data augmentation to generate diverse representations of mesh surfaces. Furthermore, it employs a combination of contrastive and clustering losses. The contrastive learning framework maximizes similarity between augmented instances of the same mesh while minimizing similarity between different meshes. We integrate this with a clustering loss, enhancing class distinction across training epochs and mitigating training variance. Our model's effectiveness is evaluated using mean Average Precision (mAP) scores and a supervised SVM linear classifier on extracted features, demonstrating its potential for various downstream tasks such as object classification and shape retrieval.

Large language models (LLMs) hold substantial promise for clinical decision support. However, their widespread adoption in medicine, particularly in healthcare, is hindered by their propensity to generate false or misleading outputs, known as hallucinations. In high-stakes domains such as women's health (obstetrics & gynaecology), where errors in clinical reasoning can have profound consequences for maternal and neonatal outcomes, ensuring the reliability of AI-generated responses is critical. Traditional methods for quantifying uncertainty, such as perplexity, fail to capture meaning-level inconsistencies that lead to misinformation. Here, we evaluate semantic entropy (SE), a novel uncertainty metric that assesses meaning-level variation, to detect hallucinations in AI-generated medical content. Using a clinically validated dataset derived from UK RCOG MRCOG examinations, we compared SE with perplexity in identifying uncertain responses. SE demonstrated superior performance, achieving an AUROC of 0.76 (95% CI: 0.75-0.78), compared to 0.62 (0.60-0.65) for perplexity. Clinical expert validation further confirmed its effectiveness, with SE achieving near-perfect uncertainty discrimination (AUROC: 0.97). While semantic clustering was successful in only 30% of cases, SE remains a valuable tool for improving AI safety in women's health. These findings suggest that SE could enable more reliable AI integration into clinical practice, particularly in resource-limited settings where LLMs could augment care. This study highlights the potential of SE as a key safeguard in the responsible deployment of AI-driven tools in women's health, leading to safer and more effective digital health interventions., Comment: 15 pages, 6 tables

We present the luminosity function and volumetric rate of a sample of Type IIP supernovae (SNe) from the Zwicky Transient Facility Census of the Local Universe survey (CLU). This is the largest sample of Type IIP SNe from a systematic volume-limited survey to-date. The final sample includes 330 Type IIP SNe and 36 low-luminosity Type II (LLIIP) SNe with $M_{\textrm{r,peak}}>-16$ mag, which triples the literature sample of LLIIP SNe. The fraction of LLIIP SNe is $19^{+3}_{-4}\%$ of the total CLU Type IIP SNe population ($8^{+1}_{-2}\%$ of all core-collapse SNe). This implies that while LLIIP SNe likely represent the fate of core-collapse SNe of $8-12$ \Msun\ progenitors, they alone cannot account for the fate of all massive stars in this mass range. To derive an absolute rate, we estimate the ZTF pipeline efficiency as a function of the apparent magnitude and the local surface brightness. We derive a volumetric rate of $(3.9_{-0.4}^{+0.4}) \times 10^{4}\ \textrm{Gpc}^{-3}\ \textrm{yr}^{-1}$ for Type IIP SNe and $(7.3_{-0.6}^{+0.6}) \times 10^{3}\ \textrm{Gpc}^{-3}\ \textrm{yr}^{-1}$ for LLIIP SNe. Now that the rate of LLIIP SNe is robustly derived, the unresolved discrepancy between core-collapse SN rates and star-formation rates cannot be explained by LLIIP SNe alone., Comment: Submitted to PASP

Large language models (LLMs) have multilingual capabilities and can solve tasks across various languages. However, we show that current LLMs make key decisions in a representation space closest to English, regardless of their input and output languages. Exploring the internal representations with a logit lens for sentences in French, German, Dutch, and Mandarin, we show that the LLM first emits representations close to English for semantically-loaded words before translating them into the target language. We further show that activation steering in these LLMs is more effective when the steering vectors are computed in English rather than in the language of the inputs and outputs. This suggests that multilingual LLMs perform key reasoning steps in a representation that is heavily shaped by English in a way that is not transparent to system users., Comment: Main paper 9 pages; including appendix 48 pages

LLM developers have imposed technical interventions to prevent fine-tuning misuse attacks, attacks where adversaries evade safeguards by fine-tuning the model using a public API. Previous work has established several successful attacks against specific fine-tuning API defences. In this work, we show that defences of fine-tuning APIs that seek to detect individual harmful training or inference samples ('pointwise' detection) are fundamentally limited in their ability to prevent fine-tuning attacks. We construct 'pointwise-undetectable' attacks that repurpose entropy in benign model outputs (e.g. semantic or syntactic variations) to covertly transmit dangerous knowledge. Our attacks are composed solely of unsuspicious benign samples that can be collected from the model before fine-tuning, meaning training and inference samples are all individually benign and low-perplexity. We test our attacks against the OpenAI fine-tuning API, finding they succeed in eliciting answers to harmful multiple-choice questions, and that they evade an enhanced monitoring system we design that successfully detects other fine-tuning attacks. We encourage the community to develop defences that tackle the fundamental limitations we uncover in pointwise fine-tuning API defences.

