Search

Sparse autoencoders (SAEs) are an unsupervised method for learning a sparse decomposition of a neural network's latent representations into seemingly interpretable features. Despite recent excitement about their potential, research applications outside of industry are limited by the high cost of training a comprehensive suite of SAEs. In this work, we introduce Gemma Scope, an open suite of JumpReLU SAEs trained on all layers and sub-layers of Gemma 2 2B and 9B and select layers of Gemma 2 27B base models. We primarily train SAEs on the Gemma 2 pre-trained models, but additionally release SAEs trained on instruction-tuned Gemma 2 9B for comparison. We evaluate the quality of each SAE on standard metrics and release these results. We hope that by releasing these SAE weights, we can help make more ambitious safety and interpretability research easier for the community. Weights and a tutorial can be found at https://huggingface.co/google/gemma-scope and an interactive demo can be found at https://www.neuronpedia.org/gemma-scope, Comment: 12 main text pages, and 14 pages of acknowledgements, references and appendices

The proliferation of misleading visualizations online, particularly during critical events like public health crises and elections, poses a significant risk. This study investigates the capability of GPT-4 models (4V, 4o, and 4o mini) to detect misleading visualizations. Utilizing a dataset of tweet-visualization pairs containing various visual misleaders, we test these models under four experimental conditions with different levels of guidance. We show that GPT-4 models can detect misleading visualizations with moderate accuracy without prior training (naive zero-shot) and that performance notably improves when provided with definitions of misleaders (guided zero-shot). However, a single prompt engineering technique does not yield the best results for all misleader types. Specifically, providing the models with misleader definitions and examples (guided few-shot) proves more effective for reasoning misleaders, while guided zero-shot performs better for design misleaders. This study underscores the feasibility of using large vision-language models to detect visual misinformation and the importance of prompt engineering for optimized detection accuracy., Comment: 5 pages, 2 figures; accepted by IEEE VIS 2024 (https://ieeevis.org/year/2024/program/paper_v-short-1177.html)

Understanding whether and to what extent large language models (LLMs) have memorised training data has important implications for the reliability of their output and the privacy of their training data. In order to cleanly measure and disentangle memorisation from other phenomena (e.g. in-context learning), we create an experimental framework that is based on repeatedly exposing LLMs to random strings. Our framework allows us to better understand the dynamics, i.e., the behaviour of the model, when repeatedly exposing it to random strings. Using our framework, we make several striking observations: (a) we find consistent phases of the dynamics across families of models (Pythia, Phi and Llama2), (b) we identify factors that make some strings easier to memorise than others, and (c) we identify the role of local prefixes and global context in memorisation. We also show that sequential exposition to different random strings has a significant effect on memorisation. Our results, often surprising, have significant downstream implications in the study and usage of LLMs.

This paper presents Hardware Description Language Generative Pre-trained Transformers (HDL-GPT), a novel approach that leverages the vast repository of open-source High Definition Language (HDL) codes to train superior quality large code models. The core premise of this paper is the hypothesis that high-quality HDL is all you need to create models with exceptional performance and broad zero-shot generalization abilities. The paper elucidates the methods employed for the curation and augmentation of large corpora from open-source HDL code, transforming highly variable quality data into high-quality data through careful prompting and context maintenance. We demonstrate that the careful selection, filtering, and augmentation of data across HDLs can yield powerful models that surpass current state-of-the-art models. We also explore the impact of different fine-tuning methods on the quality of results. We describe experimental results across a range of fine-tuned SOTA LLMs, substantiating our claims. We demonstrate improvements of 50% to 200% over SOTA HDL models on current benchmarks in tasks ranging from HDL circuit explanations, code generation, formal and simulation testbench creation, triaging bugs, and fixing them. HDL-GPT opens new avenues for the development of advanced model training techniques for circuit design tasks., Comment: DAC 2024 Invited Paper

Annotation and classification of legal text are central components of empirical legal research. Traditionally, these tasks are often delegated to trained research assistants. Motivated by the advances in language modeling, empirical legal scholars are increasingly turning to prompting commercial models, hoping that it will alleviate the significant cost of human annotation. Despite growing use, our understanding of how to best utilize large language models for legal tasks remains limited. We conduct a comprehensive study of 260 legal text classification tasks, nearly all new to the machine learning community. Starting from GPT-4 as a baseline, we show that it has non-trivial but highly varied zero-shot accuracy, often exhibiting performance that may be insufficient for legal work. We then demonstrate that a lightly fine-tuned Llama 3 model vastly outperforms GPT-4 on almost all tasks, typically by double-digit percentage points. We find that larger models respond better to fine-tuning than smaller models. A few tens to hundreds of examples suffice to achieve high classification accuracy. Notably, we can fine-tune a single model on all 260 tasks simultaneously at a small loss in accuracy relative to having a separate model for each task. Our work points to a viable alternative to the predominant practice of prompting commercial models. For concrete legal tasks with some available labeled data, researchers are better off using a fine-tuned open-source model.

