Search

There is a growing interest in training domain-expert LLMs that excel in specific technical fields compared to their general-purpose instruction-tuned counterparts. However, these expert models often experience a loss in their safety abilities in the process, making them capable of generating harmful content. As a solution, we introduce an efficient and effective merging-based alignment method called \textsc{MergeAlign} that interpolates the domain and alignment vectors, creating safer domain-specific models while preserving their utility. We apply \textsc{MergeAlign} on Llama3 variants that are experts in medicine and finance, obtaining substantial alignment improvements with minimal to no degradation on domain-specific benchmarks. We study the impact of model merging through model similarity metrics and contributions of individual models being merged. We hope our findings open new research avenues and inspire more efficient development of safe expert LLMs.

In recent years, data mining technologies have been well applied to many domains, including e-commerce. In customer relationship management (CRM), the RFM analysis model is one of the most effective approaches to increase the profits of major enterprises. However, with the rapid development of e-commerce, the diversity and abundance of e-commerce data pose a challenge to mining efficiency. Moreover, in actual market transactions, the chronological order of transactions reflects customer behavior and preferences. To address these challenges, we develop an effective algorithm called SeqRFM, which combines sequential pattern mining with RFM models. SeqRFM considers each customer's recency (R), frequency (F), and monetary (M) scores to represent the significance of the customer and identifies sequences with high recency, high frequency, and high monetary value. A series of experiments demonstrate the superiority and effectiveness of the SeqRFM algorithm compared to the most advanced RFM algorithms based on sequential pattern mining. The source code and datasets are available at GitHub https://github.com/DSI-Lab1/SeqRFM., Comment: Preprint. 5 figures, 5 tables

We report on a systematic study of the field dependence of the Hall resistivity $\rho _{yx} (B)$ as a function of doping in thin films of electron-doped superconducting cuprate Pr$_{2-x}$Ce$_{x}$CuO$_{4 \pm \delta}$ (PCCO). Across the studied doping range from $x = 0.125$ to $x = 0.20$, we observe a nonlinear dependence of $\rho_{yx}$ with $B$. The leading $B^{3}$ nonlinear term is negative, increases with decreasing temperature and peaks around optimal doping ($x = 0.15$). The observed nonlinear contribution is consistent with the presence of two different types of free carriers (electron-like and hole-like) even for doping with only an apparent hole Fermi surface as observed by angle-resolved photoemission spectroscopy. Based on an analysis using the two-carrier model, this negative nonlinear contribution to $\rho _{yx} (B)$ implies that the density of the charge carriers behaving like electrons is larger than the density of those behaving like holes for all the doping values explored. Combined with the Hall coefficient reaching $R_{H} = 0$ at specific temperatures for selected doping levels, we also conclude that the mobility of hole-like carriers is larger than that of electron-like carriers for doping around $x^{*} \approx 0.165$., Comment: 9 pages, 7 figures. Submitted to Physical Review B

At the hospital, the dispersion of information regarding anti-cancer treatment makes it difficult to extract. We proposed a solution capable of identifying dates, drugs and their temporal relationship within free-text oncology reports with very few manual annotations. We used pattern recognition for dates, dictionaries for drugs and transformer language models for the relationship, combined with a data augmentation strategy. Our models achieved good prediction F1-scores, reaching 0.872. The performance of models with data augmentation outperforms those of models without. By inferring such models, we can now identify and structure thousands of previously unavailable treatment events to better apprehend solutions and patient response., Comment: Medical Informatics Europe 2025 - Intelligent health systems -- From technology to data and knowledge, European Federation for Medical Informatics, May 2025, Glasgow, United Kingdom