Content creators often draw inspiration from multiple visual sources, combining distinct elements to craft new compositions. Modern computational approaches now aim to emulate this fundamental creative process. Although recent diffusion models excel at text-guided compositional synthesis, text as a medium often lacks precise control over visual details. Image-based composition approaches can capture more nuanced features, but existing methods are typically limited in the range of concepts they can capture, and require expensive training procedures or specialized data. We present IP-Composer, a novel training-free approach for compositional image generation that leverages multiple image references simultaneously, while using natural language to describe the concept to be extracted from each image. Our method builds on IP-Adapter, which synthesizes novel images conditioned on an input image's CLIP embedding. We extend this approach to multiple visual inputs by crafting composite embeddings, stitched from the projections of multiple input images onto concept-specific CLIP-subspaces identified through text. Through comprehensive evaluation, we show that our approach enables more precise control over a larger range of visual concept compositions., Comment: Project Page: https://ip-composer.github.io/IP-Composer/

With the rise of online applications, recommender systems (RSs) often encounter constraints in balancing exploration and exploitation. Such constraints arise when exploration is carried out by agents whose individual utility should be balanced with overall welfare. Recent work suggests that recommendations should be individually rational. Specifically, if agents have a default arm they would use, relying on the RS should yield each agent at least the reward of the default arm, conditioned on the knowledge available to the RS. Under this individual rationality constraint, striking a balance between exploration and exploitation becomes a complex planning problem. We assume a stochastic order of the rewards (e.g., Bernoulli, unit-variance Gaussian, etc.), and derive an approximately optimal algorithm. Our technique is based on an auxiliary Goal Markov Decision Process problem that is of independent interest. Additionally, we present an incentive-compatible version of our algorithm., Comment: arXiv admin note: substantial text overlap with arXiv:2006.04497

Lorentz invariance is a fundamental symmetry of spacetime and foundational to modern physics. One of its most important consequences is the constancy of the speed of light. This invariance, together with the geometry of spacetime, implies that no particle can move faster than the speed of light. In this article, we present the most stringent neutrino-based test of this prediction, using the highest energy neutrino ever detected to date, KM3-230213A. The arrival of this event, with an energy of $220^{+570}_{-110}\,\text{PeV}$, sets a constraint on $\delta \equiv c_\nu^2-1 < 4\times10^{-22}$.

Complex multi-step reasoning tasks, such as solving mathematical problems, remain challenging for large language models (LLMs). While outcome supervision is commonly used, process supervision via process reward models (PRMs) provides intermediate rewards to verify step-wise correctness in solution traces. However, as proxies for human judgement, PRMs suffer from reliability issues, including susceptibility to reward hacking. In this work, we propose leveraging uncertainty quantification (UQ) to enhance the reliability of step-wise verification with generative reward models for mathematical reasoning tasks. We introduce CoT Entropy, a novel UQ method that outperforms existing approaches in quantifying a PRM's uncertainty in step-wise verification. Our results demonstrate that incorporating uncertainty estimates improves the robustness of judge-LM PRMs, leading to more reliable verification.

Diffusion models enable high-quality and diverse visual content synthesis. However, they struggle to generate rare or unseen concepts. To address this challenge, we explore the usage of Retrieval-Augmented Generation (RAG) with image generation models. We propose ImageRAG, a method that dynamically retrieves relevant images based on a given text prompt, and uses them as context to guide the generation process. Prior approaches that used retrieved images to improve generation, trained models specifically for retrieval-based generation. In contrast, ImageRAG leverages the capabilities of existing image conditioning models, and does not require RAG-specific training. Our approach is highly adaptable and can be applied across different model types, showing significant improvement in generating rare and fine-grained concepts using different base models. Our project page is available at: https://rotem-shalev.github.io/ImageRAG

We introduce and formalize the notion of resource augmentation for maximin share allocations -- an idea that can be traced back to the seminal work of Budish [JPE 2011]. Specifically, given a fair division instance with $m$ goods and $n$ agents, we ask how many copies of the goods should be added in order to guarantee that each agent receives at least their original maximin share, or an approximation thereof. We establish a tight bound of $m/e$ copies for arbitrary monotone valuations. For additive valuations, we show that at most $\min\{n-2,\lfloor \frac{m}{3}\rfloor (1+o(1))\}$ copies suffice. For approximate-MMS in ordered instances, we give a tradeoff between the number of copies needed and the approximation guarantee. In particular, we prove that $\lfloor n/2 \rfloor$ copies suffice to guarantee a $6/7$-approximation to the original MMS, and $\lfloor n/3 \rfloor$ copies suffice for a $4/5$-approximation. Both results improve upon the best known approximation guarantees for additive valuations in the absence of copies.