Sparse autoencoders (SAEs) are a promising unsupervised approach for identifying causally relevant and interpretable linear features in a language model's (LM) activations. To be useful for downstream tasks, SAEs need to decompose LM activations faithfully; yet to be interpretable the decomposition must be sparse -- two objectives that are in tension. In this paper, we introduce JumpReLU SAEs, which achieve state-of-the-art reconstruction fidelity at a given sparsity level on Gemma 2 9B activations, compared to other recent advances such as Gated and TopK SAEs. We also show that this improvement does not come at the cost of interpretability through manual and automated interpretability studies. JumpReLU SAEs are a simple modification of vanilla (ReLU) SAEs -- where we replace the ReLU with a discontinuous JumpReLU activation function -- and are similarly efficient to train and run. By utilising straight-through-estimators (STEs) in a principled manner, we show how it is possible to train JumpReLU SAEs effectively despite the discontinuous JumpReLU function introduced in the SAE's forward pass. Similarly, we use STEs to directly train L0 to be sparse, instead of training on proxies such as L1, avoiding problems like shrinkage., Comment: v2: new appendix H comparing kernel functions & bug-fixes to pseudo-code in Appendix J v3: further bug-fix to pseudo-code in Appendix J

We present the results of an extensive evaluation of strip and pixel AC-LGAD sensors tested with a 120 GeV proton beam, focusing on the influence of design parameters on the sensor temporal and spatial resolutions. Results show that reducing the thickness of pixel sensors significantly enhances their time resolution, with 20 $\mu$m-thick sensors achieving around 20 ps. Uniform performance is attainable with optimized sheet resistance, making these sensors ideal for future timing detectors. Conversely, 20 $\mu$m-thick strip sensors exhibit higher jitter than similar pixel sensors, negatively impacting time resolution, despite reduced Landau fluctuations with respect to the 50 $\mu$m-thick versions. Additionally, it is observed that a low resistivity in strip sensors limits signal size and time resolution, whereas higher resistivity improves performance. This study highlights the importance of tuning the n$^{+}$ sheet resistance and suggests that further improvements should target specific applications like the Electron-Ion Collider or other future collider experiments. In addition, the detailed performance of four AC-LGADs sensor designs is reported as examples of possible candidates for specific detector applications. These advancements position AC-LGADs as promising candidates for future 4D tracking systems, pending the development of specialized readout electronics.

The mechanism of antiproton-nucleus annihilation at rest is not fully understood, despite substantial previous experimental and theoretical work. In this study we used slow extracted, sub-keV antiprotons from the ASACUSA apparatus at CERN to measure the charged particle multiplicities and their energy deposits from antiproton annihilations at rest on three different nuclei: carbon, molybdenum and gold. The results are compared with predictions from different models in the simulation tools Geant4 and FLUKA. A model that accounts for all the observed features is still missing, as well as measurements at low energies, to validate such models., Comment: 25 pages + 2 cover pages, 16 figures

Language Models (LMs) excel in natural language processing tasks for English but show reduced performance in most other languages. This problem is commonly tackled by continually pre-training and fine-tuning these models for said languages. A significant issue in this process is the limited vocabulary coverage in the original model's tokenizer, leading to inadequate representation of new languages and necessitating an expansion of the tokenizer. The initialization of the embeddings corresponding to new vocabulary items presents a further challenge. Current strategies require cross-lingual embeddings and lack a solid theoretical foundation as well as comparisons with strong baselines. In this paper, we first establish theoretically that initializing within the convex hull of existing embeddings is a good initialization, followed by a novel but simple approach, Constrained Word2Vec (CW2V), which does not require cross-lingual embeddings. Our study evaluates different initialization methods for expanding RoBERTa and LLaMA 2 across four languages and five tasks. The results show that CW2V performs equally well or even better than more advanced techniques. Additionally, simpler approaches like multivariate initialization perform on par with these advanced methods indicating that efficient large-scale multilingual continued pretraining can be achieved even with simpler initialization methods., Comment: Under review

Transfer learning is a powerful technique for knowledge-sharing between different tasks. Recent work has found that the representations of models with certain invariances, such as to adversarial input perturbations, achieve higher performance on downstream tasks. These findings suggest that invariance may be an important property in the context of transfer learning. However, the relationship of invariance with transfer performance is not fully understood yet and a number of questions remain. For instance, how important is invariance compared to other factors of the pretraining task? How transferable is learned invariance? In this work, we systematically investigate the importance of representational invariance for transfer learning, as well as how it interacts with other parameters during pretraining. To do so, we introduce a family of synthetic datasets that allow us to precisely control factors of variation both in training and test data. Using these datasets, we a) show that for learning representations with high transfer performance, invariance to the right transformations is as, or often more, important than most other factors such as the number of training samples, the model architecture and the identity of the pretraining classes, b) show conditions under which invariance can harm the ability to transfer representations and c) explore how transferable invariance is between tasks. The code is available at \url{https://github.com/tillspeicher/representation-invariance-transfer}., Comment: Published at TMLR 2024