The Hall coefficient $R_\mathrm{H}$ as a function of temperature for Pr$_{2-x}$Ce$_{x}$CuO$_{4}$ (PCCO) at $x = 0.17$ shows two sign changes between 0 and 300K that can be explained qualitatively using the two-carrier model. Using this two-carrier model, one predicts the presence of a nonlinear $B^{3}$ contribution to the Hall resistivity $\rho_{yx}$ that should be easily observed when the linear coefficient $R_{H}$ is approaching zero. We present the measurement of this nonlinear Hall resistivity for overdoped PCCO thin films at $x=0.18$. We show that this weak nonlinear term persists at all temperatures despite being dominated by the linear term ($R_{H}B$). It is strongly temperature dependent and, analyzing it with the two-carrier model, we conclude that the density of electron carriers is larger than the hole density, but that the hole mobility is larger than that of the electrons. This nonlinear Hall resistivity should be observable for similar materials for which $R_{H}(T)$ goes through sign reversals as does overdoped PCCO., Comment: Accepted for publication in Phys. Rev. B (Oct. 2024), 6 pages, 5 figures

Cyanobacteria are the most frequent dominant species of algal blooms in inland waters, threatening ecosystem function and water quality, especially when toxin-producing strains predominate. Enhanced by anthropogenic activities and global warming, cyanobacterial blooms are expected to increase in frequency and global distribution. Early warning systems (EWS) for cyanobacterial blooms development allow timely implementation of management measures, reducing the risks associated to these blooms. In this paper, we propose an effective EWS for cyanobacterial bloom forecasting, which uses 6 years of incomplete high-frequency spatio-temporal data from multiparametric probes, including phycocyanin (PC) fluorescence as a proxy for cyanobacteria. A probe agnostic and replicable method is proposed to pre-process the data and to generate time series specific for cyanobacterial bloom forecasting. Using these pre-processed data, six different non-site/species-specific predictive models were compared including the autoregressive and multivariate versions of Linear Regression, Random Forest, and Long-Term Short-Term (LSTM) neural networks. Results were analyzed for seven forecasting time horizons ranging from 4 to 28 days evaluated with a hybrid system that combined regression metrics (MSE, R2, MAPE) for PC values, classification metrics (Accuracy, F1, Kappa) for a proposed alarm level of 10 ug PC/L, and a forecasting-specific metric to measure prediction improvement over the displaced signal (skill). The multivariate version of LSTM showed the best and most consistent results across all forecasting horizons and metrics, achieving accuracies of up to 90% in predicting the proposed PC alarm level. Additionally, positive skill values indicated its outstanding effectiveness to forecast cyanobacterial blooms from 16 to 28 days in advance.

With sizable volatile envelopes but smaller radii than the solar system ice giants, sub-Neptunes have been revealed as one of the most common types of planet in the galaxy. While the spectroscopic characterization of larger sub-Neptunes (2.5-4R$_\oplus$) has revealed hydrogen-dominated atmospheres, smaller sub-Neptunes (1.6--2.5R$_\oplus$) could either host thin, rapidly evaporating hydrogen-rich atmospheres or be stable metal-rich "water worlds" with high mean molecular weight atmospheres and a fundamentally different formation and evolutionary history. Here, we present the 0.6--2.8$\mu$m JWST NIRISS/SOSS transmission spectrum of GJ 9827 d, the smallest (1.98 R$_\oplus$) warm (T$_\mathrm{eq, A_B=0.3} \sim 620$K) sub-Neptune where atmospheric absorbers have been detected to date. Our two transit observations with NIRISS/SOSS, combined with the existing HST/WFC3 spectrum, enable us to break the clouds-metallicity degeneracy. We detect water in a highly metal-enriched "steam world" atmosphere (O/H of $\sim 4$ by mass and H$_2$O found to be the background gas with a volume mixing ratio of >31%). We further show that these results are robust to stellar contamination through the transit light source effect. We do not detect escaping metastable He, which, combined with previous nondetections of escaping He and H, supports the steam atmosphere scenario. In water-rich atmospheres, hydrogen loss driven by water photolysis happens predominantly in the ionized form which eludes observational constraints. We also detect several flares in the NIRISS/SOSS light-curves with far-UV energies of the order of 10$^{30}$ erg, highlighting the active nature of the star. Further atmospheric characterization of GJ 9827 d probing carbon or sulfur species could reveal the origin of its high metal enrichment., Comment: 37 pages, 18 figures, Accepted for publication in ApJL