The methylidyne cation (CH$^+$) and the methyl cation (CH$_3^+$) are building blocks of organic molecules, yet their coupled formation and excitation mechanisms remain mainly unprobed. The James Webb Space Telescope (JWST), with its high spatial resolution and good spectral resolution, provides unique access to the detection of these molecules. Our goal is to use the first detection of CH$^+$ and CH$_3^+$ rovibrational emission in the Orion Bar and in the protoplanetary disk d203-506, irradiated by the Trapezium cluster, to probe their formation and excitation mechanisms and constrain the physico-chemical conditions. We use spectro-imaging acquired using both the NIRSpec and MIRI-MRS instruments to study the CH$^+$ and CH$_3^+$ spatial distribution at very small scales, and compare it to excited H$_2$ emission. CH$^+$ and CH$_3^+$ emissions originate from the same region as highly excited H$_2$. Our comparison between the Bar and d203-506 reveals that both CH$^+$ and CH$_3^+$ excitation and/or formation are highly dependent on gas density. The excitation temperature of the observed CH$^+$ and CH$_3^+$ rovibrational lines is around $T$ ~ 1500 K in the Bar and $T$ ~ 800 K in d203-506. Moreover, the column densities derived from the rovibrational emission are less than 0.1 % of the total known (CH$^+$) and expected (CH$_3^+$) column densities. These results show that CH$^+$ and CH$_3^+$ level populations strongly deviate from ETL. CH$^+$ rovibrational emission can be explained by chemical formation pumping with excited H$_2$ via C$^+$ + H$_2^*$ = CH$^+$ + H. These results support a gas phase formation pathway of CH$^+$ and CH$_3^+$ via successive hydrogen abstraction reactions. However, we do not find any evidence of CH$_2^+$ emission in the JWST spectrum. Finally, observed CH$^+$ intensities coupled with chemical formation pumping model provide a diagnostic tool to trace the local density., Comment: Accepted on the 11/02/2025 in A&A

On February 13th, 2023, the KM3NeT/ARCA telescope detected a neutrino candidate with an estimated energy in the hundreds of PeVs. In this article, the observation of this ultra-high-energy neutrino is discussed in light of null observations above tens of PeV from the IceCube and Pierre Auger observatories. Performing a joint fit of all experiments under the assumption of an isotropic $E^{-2}$ flux, the best-fit single-flavour flux normalisation is $E^2 \Phi^{\rm 1f}_{\nu + \bar \nu} = 7.5 \times 10^{-10}~{\rm GeV cm^{-2} s^{-1} sr^{-1}}$ in the 90% energy range of the KM3NeT event. Furthermore, the ultra-high-energy data are then fit together with the IceCube measurements at lower energies, either with a single power law or with a broken power law, allowing for the presence of a new component in the spectrum. The joint fit including non-observations by other experiments in the ultra-high-energy region shows a slight preference for a break in the PeV regime if the ``High-Energy Starting Events'' sample is included, and no such preference for the other two IceCube samples investigated. A stronger preference for a break appears if only the KM3NeT data is considered in the ultra-high-energy region, though the flux resulting from such a fit would be inconsistent with null observations from IceCube and Pierre Auger. In all cases, the observed tension between KM3NeT and other datasets is of the order of $2.5\sigma-3\sigma$, and increased statistics are required to resolve this apparent tension and better characterise the neutrino landscape at ultra-high energies., Comment: 11 pages, 4 figures

We study the problem of semi-supervised learning on graphs in the regime where data labels are scarce or possibly corrupted. We propose an approach called $p$-conductance learning that generalizes the $p$-Laplace and Poisson learning methods by introducing an objective reminiscent of $p$-Laplacian regularization and an affine relaxation of the label constraints. This leads to a family of probability measure mincut programs that balance sparse edge removal with accurate distribution separation. Our theoretical analysis connects these programs to well-known variational and probabilistic problems on graphs (including randomized cuts, effective resistance, and Wasserstein distance) and provides motivation for robustness when labels are diffused via the heat kernel. Computationally, we develop a semismooth Newton-conjugate gradient algorithm and extend it to incorporate class-size estimates when converting the continuous solutions into label assignments. Empirical results on computer vision and citation datasets demonstrate that our approach achieves state-of-the-art accuracy in low label-rate, corrupted-label, and partial-label regimes., Comment: 29 pages, 7 figures

The KM3NeT observatory detected the most energetic neutrino candidate ever observed, with an energy between 72 PeV and 2.6 EeV at the 90% confidence level. The observed neutrino is likely of cosmic origin. In this article, it is investigated if the neutrino could have been produced within the Milky Way. Considering the low fluxes of the Galactic diffuse emission at these energies, the lack of a nearby potential Galactic particle accelerator in the direction of the event and the difficulty to accelerate particles to such high energies in Galactic systems, we conclude that if the event is indeed cosmic, it is most likely of extragalactic origin.

We present novel white-hat search engine optimization techniques based on genAI and demonstrate their empirical merits.