We aim at characterizing the population of low-luminosity X-ray sources in the Galactic plane by studying their X-ray spectra and periodic signals in the light curves. We are performing an X-ray survey of the Galactic disk using XMM-Newton, and the source XMMU J173029.8-330920 was serendipitously discovered in our campaign. We performed a follow-up observation of the source using our pre-approved NuSTAR target of opportunity time. We used various phenomenological models in xspec for the X-ray spectral modeling. We also computed the Lomb-Scargle periodogram to search for X-ray periodicity. A Monte Carlo method was used to simulate 1000 artificial light curves to estimate the significance of the detected period. We also searched for X-ray, optical, and infrared counterparts of the source in various catalogs. The spectral modeling indicates the presence of an intervening cloud with $N_{\rm H}\sim(1.5-2.3)\times10^{23}\ \rm cm^{-2}$ that partially absorbs the incoming X-ray photons. The X-ray spectra are best fit by a model representing emission from a collisionally ionized diffuse gas with plasma temperature $kT=26^{+11}_{-5}$ keV. Furthermore, an Fe $K_{\alpha}$ line at $6.47^{+0.13}_{-0.06}$ keV was detected with an equivalent width of the line of $312\pm104$ eV. We discovered a coherent pulsation with a period of $521.7\pm0.8$ s. The 3-10 keV pulsed fraction of the source is around $\sim$50-60\%. The hard X-ray emission with plasma temperature $kT=26^{+11}_{-5}$ keV, iron $K_{\alpha}$ emission at 6.4 keV and a periodic behavior of $521.7\pm0.8$ s suggest XMMU J173029.8-33092 to be an intermediate polar. We estimated the mass of the central white dwarf to be $0.94-1.4\ M_{\odot}$ by assuming a distance to the source of $\sim1.4-5$ kpc., Comment: 7 pages, 4 figures, accepted for publication in A&A

It has been suggested that a $dS_{d+1}$ spacetime of radius $R_{ds}$ has a holographic dual, living at future space-like infinity ${\cal I}^+$, with the bulk wave function being dual to the partition function of the boundary theory, [arXiv:astro-ph/0210603v5]. We consider some aspects of this correspondence. For under damped scalars with mass $M^2R_{ds}^2>{d^2\over4}$, belonging to the principal series, we show that for the Bunch Davies vacuum a suitable source in the boundary theory can be identified in terms of the coherent state representation of the wave function. We argue that terms in the resulting correlation functions, which are independent of the late time cut-off, satisfy the Ward identities of a conformal field theory. We also discuss other ways to identify sources, both in the under damped and the over damped case, where $M^2R_{ds}^2<{d^2\over4}$, and argue that these too can lead to correlators satisfying the Ward identities of a CFT. Some comments on the violation of reflection positivity, and the cut-off dependent terms, along with some explicit checks and sample calculations, are also included.

In this paper, we introduce some new notions of aging based on geometric, harmonic means of failure rate and aging intensity function. We define a generalized version of aging functions called specific interval-average geometric hazard rate, specific interval-average harmonic hazard rate. We focus on some characterization results and their inter-relationships among the resulting non-parametric classes of distributions. Monotonic nature of so defined aging classes are exhibited by some well known probability distributions. Probabilistic orders based on these functions are taken up for further study. The work is illustrated through case studies and a simulated data having applications in reliability/survival analysis., Comment: 22 pages, 3 figures

Decomposing model activations into interpretable components is a key open problem in mechanistic interpretability. Sparse autoencoders (SAEs) are a popular method for decomposing the internal activations of trained transformers into sparse, interpretable features, and have been applied to MLP layers and the residual stream. In this work we train SAEs on attention layer outputs and show that also here SAEs find a sparse, interpretable decomposition. We demonstrate this on transformers from several model families and up to 2B parameters. We perform a qualitative study of the features computed by attention layers, and find multiple families: long-range context, short-range context and induction features. We qualitatively study the role of every head in GPT-2 Small, and estimate that at least 90% of the heads are polysemantic, i.e. have multiple unrelated roles. Further, we show that Sparse Autoencoders are a useful tool that enable researchers to explain model behavior in greater detail than prior work. For example, we explore the mystery of why models have so many seemingly redundant induction heads, use SAEs to motivate the hypothesis that some are long-prefix whereas others are short-prefix, and confirm this with more rigorous analysis. We use our SAEs to analyze the computation performed by the Indirect Object Identification circuit (Wang et al.), validating that the SAEs find causally meaningful intermediate variables, and deepening our understanding of the semantics of the circuit. We open-source the trained SAEs and a tool for exploring arbitrary prompts through the lens of Attention Output SAEs.