The effectiveness of Recommender Systems (RS) is closely tied to the quality and distinctiveness of user profiles, yet despite many advancements in raw performance, the sensitivity of RS to user profile quality remains under-researched. This paper introduces novel information-theoretic measures for understanding recommender systems: a "surprise" measure quantifying users' deviations from popular choices, and a "conditional surprise" measure capturing user interaction coherence. We evaluate 7 recommendation algorithms across 9 datasets, revealing the relationships between our measures and standard performance metrics. Using a rigorous statistical framework, our analysis quantifies how much user profile density and information measures impact algorithm performance across domains. By segmenting users based on these measures, we achieve improved performance with reduced data and show that simpler algorithms can match complex ones for low-coherence users. Additionally, we employ our measures to analyze how well different recommendation algorithms maintain the coherence and diversity of user preferences in their predictions, providing insights into algorithm behavior. This work advances the theoretical understanding of user behavior and practical heuristics for personalized recommendation systems, promoting more efficient and adaptive architectures.

Fully connected Graph Transformers (GT) have rapidly become prominent in the static graph community as an alternative to Message-Passing models, which suffer from a lack of expressivity, oversquashing, and under-reaching. However, in a dynamic context, by interconnecting all nodes at multiple snapshots with self-attention, GT loose both structural and temporal information. In this work, we introduce Supra-LAplacian encoding for spatio-temporal TransformErs (SLATE), a new spatio-temporal encoding to leverage the GT architecture while keeping spatio-temporal information. Specifically, we transform Discrete Time Dynamic Graphs into multi-layer graphs and take advantage of the spectral properties of their associated supra-Laplacian matrix. Our second contribution explicitly model nodes' pairwise relationships with a cross-attention mechanism, providing an accurate edge representation for dynamic link prediction. SLATE outperforms numerous state-of-the-art methods based on Message-Passing Graph Neural Networks combined with recurrent models (e.g LSTM), and Dynamic Graph Transformers, on 9 datasets. Code is available at: github.com/ykrmm/SLATE.

This paper evaluates and compares the performance of model-free and model-based reinforcement learning for the attitude control of fixed-wing unmanned aerial vehicles using PID as a reference point. The comparison focuses on their ability to handle varying flight dynamics and wind disturbances in a simulated environment. Our results show that the Temporal Difference Model Predictive Control agent outperforms both the PID controller and other model-free reinforcement learning methods in terms of tracking accuracy and robustness over different reference difficulties, particularly in nonlinear flight regimes. Furthermore, we introduce actuation fluctuation as a key metric to assess energy efficiency and actuator wear, and we test two different approaches from the literature: action variation penalty and conditioning for action policy smoothness. We also evaluate all control methods when subject to stochastic turbulence and gusts separately, so as to measure their effects on tracking performance, observe their limitations and outline their implications on the Markov decision process formalism., Comment: Published at ICINCO 2024

We compute the cobordism group $\Omega^{\operatorname{lag}}(M)$ of Lagrangian immersions into a symplectic manifold $(M, \omega)$ in terms of a stable homotopy group of a Thom spectrum constructed from $M$. This generalizes a result of Eliashberg in the case of exact symplectic manifolds. As applications, we compute $\Omega^{\operatorname{lag}}(M)$ when $M$ is a closed surface and give a partial description of $\Omega^{\operatorname{lag}}(M)$ when $M$ is a monotone manifold., Comment: 28 pages

The Sun's global inertial modes are very sensitive to the solar differential rotation and to properties of the deep solar convection zone which are currently poorly constrained. These properties include the superadiabatic temperature gradient, the latitudinal entropy gradient, and the turbulent viscosity. The inertial modes also play a key role in controlling the Sun's large-scale structure and dynamics, in particular the solar differential rotation. This paper summarizes recent observations and advances in the (linear and nonlinear) modeling of the solar inertial modes., Comment: Submitted on 05-Mar-2024 for publication in procs. IAU Symp 365, Eds. A. Getling and L. Kitchatinov