Self-consistency decoding enhances LLMs' performance on reasoning tasks by sampling diverse reasoning paths and selecting the most frequent answer. However, it is computationally expensive, as sampling many of these (lengthy) paths is required to increase the chances that the correct answer emerges as the most frequent one. To address this, we introduce Confidence-Informed Self-Consistency (CISC). CISC performs a weighted majority vote based on confidence scores obtained directly from the model. By prioritizing high-confidence paths, it can identify the correct answer with a significantly smaller sample size. When tested on nine models and four datasets, CISC outperforms self-consistency in nearly all configurations, reducing the required number of reasoning paths by over 40% on average. In addition, we introduce the notion of within-question confidence evaluation, after showing that standard evaluation methods are poor predictors of success in distinguishing correct and incorrect answers to the same question. In fact, the most calibrated confidence method proved to be the least effective for CISC. Lastly, beyond these practical implications, our results and analyses show that LLMs can effectively judge the correctness of their own outputs, contributing to the ongoing debate on this topic.

We study an interacting spinless quadratic band touching model that realizes a topological Mott insulating state. We quench the interaction from a value corresponding to the nematic insulator to that of the quantum anomalous Hall (QAH) ordered phase. We perform time-dependent Hartree-Fock simulations and show that after the quench the system realizes an excited Dirac semimetal state, which is however unstable and spontaneously evolves to a state with inhomogeneous nematic and QAH order parameters. The modulations form a stripe pattern that grows exponentially with time until the local Chern marker reaches unity. The alternating QAH order defines a domain structure with boundaries that host chiral sublattice currents., Comment: Comments are welcome

The study of fibre networks is of great importance due to their appearance in a wide range of natural and industrial processes. These networks are particularly complex because their properties depend not only on the characteristics of individual fibres but also on their collective arrangements within the network. In this study, we investigate the case of nylon fibre aggregates that were generated in a turbulent von Karman flow. The key properties of these aggregates are characterized, including their three-dimensional structures, analysed through X-ray tomography, and their mechanical responses. The results are compared with those observed for aegagropilae, a natural fibre aggregate commonly found on the Mediterranean sea shore, Comment: 17 pages, 8 figures

Superconducting traveling-wave parametric amplifiers (TWPAs) are increasingly used in various applications, including quantum computing, quantum sensing, and dark matter detection. However, one important characteristic of these amplifiers, gain compression, has not received much attention. As a result, there is a lack of comprehensive experimental exploration of this phenomenon in the existing literature. In this study, we present an experimental investigation of gain compression in a Josephson traveling-wave parametric amplifier based on a four-wave mixing process. We have implemented a novel setup to monitor the complex transmission of both the pump and signal tones, which allows us to simultaneously track pump depletion and signal amplification as functions of signal power and frequency across the entire bandwidth of the device. Our findings indicate that, while pump depletion occurs during gain compression, it is not the only mechanism involved in the saturation of a TWPA. Power-induced phase-matching processes also take place within the device. This study provides valuable insights for optimizing TWPAs for applications that require high total input power, such as multiplexed qubit readout or broadband photon emission.

Synthetic datasets are widely used in many applications, such as missing data imputation, examining non-stationary scenarios, in simulations, training data-driven models, and analyzing system robustness. Typically, synthetic data are based on historical data obtained from the observed system. The data needs to represent a specific behavior of the system, yet be new and diverse enough so that the system is challenged with a broad range of inputs. This paper presents a method, based on discrete Fourier transform, for generating synthetic time series with similar statistical moments for any given signal. The suggested method makes it possible to control the level of similarity between the given signal and the generated synthetic signals. Proof shows analytically that this method preserves the first two statistical moments of the input signal, and its autocorrelation function. The method is compared to known methods, ARMA, GAN, and CoSMoS. A large variety of environmental datasets with different temporal resolutions, and from different domains are used, testing the generality and flexibility of the method. A Python library implementing this method is made available as open-source software., Comment: Accepted for publication 27 November 2024. Code available at https://github.com/Al-Ofek/stsg.git

Evaluations of large language model (LLM) risks and capabilities are increasingly being incorporated into AI risk management and governance frameworks. Currently, most risk evaluations are conducted by designing inputs that elicit harmful behaviors from the system. However, a fundamental limitation of this approach is that the harmfulness of the behaviors identified during any particular evaluation can only lower bound the model's worst-possible-case behavior. As a complementary method for eliciting harmful behaviors, we propose evaluating LLMs with model tampering attacks which allow for modifications to latent activations or weights. We pit state-of-the-art techniques for removing harmful LLM capabilities against a suite of 5 input-space and 6 model tampering attacks. In addition to benchmarking these methods against each other, we show that (1) model resilience to capability elicitation attacks lies on a low-dimensional robustness subspace; (2) the attack success rate of model tampering attacks can empirically predict and offer conservative estimates for the success of held-out input-space attacks; and (3) state-of-the-art unlearning methods can easily be undone within 16 steps of fine-tuning. Together these results highlight the difficulty of removing harmful LLM capabilities and show that model tampering attacks enable substantially more rigorous evaluations than input-space attacks alone. We release models at https://huggingface.co/LLM-GAT