Despite their widespread use, the mechanisms by which large language models (LLMs) represent and regulate uncertainty in next-token predictions remain largely unexplored. This study investigates two critical components believed to influence this uncertainty: the recently discovered entropy neurons and a new set of components that we term token frequency neurons. Entropy neurons are characterized by an unusually high weight norm and influence the final layer normalization (LayerNorm) scale to effectively scale down the logits. Our work shows that entropy neurons operate by writing onto an unembedding null space, allowing them to impact the residual stream norm with minimal direct effect on the logits themselves. We observe the presence of entropy neurons across a range of models, up to 7 billion parameters. On the other hand, token frequency neurons, which we discover and describe here for the first time, boost or suppress each token's logit proportionally to its log frequency, thereby shifting the output distribution towards or away from the unigram distribution. Finally, we present a detailed case study where entropy neurons actively manage confidence in the setting of induction, i.e. detecting and continuing repeated subsequences., Comment: 25 pages, 14 figures

A key goal in mechanistic interpretability is circuit analysis: finding sparse subgraphs of models corresponding to specific behaviors or capabilities. However, MLP sublayers make fine-grained circuit analysis on transformer-based language models difficult. In particular, interpretable features -- such as those found by sparse autoencoders (SAEs) -- are typically linear combinations of extremely many neurons, each with its own nonlinearity to account for. Circuit analysis in this setting thus either yields intractably large circuits or fails to disentangle local and global behavior. To address this we explore transcoders, which seek to faithfully approximate a densely activating MLP layer with a wider, sparsely-activating MLP layer. We successfully train transcoders on language models with 120M, 410M, and 1.4B parameters, and find them to perform at least on par with SAEs in terms of sparsity, faithfulness, and human-interpretability. We then introduce a novel method for using transcoders to perform weights-based circuit analysis through MLP sublayers. The resulting circuits neatly factorize into input-dependent and input-invariant terms. Finally, we apply transcoders to reverse-engineer unknown circuits in the model, and we obtain novel insights regarding the greater-than circuit in GPT2-small. Our results suggest that transcoders can prove effective in decomposing model computations involving MLPs into interpretable circuits. Code is available at https://github.com/jacobdunefsky/transcoder_circuits., Comment: 28 pages, 6 figures, 3 tables, 2 algorithms. Under review

Conversational large language models are fine-tuned for both instruction-following and safety, resulting in models that obey benign requests but refuse harmful ones. While this refusal behavior is widespread across chat models, its underlying mechanisms remain poorly understood. In this work, we show that refusal is mediated by a one-dimensional subspace, across 13 popular open-source chat models up to 72B parameters in size. Specifically, for each model, we find a single direction such that erasing this direction from the model's residual stream activations prevents it from refusing harmful instructions, while adding this direction elicits refusal on even harmless instructions. Leveraging this insight, we propose a novel white-box jailbreak method that surgically disables refusal with minimal effect on other capabilities. Finally, we mechanistically analyze how adversarial suffixes suppress propagation of the refusal-mediating direction. Our findings underscore the brittleness of current safety fine-tuning methods. More broadly, our work showcases how an understanding of model internals can be leveraged to develop practical methods for controlling model behavior.

Visual Question Answering (VQA) is an important task in multimodal AI, and it is often used to test the ability of vision-language models to understand and reason on knowledge present in both visual and textual data. However, most of the current VQA models use datasets that are primarily focused on English and a few major world languages, with images that are typically Western-centric. While recent efforts have tried to increase the number of languages covered on VQA datasets, they still lack diversity in low-resource languages. More importantly, although these datasets often extend their linguistic range via translation or some other approaches, they usually keep images the same, resulting in narrow cultural representation. To address these limitations, we construct CVQA, a new Culturally-diverse multilingual Visual Question Answering benchmark, designed to cover a rich set of languages and cultures, where we engage native speakers and cultural experts in the data collection process. As a result, CVQA includes culturally-driven images and questions from across 28 countries on four continents, covering 26 languages with 11 scripts, providing a total of 9k questions. We then benchmark several Multimodal Large Language Models (MLLMs) on CVQA, and show that the dataset is challenging for the current state-of-the-art models. This benchmark can serve as a probing evaluation suite for assessing the cultural capability and bias of multimodal models and hopefully encourage more research efforts toward increasing cultural awareness and linguistic diversity in this field.

We demonstrate the formation of orbital and spin Hall effects (OHE/SHE) in the 1T phase of non-magnetic transition metal dichalcogenides. With the aid of density functional theory calculations and model Hamiltonian studies on MX$_2$ (M = Pt, Pd and X = S, Se, and Te), we show an intrinsic orbital Hall conductivity ($\sim 10^3 \hbar /e\ \Omega^{-1}cm^{-1}$) , which primarily emerges due to the orbital texture around the valleys in the momentum space. The robust spin-orbit coupling in these systems induces a sizable SHE out of OHE. Furthermore, to resemble the typical experimental setups, where the magnetic overlayers produce a proximity magnetic field, we examine the effect of magnetic field on OHE and SHE and showed that the latter can be doubled in these class of compounds. With a giant OHE and tunable SHE, the 1T-TMDs are promising candidates for spin and orbital driven quantum devices such as SOT-MRAM, spin nano-oscillators, spin logic devices etc., and to carry out spin-charge conversion experiments for fundamental research.