The clustering of categorical data is a common and important task in computer science, offering profound implications across a spectrum of applications. Unlike purely numerical data, categorical data often lack inherent ordering as in nominal data, or have varying levels of order as in ordinal data, thus requiring specialized methodologies for efficient organization and analysis. This review provides a comprehensive synthesis of categorical data clustering in the past twenty-five years, starting from the introduction of K-modes. It elucidates the pivotal role of categorical data clustering in diverse fields such as health sciences, natural sciences, social sciences, education, engineering and economics. Practical comparisons are conducted for algorithms having public implementations, highlighting distinguishing clustering methodologies and revealing the performance of recent algorithms on several benchmark categorical datasets. Finally, challenges and opportunities in the field are discussed.

Large language models (LLMs) play a vital role in almost every domain in today's organizations. In the context of this work, we highlight the use of LLMs for sentiment analysis (SA) and explainability. Specifically, we contribute a novel technique to leverage LLMs as a post-hoc model-independent tool for the explainability of SA. We applied our technique in the financial domain for currency-pair price predictions using open news feed data merged with market prices. Our application shows that the developed technique is not only a viable alternative to using conventional eXplainable AI but can also be fed back to enrich the input to the machine learning (ML) model to better predict future currency-pair values. We envision our results could be generalized to employing explainability as a conventional enrichment for ML input for better ML predictions in general., Comment: 7 pages, 3 figures, AIFin@ECAI 2024

We give a criterion for separability of subgroups of certain outer automorphism groups. This answers questions of Hagen and Sisto, by strengthening and generalizing a result of theirs on mapping class groups., Comment: 5 pages

Dynamic link prediction is a critical task in the analysis of evolving networks, with applications ranging from recommender systems to economic exchanges. However, the concept of the temporal receptive field, which refers to the temporal context that models use for making predictions, has been largely overlooked and insufficiently analyzed in existing research. In this study, we present a comprehensive analysis of the temporal receptive field in dynamic graph learning. By examining multiple datasets and models, we formalize the role of temporal receptive field and highlight their crucial influence on predictive accuracy. Our results demonstrate that appropriately chosen temporal receptive field can significantly enhance model performance, while for some models, overly large windows may introduce noise and reduce accuracy. We conduct extensive benchmarking to validate our findings, ensuring that all experiments are fully reproducible. Code is available at https://github.com/ykrmm/BenchmarkTW .

The increasing scale of Transformer models has led to an increase in their pre-training computational requirements. While quantization has proven to be effective after pre-training and during fine-tuning, applying quantization in Transformers during pre-training has remained largely unexplored at scale for language modeling. This study aims to explore the impact of quantization for efficient pre-training of Transformers, with a focus on linear layer components. By systematically applying straightforward linear quantization to weights, activations, gradients, and optimizer states, we assess its effects on model efficiency, stability, and performance during training. By offering a comprehensive recipe of effective quantization strategies to be applied during the pre-training of Transformers, we promote high training efficiency from scratch while retaining language modeling ability. Code is available at https://github.com/chandar-lab/EfficientLLMs.

The point-curvature model for membrane protein inclusions is shown to capture multibody interactions very well. Using this model, we find that the interplay between membrane tension and multibody interactions results in a collective attraction of oppositely curved inclusions tending to form antiferromagnetic structures with a square lattice. This attraction can produce a phase separation between curved and non-curved proteins, resulting in the clustering of curved proteins. We also show that the many-body repulsion between identical curved proteins is enhanced by membrane tension. This can lead to the dissolution of clusters stabilized by short-range forces when the tension is increased. These new phenomena are biologically relevant and could be investigated experimentally, Comment: 6 pages, 7 figures