Standard reinforcement learning (RL) assumes that an agent can observe a reward for each state-action pair. However, in practical applications, it is often difficult and costly to collect a reward for each state-action pair. While there have been several works considering RL with trajectory feedback, it is unclear if trajectory feedback is inefficient for learning when trajectories are long. In this work, we consider a model named RL with segment feedback, which offers a general paradigm filling the gap between per-state-action feedback and trajectory feedback. In this model, we consider an episodic Markov decision process (MDP), where each episode is divided into $m$ segments, and the agent observes reward feedback only at the end of each segment. Under this model, we study two popular feedback settings: binary feedback and sum feedback, where the agent observes a binary outcome and a reward sum according to the underlying reward function, respectively. To investigate the impact of the number of segments $m$ on learning performance, we design efficient algorithms and establish regret upper and lower bounds for both feedback settings. Our theoretical and experimental results show that: under binary feedback, increasing the number of segments $m$ decreases the regret at an exponential rate; in contrast, surprisingly, under sum feedback, increasing $m$ does not reduce the regret significantly.

Oscillations of atmospheric muon and electron neutrinos produce tau neutrinos with energies in the GeV range, which can be observed by the ORCA detector of the KM3NeT neutrino telescope in the Mediterranean Sea. First measurements with ORCA6, an early subarray corresponding to about 5$\%$ of the final detector, are presented. A sample of 5828 neutrino candidates has been selected from the analysed exposure of 433 kton-years. The $\nu_\tau$ normalisation, defined as the ratio between the number of observed and expected tau neutrino events, is measured to be $S_\tau = 0.48^{+0.5}_{-0.33}$. This translates into a $\nu_\tau$ charged-current cross section measurement of $\sigma_{\tau}^{\text{meas}} = (2.5 ^{+2.6}_{-1.8}) \times 10^{-38}$ cm$^{2}$ nucleon$^{-1}$ at the median $\nu_\tau$ energy of 20.3 GeV. The result is consistent with the measurements of other experiments. In addition, the current limit on the non-unitarity parameter affecting the $\tau$-row of the neutrino mixing matrix was improved, with $\alpha_{33}>$ 0.95 at the 95$\%$ confidence level., Comment: 25 pages, 8 figures

This paper presents a comprehensive evaluation of GitHub Copilot's deployment and impact on developer productivity at Zoominfo, a leading Go-To-Market (GTM) Intelligence Platform. We describe our systematic four-phase approach to evaluating and deploying GitHub Copilot across our engineering organization, involving over 400 developers. Our analysis combines both quantitative metrics, focusing on acceptance rates of suggestions given by GitHub Copilot and qualitative feedback given by developers through developer satisfaction surveys. The results show an average acceptance rate of 33% for suggestions and 20% for lines of code, with high developer satisfaction scores of 72%. We also discuss language-specific performance variations, limitations, and lessons learned from this medium-scale enterprise deployment. Our findings contribute to the growing body of knowledge about AI-assisted software development in enterprise settings., Comment: 25 pages, 11 figures

Video encoders optimize compression for human perception by minimizing reconstruction error under bit-rate constraints. In many modern applications such as autonomous driving, an overwhelming majority of videos serve as input for AI systems performing tasks like object recognition or segmentation, rather than being watched by humans. It is therefore useful to optimize the encoder for a downstream task instead of for perceptual image quality. However, a major challenge is how to combine such downstream optimization with existing standard video encoders, which are highly efficient and popular. Here, we address this challenge by controlling the Quantization Parameters (QPs) at the macro-block level to optimize the downstream task. This granular control allows us to prioritize encoding for task-relevant regions within each frame. We formulate this optimization problem as a Reinforcement Learning (RL) task, where the agent learns to balance long-term implications of choosing QPs on both task performance and bit-rate constraints. Notably, our policy does not require the downstream task as an input during inference, making it suitable for streaming applications and edge devices such as vehicles. We demonstrate significant improvements in two tasks, car detection, and ROI (saliency) encoding. Our approach improves task performance for a given bit rate compared to traditional task agnostic encoding methods, paving the way for more efficient task-aware video compression.