Disentangling model activations into meaningful features is a central problem in interpretability. However, the absence of ground-truth for these features in realistic scenarios makes validating recent approaches, such as sparse dictionary learning, elusive. To address this challenge, we propose a framework for evaluating feature dictionaries in the context of specific tasks, by comparing them against \emph{supervised} feature dictionaries. First, we demonstrate that supervised dictionaries achieve excellent approximation, control, and interpretability of model computations on the task. Second, we use the supervised dictionaries to develop and contextualize evaluations of unsupervised dictionaries along the same three axes. We apply this framework to the indirect object identification (IOI) task using GPT-2 Small, with sparse autoencoders (SAEs) trained on either the IOI or OpenWebText datasets. We find that these SAEs capture interpretable features for the IOI task, but they are less successful than supervised features in controlling the model. Finally, we observe two qualitative phenomena in SAE training: feature occlusion (where a causally relevant concept is robustly overshadowed by even slightly higher-magnitude ones in the learned features), and feature over-splitting (where binary features split into many smaller, less interpretable features). We hope that our framework will provide a useful step towards more objective and grounded evaluations of sparse dictionary learning methods.

Recent work has found that sparse autoencoders (SAEs) are an effective technique for unsupervised discovery of interpretable features in language models' (LMs) activations, by finding sparse, linear reconstructions of LM activations. We introduce the Gated Sparse Autoencoder (Gated SAE), which achieves a Pareto improvement over training with prevailing methods. In SAEs, the L1 penalty used to encourage sparsity introduces many undesirable biases, such as shrinkage -- systematic underestimation of feature activations. The key insight of Gated SAEs is to separate the functionality of (a) determining which directions to use and (b) estimating the magnitudes of those directions: this enables us to apply the L1 penalty only to the former, limiting the scope of undesirable side effects. Through training SAEs on LMs of up to 7B parameters we find that, in typical hyper-parameter ranges, Gated SAEs solve shrinkage, are similarly interpretable, and require half as many firing features to achieve comparable reconstruction fidelity., Comment: 15 main text pages, 22 appendix pages

Activation patching is a popular mechanistic interpretability technique, but has many subtleties regarding how it is applied and how one may interpret the results. We provide a summary of advice and best practices, based on our experience using this technique in practice. We include an overview of the different ways to apply activation patching and a discussion on how to interpret the results. We focus on what evidence patching experiments provide about circuits, and on the choice of metric and associated pitfalls., Comment: A tutorial on activation patching. 13 pages, 2 figures

We present the design and characterization of the first full chain precision timing prototype ASIC, named ETL Readout Chip version 1 (ETROC1) for the CMS MTD endcap timing layer (ETL) upgrade. The ETL utilizes Low Gain Avalanche Diode (LGAD) sensors to detect charged particles, with the goal to achieve a time resolution of 40 - 50 ps per hit, and 30 - 40 ps per track with hits from two detector layers. The ETROC1 is composed of a 5 x 5 pixel array and peripheral circuits. The pixel array includes a 4 x 4 active pixel array with an H-tree shaped network delivering clock and charge injection signals. Each active pixel is composed of various components, including a bump pad, a charge injection circuit, a pre-amplifier, a discriminator, a digital-to-analog converter, and a time-to-digital converter. These components play essential roles as the front-end link in processing LGAD signals and measuring timing-related information. The peripheral circuits provide clock signals and readout functionalities. The size of the ETROC1 chip is 7 mm x 9 mm. ETROC1 has been fabricated in a 65 nm CMOS process, and extensively tested under stimuli of charge injection, infrared laser, and proton beam. The time resolution of bump-bonded ETROC1 + LGAD chipsets reaches 42 - 46 ps per hit in the beam test., Comment: 14 pages, 15 figures

We propose an approach for estimating the latent knowledge embedded inside large language models (LLMs). We leverage the in-context learning (ICL) abilities of LLMs to estimate the extent to which an LLM knows the facts stored in a knowledge base. Our knowledge estimator avoids reliability concerns with previous prompting-based methods, is both conceptually simpler and easier to apply, and we demonstrate that it can surface more of the latent knowledge embedded in LLMs. We also investigate how different design choices affect the performance of ICL-based knowledge estimation. Using the proposed estimator, we perform a large-scale evaluation of the factual knowledge of a variety of open source LLMs, like OPT, Pythia, Llama(2), Mistral, Gemma, etc. over a large set of relations and facts from the Wikidata knowledge base. We observe differences in the factual knowledge between different model families and models of different sizes, that some relations are consistently better known than others but that models differ in the precise facts they know, and differences in the knowledge of base models and their finetuned counterparts.

The challenge of selecting the most relevant features of a given dataset arises ubiquitously in data analysis and dimensionality reduction. However, features found to be of high importance for the entire dataset may not be relevant to subsets of interest, and vice versa. Given a feature selector and a fixed decomposition of the data into subsets, we describe a method for identifying selected features which are compatible with the decomposition into subsets. We achieve this by re-framing the problem of finding compatible features to one of finding sections of a suitable quiver representation. In order to approximate such sections, we then introduce a Laplacian operator for quiver representations valued in Hilbert spaces. We provide explicit bounds on how the spectrum of a quiver Laplacian changes when the representation and the underlying quiver are modified in certain natural ways. Finally, we apply this machinery to the study of peak-calling algorithms which measure chromatin accessibility in single-cell data. We demonstrate that eigenvectors of the associated quiver Laplacian yield locally and globally compatible features., Comment: 40 pages, 7 figures