In the era of precision measurements of neutrino oscillation parameters, it is necessary for experiments to disentangle discrepancies that may indicate physics beyond the Standard Model in the neutrino sector. KM3NeT/ORCA is a water Cherenkov neutrino detector under construction and anchored at the bottom of the Mediterranean Sea. The detector is designed to study the oscillations of atmospheric neutrinos and determine the neutrino mass ordering. This paper focuses on the initial configuration of ORCA, referred to as ORCA6, which comprises six out of the foreseen 115 detection units of photosensors. A high-purity neutrino sample was extracted during 2020 and 2021, corresponding to an exposure of 433 kton-years. This sample is analysed following a binned log-likelihood approach to search for invisible neutrino decay, in a three-flavour neutrino oscillation scenario, where the third neutrino mass state $\nu_3$ decays into an invisible state, e.g. a sterile neutrino. The resulting best fit of the invisible neutrino decay parameter is $\alpha_3 = 0.92^{+1.08}_{-0.57}\times 10^{-4}~\mathrm{eV^2}$, corresponding to a scenario with $\theta_{23}$ in the second octant and normal neutrino mass ordering. The results are consistent with the Standard Model, within a $2.1\,\sigma$ interval., Comment: 20 pages, 8 figures

In this paper, we study $C^*$-envelopes of finite-dimensional operator algebras arising from constrained interpolation problems on the unit disc. In particular, we consider interpolation problems for the algebra $H^\infty_{\text{node}}$ that consists of bounded analytic functions on the unit disk that satisfy $ f(0) = f(\lambda)$ for some $0 \neq \lambda \in \mathbb{D}$. We show that there exist choices of four interpolation nodes that exclude both $0$ and $\lambda$, such that if $I$ is the ideal of functions that vanish at the interpolation nodes, then $C^*_e(H^\infty_{\text{node}}/I)$ is infinite-dimensional. This differs markedly from the behavior of the algebra corresponding to interpolation nodes that contain the constrained points studied in the literature. Additionally, we use the distance formula to provide a completely isometric embedding of $C^*_e(H^\infty_{\text{node}}/I)$ for any choice of $n$ interpolation nodes that do not contain the constrained points into $M_n(G^2_{nc})$, where $G^2_{nc}$ is Brown's noncommutative Grassmannian.

Current visual SLAM systems face significant challenges in balancing computational efficiency with robust loop closure handling. Traditional approaches require careful manual tuning and incur substantial computational overhead, while learning-based methods either lack explicit loop closure capabilities or implement them through computationally expensive methods. We present AutoLoop, a novel approach that combines automated curriculum learning with efficient fine-tuning for visual SLAM systems. Our method employs a DDPG (Deep Deterministic Policy Gradient) agent to dynamically adjust loop closure weights during training, eliminating the need for manual hyperparameter search while significantly reducing the required training steps. The approach pre-computes potential loop closure pairs offline and leverages them through an agent-guided curriculum, allowing the model to adapt efficiently to new scenarios. Experiments conducted on TartanAir for training and validated across multiple benchmarks including KITTI, EuRoC, ICL-NUIM and TUM RGB-D demonstrate that AutoLoop achieves comparable or superior performance while reducing training time by an order of magnitude compared to traditional approaches. AutoLoop provides a practical solution for rapid adaptation of visual SLAM systems, automating the weight tuning process that traditionally requires multiple manual iterations. Our results show that this automated curriculum strategy not only accelerates training but also maintains or improves the model's performance across diverse environmental conditions.

The detection of deepfake speech has become increasingly challenging with the rapid evolution of deepfake technologies. In this paper, we propose a hybrid architecture for deepfake speech detection, combining a self-supervised learning framework for feature extraction with a classifier head to form an end-to-end model. Our approach incorporates both audio-level and feature-level augmentation techniques. Specifically, we introduce and analyze various masking strategies for augmenting raw audio spectrograms and for enhancing feature representations during training. We incorporate compression augmentations during the pretraining phase of the feature extractor to address the limitations of small, single-language datasets. We evaluate the model on the ASVSpoof5 (ASVSpoof 2024) challenge, achieving state-of-the-art results in Track 1 under closed conditions with an Equal Error Rate of 4.37%. By employing different pretrained feature extractors, the model achieves an enhanced EER of 3.39%. Our model demonstrates robust performance against unseen deepfake attacks and exhibits strong generalization across different codecs.

As AI systems become more capable, widely deployed, and increasingly autonomous in critical areas such as cybersecurity, biological research, and healthcare, ensuring their safety and alignment with human values is paramount. Machine unlearning -- the ability to selectively forget or suppress specific types of knowledge -- has shown promise for privacy and data removal tasks, which has been the primary focus of existing research. More recently, its potential application to AI safety has gained attention. In this paper, we identify key limitations that prevent unlearning from serving as a comprehensive solution for AI safety, particularly in managing dual-use knowledge in sensitive domains like cybersecurity and chemical, biological, radiological, and nuclear (CBRN) safety. In these contexts, information can be both beneficial and harmful, and models may combine seemingly harmless information for harmful purposes -- unlearning this information could strongly affect beneficial uses. We provide an overview of inherent constraints and open problems, including the broader side effects of unlearning dangerous knowledge, as well as previously unexplored tensions between unlearning and existing safety mechanisms. Finally, we investigate challenges related to evaluation, robustness, and the preservation of safety features during unlearning. By mapping these limitations and open challenges, we aim to guide future research toward realistic applications of unlearning within a broader AI safety framework, acknowledging its limitations and highlighting areas where alternative approaches may be required.