Neutron stars are the dense and highly magnetic relics of supernova explosions of massive stars. The quest to constrain the Equation of State (EoS) of ultra-dense matter and thereby probe the behavior of matter inside neutron stars, is one of the core goals of modern physics and astrophysics. A promising method involves investigating the long-term cooling of neutron stars, and comparing theoretical predictions with various sources at different ages. However, limited observational data, and uncertainties in source ages and distances, have hindered this approach. In this work, re-analyzing XMM-Newton and Chandra data from dozens of thermally emitting isolated neutron stars, we have identified three sources with unexpectedly cold surface temperatures for their young ages. To investigate these anomalies, we conducted magneto-thermal simulations across diverse mass and magnetic fields, considering three different EoS. We found that the "minimal" cooling model, failed to explain the observations, regardless the mass and the magnetic field, as validated by a machine learning classification method. The existence of these young cold neutron stars suggests that any dense matter EoS must be compatible with a fast cooling process at least in certain mass ranges, eliminating a significant portion of current EoS options according to recent meta-modelling analysis., Comment: 60 pages, 10 figures, 7 tables; Accepted for publication; first four authors are corresponding authors

Realizations of the holographic correspondence in String/M theory typically involve spacetimes of the form $AdS \times Y$ where $Y$ is some internal space which geometrizes an internal symmetry of the dual field theory, hereafter referred to as an "$R$ symmetry". It has been speculated that areas of Ryu-Takayanagi surfaces anchored on the boundary of a subregion of $Y$, and smeared over the base space of the dual field theory, quantify entanglement of internal degrees of freedom. A natural candidate for the corresponding operators are linear combinations of operators with definite $R$ charge with coefficients given by the "spherical harmonics'' of the internal space: this is natural when the product spaces appear as IR geometries of higher dimensional AdS spaces. We study clustering properties of such operators both for pure $AdS \times Y$ and for flow geometries, where $AdS \times Y$ arises in the IR from a different spacetime in the UV, for example higher dimensional AdS or asymptotically flat spacetime. We show, in complete generality, that the two point functions of such operators separated along the internal space obey clustering properties at scales larger than the $AdS$ scale. For non-compact $Y$, this provides a notion of approximate locality. When $Y$ is compact, clustering happens only when the size of $Y$ is parametrically larger than the $AdS$ scale. This latter situation is realized in flow geometries where the product spaces arise in the IR from an asymptotically AdS geometry at UV, but not typically when they arise near black hole horizons in asymptotically flat spacetimes. We discuss the significance of this result for entanglement and comment on the role of color degrees of freedom., Comment: 24 pages, no figures

In the domain of computer vision, semantic segmentation emerges as a fundamental application within machine learning, wherein individual pixels of an image are classified into distinct semantic categories. This task transcends traditional accuracy metrics by incorporating uncertainty quantification, a critical measure for assessing the reliability of each segmentation prediction. Such quantification is instrumental in facilitating informed decision-making, particularly in applications where precision is paramount. Within this nuanced framework, the metric known as PAvPU (Patch Accuracy versus Patch Uncertainty) has been developed as a specialized tool for evaluating entropy-based uncertainty in image segmentation tasks. However, our investigation identifies three core deficiencies within the PAvPU framework and proposes robust solutions aimed at refining the metric. By addressing these issues, we aim to enhance the reliability and applicability of uncertainty quantification, especially in scenarios that demand high levels of safety and accuracy, thus contributing to the advancement of semantic segmentation methodologies in critical applications., Comment: Premature Submission: accidentally submitted before it was ready

We present a Rabi-type measurement of two ground-state hydrogen hyperfine transitions performed in two opposite external magnetic field directions. This puts first constraints at the level of 2.3 10^-21 GeV on a set of coefficients of the Standard Model Extension, which were not measured by previous experiments. Moreover, we introduce a novel method, applicable to antihydrogen hyperfine spectroscopy in a beam, that determines the zero-field hyperfine transition frequency from the two transitions measured at the same magnetic field. Our value, nu_0 = 1.420 405 751 63(63) GHz, is in agreement with literature at a relative precision of 0.44 ppb. This is the highest precision achieved on hydrogen in a beam, improving over previous results by a factor of 6., Comment: 8 pages, 5 figures, 2 tables, corresponding author: chloe.m@cern.ch (C. Malbrunot)

For many years, it has been claimed that the Galactic ridge X-ray emission at the Galactic Center (GC) is truly diffuse in nature. However, with the advancement of modern X-ray satellites, it has been found that most of the diffuse emission is actually comprised of thousands of previously unresolved X-ray point sources. Further, many studies suggest that a vast majority of these X-ray point sources are magnetic cataclysmic variables (mCVs) and active binaries. One unambiguous way to identify these mCVs and other sources is by detecting their X-ray periodicity. Therefore, we systematically searched for periodic X-ray sources in the inner Galactic disk, including the GC region. We have used data from our ongoing XMM-Newton Heritage survey of the inner Galactic disk ($350^{\circ}\lesssim l\lesssim+7^{\circ}$ and $-1^{\circ}\lesssim b\lesssim +1^{\circ}$) plus the XMM-Newton archival observations of the GC. We computed the Lomb-Scargle periodogram of the light curves for the periodicity search. We fitted the energy spectra of the sources using a simple power-law model plus three Gaussians at 6.4, 6.7, and 6.9 keV for the iron $K$ emission complex. We detected periodicity in 26 sources. For 14 of them, this is the first discovery of periodicity. For the other 12 sources, we found periods similar to those already known, indicating no significant period evolution. We also searched for the Gaia counterparts of the periodic sources to estimate their distances using the Gaia parallax. We found a likely Gaia counterpart for seven sources. We have classified the sources into four categories based on the periodicity, hardness ratio, and the equivalent width of Fe $K$ line emission. Of the 14 sources where we detect the periodicity for the first time, four are likely to be intermediate polars, five are likely to be polars, two are neutron star X-ray binaries, and three are of unknown nature., Comment: 19 pages, 9 figures, accepted for publication in A&A

Electric vehicles (EVs) play a crucial role in the transition towards sustainable modes of transportation and thus are critical to the energy transition. As their number grows, managing the aggregate power of EV charging is crucial to maintain grid stability and mitigate congestion. This study analyses more than 500 thousand real charging transactions in the Netherlands to explore the challenge and opportunity for the energy system presented by EV growth and smart charging flexibility. Specifically, it analyses the collective ability to provide congestion management services according to the specifications of those services in the Netherlands. In this study, a data-driven model of charging behaviour is created to explore the implications of delivering dependable congestion management services at various aggregation levels and types of service. The probability of offering specific grid services by different categories of charging stations (CS) is analysed. These probabilities can help EV aggregators, such as charging point operators, make informed decisions about offering congestion mitigation products per relevant regulations and distribution system operators to assess their potential. The ability to offer different flexibility products, namely re-dispatch and capacity limitation, for congestion management, is assessed using various dispatch strategies. Next, machine learning models are used to predict the probability of CSs being able to deliver these products, accounting for uncertainties. Results indicate that residential charging locations have significant potential to provide both products during evening peak hours. While shared EVs offer better certainty regarding arrival and departure times, their small fleet size currently restricts their ability to meet the minimum order size of flexible products., Comment: 18 Pages,13 figures, 2 tables

Aggregate peak Electric Vehicle (EV) charging demand is a matter of growing concern for network operators as it severely limits the network's capacity, preventing its reliable operation. Various tariff schemes have been proposed to limit peak demand by incentivizing flexible asset users to shift their demand from peak periods. However, fewer studies quantify the effect of these tariff schemes on the aggregate level. In this paper, we compare the effect of a multi-level segmented network tariff with and without dynamic energy prices for individual EV users on the aggregate peak demand. Results based on real charging transactions from over 1200 public charging points in the Netherlands show that the segmented network tariff with flat energy prices results in more diverse load profiles with increasing aggregation, as compared to cost-optimized dispatch based on only dynamic day-ahead energy prices. When paired with dynamic energy prices, the segmented tariff still outperforms only dynamic energy price-based tariffs in reducing peaks. Results show that a balance between power thresholds and price per threshold is crucial in designing a suitable tariff, taking into account the needs of the power network. We also provide valuable insights to network operators by calculating the diversity factor for various peak demands per charging point., Comment: 6 pages, 4 figures, 2 columns

The Antiproton Decelerator (AD) at CERN provides antiproton bunches with a kinetic energy of 5.3 MeV. The Extra-Low ENergy Antiproton ring at CERN, commissioned at the AD in 2018, now supplies a bunch of electron-cooled antiprotons at a fixed energy of 100 keV. The MUSASHI antiproton trap was upgraded by replacing the radio-frequency quadrupole decelerator with a pulsed drift tube to re-accelerate antiprotons and optimize the injection energy into the degrader foils. By increasing the beam energy to 119 keV, a cooled antiproton accumulation efficiency of (26 +- 6)% was achieved., Comment: 22 pages, 8 figures

We do the error analysis in reliability measures due to the assumption of independence amongst the component lifetimes. In reliability theory, we come across different n-component structures like series, parallel, and k-out-of-n systems. A n component series system works only if all the n components work. While studying the reliability measures of a n-component series system, we mostly assume that all the components have independent lifetimes. Such an assumption eases mathematical complexity while analyzing the data and hence is very common. But in reality, the lifetimes of the components are very much interdependent. Such an assumption of independence hence leads to inaccurate analysis of data. In multiple situations like studying a complex system with many components, we turn to assuming independence keeping some room for error. However, if we have some knowledge of the behaviour of errors or some estimate on the error bound, we could decide if we assume independence and prefer mathematical simplicity (if we know the error is within our allowed limit), or keep the mathematical complexity and get accurate results without assuming independence. We aim to find the relative errors in the reliability measures for a n-component series system.