Inspired by bouncing drop experiments that revealed how macroscopic systems can exhibit wave-particle properties akin to quantum systems, we explore here a new wave-particle system based on internal gravity waves propagating in a density-stratified fluid. Recent experiments on oscillating particles (called ludions) in a density-stratified fluid have shown that fluid-particle interactions lead to a symmetry breaking, resulting in the particle self-locomotion in the horizontal plane. Here, we show using 2D numerical simulations that this symmetry breaking is driven by a Doppler-effect force, which we also predict analytically using a minimal theoretical model. In wall-bounded domains, this Doppler force interacts with a Casimir-like potential rapidly developing due to wave reflection on the distant walls. Despite the Doppler force, this potential captures the ludion trajectories in fixed points or chaotic attractors located at or around the potential wells. We finally propose a minimal dynamical model recapitulating the main aspects of the ludion dynamics. Our work introduces the ludion as a new promising system for the exploration of macroscopic wave-particle duality., Comment: 8 pages,6 figures

In many applications, both data samples and features have underlying hierarchical structures. However, existing methods for learning these latent structures typically focus on either samples or features, ignoring possible coupling between them. In this paper, we introduce a coupled hierarchical structure learning method using tree-Wasserstein distance (TWD). Our method jointly computes TWDs for samples and features, representing their latent hierarchies as trees. We propose an iterative, unsupervised procedure to build these sample and feature trees based on diffusion geometry, hyperbolic geometry, and wavelet filters. We show that this iterative procedure converges and empirically improves the quality of the constructed trees. The method is also computationally efficient and scales well in high-dimensional settings. Our method can be seamlessly integrated with hyperbolic graph convolutional networks (HGCN). We demonstrate that our method outperforms competing approaches in sparse approximation and unsupervised Wasserstein distance learning on several word-document and single-cell RNA-sequencing datasets. In addition, integrating our method into HGCN enhances performance in link prediction and node classification tasks.

A natural question, inspired by the famous Ryser-Brualdi-Stein Conjecture, is to determine the largest positive integer $g(r,n)$ such that every collection of $n$ matchings, each of size $n$, in an $r$-partite $r$-uniform hypergraph contains a rainbow matching of size $g(r,n)$. The parameter $g'(r,n)$ is defined identically with the exception that the host hypergraph is not required to be $r$-partite. In this note, we improve the best known lower bounds on $g'(r,n)$ for all $r \geq 4$ and the upper bounds on $g(r,n)$ for all $r \geq 3$, provided $n$ is sufficiently large. More precisely, we show that if $r\ge3$ then $$\frac{2n}{r+1}-\Theta_r(1)\le g'(r,n)\le g(r,n)\le n-\Theta_r(n^{1-\frac{1}{r}}).$$ Interestingly, while it has been conjectured that $g(2,n)=g'(2,n)=n-1$, our results show that if $r\ge3$ then $g(r,n)$ and $g'(r,n)$ are bounded away from $n$ by a function which grows in $n$. We also prove analogous bounds for the related problem where we are interested in the smallest size $s$ for which any collection of $n$ matchings of size $s$ in an ($r$-partite) $r$-uniform hypergraph contains a rainbow matching of size $n$.

Most current captioning systems use language models trained on data from specific settings, such as image-based captioning via Amazon Mechanical Turk, limiting their ability to generalize to other modality distributions and contexts. This limitation hinders performance in tasks like audio or video captioning, where different semantic cues are needed. Addressing this challenge is crucial for creating more adaptable and versatile captioning frameworks applicable across diverse real-world contexts. In this work, we introduce a method to adapt captioning networks to the semantics of alternative settings, such as capturing audibility in audio captioning, where it is crucial to describe sounds and their sources. Our framework consists of two main components: (i) a frozen captioning system incorporating a language model (LM), and (ii) a text classifier that guides the captioning system. The classifier is trained on a dataset automatically generated by GPT-4, using tailored prompts specifically designed to enhance key aspects of the generated captions. Importantly, the framework operates solely during inference, eliminating the need for further training of the underlying captioning model. We evaluate the framework on various models and modalities, with a focus on audio captioning, and report promising results. Notably, when combined with an existing zero-shot audio captioning system, our framework improves its quality and sets state-of-the-art performance in zero-shot audio captioning.

Personalizing text-to-image models to generate images of specific subjects across diverse scenes and styles is a rapidly advancing field. Current approaches often face challenges in maintaining a balance between identity preservation and alignment with the input text prompt. Some methods rely on a single textual token to represent a subject, which limits expressiveness, while others employ richer representations but disrupt the model's prior, diminishing prompt alignment. In this work, we introduce Nested Attention, a novel mechanism that injects a rich and expressive image representation into the model's existing cross-attention layers. Our key idea is to generate query-dependent subject values, derived from nested attention layers that learn to select relevant subject features for each region in the generated image. We integrate these nested layers into an encoder-based personalization method, and show that they enable high identity preservation while adhering to input text prompts. Our approach is general and can be trained on various domains. Additionally, its prior preservation allows us to combine multiple personalized subjects from different domains in a single image., Comment: Project page at https://snap-research.github.io/NestedAttention/

Text-to-video diffusion models have shown remarkable progress in generating coherent video clips from textual descriptions. However, the interplay between motion, structure, and identity representations in these models remains under-explored. Here, we investigate how self-attention query features (a.k.a. Q features) simultaneously govern motion, structure, and identity and examine the challenges arising when these representations interact. Our analysis reveals that Q affects not only layout, but that during denoising Q also has a strong effect on subject identity, making it hard to transfer motion without the side-effect of transferring identity. Understanding this dual role enabled us to control query feature injection (Q injection) and demonstrate two applications: (1) a zero-shot motion transfer method that is 20 times more efficient than existing approaches, and (2) a training-free technique for consistent multi-shot video generation, where characters maintain identity across multiple video shots while Q injection enhances motion fidelity., Comment: (1) Project page: https://research.nvidia.com/labs/par/MotionByQueries/ (2) The methods and results in section 5, "Consistent multi-shot video generation", are based on the arXiv version 1 (v1) of this work. Here, in version 2 (v2), we extend and further analyze those findings to efficient motion transfer

Adding Object into images based on text instructions is a challenging task in semantic image editing, requiring a balance between preserving the original scene and seamlessly integrating the new object in a fitting location. Despite extensive efforts, existing models often struggle with this balance, particularly with finding a natural location for adding an object in complex scenes. We introduce Add-it, a training-free approach that extends diffusion models' attention mechanisms to incorporate information from three key sources: the scene image, the text prompt, and the generated image itself. Our weighted extended-attention mechanism maintains structural consistency and fine details while ensuring natural object placement. Without task-specific fine-tuning, Add-it achieves state-of-the-art results on both real and generated image insertion benchmarks, including our newly constructed "Additing Affordance Benchmark" for evaluating object placement plausibility, outperforming supervised methods. Human evaluations show that Add-it is preferred in over 80% of cases, and it also demonstrates improvements in various automated metrics., Comment: Project page is at https://research.nvidia.com/labs/par/addit/

The practical use of text-to-image generation has evolved from simple, monolithic models to complex workflows that combine multiple specialized components. While workflow-based approaches can lead to improved image quality, crafting effective workflows requires significant expertise, owing to the large number of available components, their complex inter-dependence, and their dependence on the generation prompt. Here, we introduce the novel task of prompt-adaptive workflow generation, where the goal is to automatically tailor a workflow to each user prompt. We propose two LLM-based approaches to tackle this task: a tuning-based method that learns from user-preference data, and a training-free method that uses the LLM to select existing flows. Both approaches lead to improved image quality when compared to monolithic models or generic, prompt-independent workflows. Our work shows that prompt-dependent flow prediction offers a new pathway to improving text-to-image generation quality, complementing existing research directions in the field., Comment: Project website: https://comfygen-paper.github.io/

Recent advancements in generative modeling now enable the creation of 4D content (moving 3D objects) controlled with text prompts. 4D generation has large potential in applications like virtual worlds, media, and gaming, but existing methods provide limited control over the appearance and geometry of generated content. In this work, we introduce a method for animating user-provided 3D objects by conditioning on textual prompts to guide 4D generation, enabling custom animations while maintaining the identity of the original object. We first convert a 3D mesh into a ``static" 4D Neural Radiance Field (NeRF) that preserves the visual attributes of the input object. Then, we animate the object using an Image-to-Video diffusion model driven by text. To improve motion realism, we introduce an incremental viewpoint selection protocol for sampling perspectives to promote lifelike movement and a masked Score Distillation Sampling (SDS) loss, which leverages attention maps to focus optimization on relevant regions. We evaluate our model in terms of temporal coherence, prompt adherence, and visual fidelity and find that our method outperforms baselines that are based on other approaches, achieving up to threefold improvements in identity preservation measured using LPIPS scores, and effectively balancing visual quality with dynamic content.

Diffusion-based generative models have become the standard for image generation. ODE-based samplers and flow matching models improve efficiency, in comparison to diffusion models, by reducing sampling steps through learned vector fields. However, the theoretical foundations of flow matching models remain limited, particularly regarding the convergence of individual sample trajectories at terminal time - a critical property that impacts sample quality and being critical assumption for models like the consistency model. In this paper, we advance the theory of flow matching models through a comprehensive analysis of sample trajectories, centered on the denoiser that drives ODE dynamics. We establish the existence, uniqueness and convergence of ODE trajectories at terminal time, ensuring stable sampling outcomes under minimal assumptions. Our analysis reveals how trajectories evolve from capturing global data features to local structures, providing the geometric characterization of per-sample behavior in flow matching models. We also explain the memorization phenomenon in diffusion-based training through our terminal time analysis. These findings bridge critical gaps in understanding flow matching models, with practical implications for sampling stability and model design., Comment: Fixed typos and improved organization for clarity