Understanding the modus operandi of adversaries aids organizations in employing efficient defensive strategies and sharing intelligence in the community. This knowledge is often present in unstructured natural language text within threat analysis reports. A translation tool is needed to interpret the modus operandi explained in the sentences of the threat report and translate it into a structured format. This research introduces a methodology named TTPXHunter for the automated extraction of threat intelligence in terms of Tactics, Techniques, and Procedures (TTPs) from finished cyber threat reports. It leverages cyber domain-specific state-of-the-art natural language processing (NLP) to augment sentences for minority class TTPs and refine pinpointing the TTPs in threat analysis reports significantly. The knowledge of threat intelligence in terms of TTPs is essential for comprehensively understanding cyber threats and enhancing detection and mitigation strategies. We create two datasets: an augmented sentence-TTP dataset of 39,296 samples and a 149 real-world cyber threat intelligence report-to-TTP dataset. Further, we evaluate TTPXHunter on the augmented sentence dataset and the cyber threat reports. The TTPXHunter achieves the highest performance of 92.42% f1-score on the augmented dataset, and it also outperforms existing state-of-the-art solutions in TTP extraction by achieving an f1-score of 97.09% when evaluated over the report dataset. TTPXHunter significantly improves cybersecurity threat intelligence by offering quick, actionable insights into attacker behaviors. This advancement automates threat intelligence analysis, providing a crucial tool for cybersecurity professionals fighting cyber threats., Comment: Under Review

Activation Patching is a method of directly computing causal attributions of behavior to model components. However, applying it exhaustively requires a sweep with cost scaling linearly in the number of model components, which can be prohibitively expensive for SoTA Large Language Models (LLMs). We investigate Attribution Patching (AtP), a fast gradient-based approximation to Activation Patching and find two classes of failure modes of AtP which lead to significant false negatives. We propose a variant of AtP called AtP*, with two changes to address these failure modes while retaining scalability. We present the first systematic study of AtP and alternative methods for faster activation patching and show that AtP significantly outperforms all other investigated methods, with AtP* providing further significant improvement. Finally, we provide a method to bound the probability of remaining false negatives of AtP* estimates.

Prior interpretability research studying narrow distributions has preliminarily identified self-repair, a phenomena where if components in large language models are ablated, later components will change their behavior to compensate. Our work builds off this past literature, demonstrating that self-repair exists on a variety of models families and sizes when ablating individual attention heads on the full training distribution. We further show that on the full training distribution self-repair is imperfect, as the original direct effect of the head is not fully restored, and noisy, since the degree of self-repair varies significantly across different prompts (sometimes overcorrecting beyond the original effect). We highlight two different mechanisms that contribute to self-repair, including changes in the final LayerNorm scaling factor and sparse sets of neurons implementing Anti-Erasure. We additionally discuss the implications of these results for interpretability practitioners and close with a more speculative discussion on the mystery of why self-repair occurs in these models at all, highlighting evidence for the Iterative Inference hypothesis in language models, a framework that predicts self-repair., Comment: ICML 2024

The polytropic index of electrons in a magnetized plasma is experimentally investigated in the presence of external heating and anisotropic work done on the system by the magnetic field, incorporating the effective dimensionality in an anisotropic plasma. The measurements clearly demonstrate localized regions of superadiabatic electrons due to particle acceleration and energization through non-resonant betatron and Fermi heating processes. In regions where wave-induced heating through electron cyclotron resonance dominates, the plasma electrons behave adiabatically. The realization of superadiabaticity is universal where heating dominates the work done and has broader implications for energy exchange processes in astrophysically-relevant plasmas., Comment: 7 pages, 4 figures

How do transformer-based large language models (LLMs) store and retrieve knowledge? We focus on the most basic form of this task -- factual recall, where the model is tasked with explicitly surfacing stored facts in prompts of form `Fact: The Colosseum is in the country of'. We find that the mechanistic story behind factual recall is more complex than previously thought. It comprises several distinct, independent, and qualitatively different mechanisms that additively combine, constructively interfering on the correct attribute. We term this generic phenomena the additive motif: models compute through summing up multiple independent contributions. Each mechanism's contribution may be insufficient alone, but summing results in constructive interfere on the correct answer. In addition, we extend the method of direct logit attribution to attribute an attention head's output to individual source tokens. We use this technique to unpack what we call `mixed heads' -- which are themselves a pair of two separate additive updates from different source tokens., Comment: NeurIPS 2023 Attributing Model Behaviour at Scale Workshop

We study the phenomenology of a vector dark matter (VDM) in a $U(1)_X$ gauged extension of the Standard Model (SM) which is connected to the type II seesaw framework via the Higgs portal. When this $U(1)_X$ symmetry is spontaneously broken by the vacuum expectation value (VEV) of a complex scalar singlet, the gauge boson $Z^\prime$ becomes massive. The stability of the dark matter (DM) is ensured by the introduction of an exact charge conjugation symmetry. On the other hand, the $SU(2)_L$ triplet scalar generates light neutrino masses through the type II seesaw mechanism. We have studied the phenomenology of the usual WIMP DM considering all possible theoretical and experimental constraints that are applicable. Due to the presence of triplet scalar, our scenario can accommodate the observed $2\sigma$ deviation in $h \to Z \gamma $ decay. We have also briefly discussed the possibility of non-thermal production of DM from the decay of the same complex scalar that is responsible for the breaking of this $U(1)_X$ symmetry., Comment: 33 pages, 11 figures, 1 table, matches the version accepted for publication in PRD