Realizing tunable functional materials with built-in nanoscale heat flow directionality represents a significant challenge that could advance thermal management strategies. Here we use spatiotemporally-resolved thermoreflectance to visualize lateral thermal transport anisotropy in self-assembled supercrystals of anisotropic Au nanocrystals. Correlative electron and thermoreflectance microscopy reveal that nano- to meso-scale heat predominantly flows along the long-axis of the anisotropic nanocrystals, and does so across grain boundaries and curved assemblies while voids disrupt heat flow. We finely control the anisotropy via the aspect ratio of constituent nanorods, and it exceeds the aspect ratio for nano-bipyramid supercrystals and certain nanorod arrangements. Finite element simulations and effective medium modeling rationalize the emergent anisotropic behavior in terms of a simple series resistance model, further providing a framework for estimating thermal anisotropy as a function of material and structural parameters. Self-assembly of colloidal nanocrystals promises an interesting route to direct heat flow in a wide range of applications that utilize this important class of materials.

The nonequilibrium dynamics of a cycling three-state Potts model is studied on a square lattice using Monte Carlo simulations and continuum theory. This model is relevant to chemical reactions on a catalytic surface and to molecular transport across a membrane. Several characteristic modes are formed depending on the flipping energies between successive states and the contact energies between neighboring sites. Under cyclic symmetry conditions, cycling homogeneous phases and spiral waves form at low and high flipping energies, respectively. In the intermediate flipping energy regime, these two modes coexist temporally in small systems and/or at low contact energies. Under asymmetric conditions, we observed small biphasic domains exhibiting amoeba-like locomotion and temporal coexistence of spiral waves and a dominant non-cyclic one-state phase. An increase in the flipping energy between two successive states, say state 0 and state 1, while keeping the other flipping energies constant, induces the formation of the third phase (state 2), owing to the suppression of the nucleation of state 0 domains. Under asymmetric conditions regarding the contact energies, two different modes can appear depending on the initial state, due to a hysteresis phenomenon., Comment: 14 pages, 15 figures

Face recognition and verification are two computer vision tasks whose performance has progressed with the introduction of deep representations. However, ethical, legal, and technical challenges due to the sensitive character of face data and biases in real training datasets hinder their development. Generative AI addresses privacy by creating fictitious identities, but fairness problems persist. We promote fairness by introducing a demographic attributes balancing mechanism in generated training datasets. We experiment with an existing real dataset, three generated training datasets, and the balanced versions of a diffusion-based dataset. We propose a comprehensive evaluation that considers accuracy and fairness equally and includes a rigorous regression-based statistical analysis of attributes. The analysis shows that balancing reduces demographic unfairness. Also, a performance gap persists despite generation becoming more accurate with time. The proposed balancing method and comprehensive verification evaluation promote fairer and transparent face recognition and verification.

LHS 1140 b is the second-closest temperate transiting planet to the Earth with an equilibrium temperature low enough to support surface liquid water. At 1.730$\pm$0.025 R$_\oplus$, LHS 1140 b falls within the radius valley separating H$_2$-rich mini-Neptunes from rocky super-Earths. Recent mass and radius revisions indicate a bulk density significantly lower than expected for an Earth-like rocky interior, suggesting that LHS 1140 b could either be a mini-Neptune with a small envelope of hydrogen ($\sim$0.1% by mass) or a water world (9--19% water by mass). Atmospheric characterization through transmission spectroscopy can readily discern between these two scenarios. Here, we present two JWST/NIRISS transit observations of LHS 1140 b, one of which captures a serendipitous transit of LHS 1140 c. The combined transmission spectrum of LHS 1140 b shows a telltale spectral signature of unocculted faculae (5.8 $\sigma$), covering $\sim$20% of the visible stellar surface. Besides faculae, our spectral retrieval analysis reveals tentative evidence of residual spectral features, best-fit by Rayleigh scattering from an N$_2$-dominated atmosphere (2.3 $\sigma$), irrespective of the consideration of atmospheric hazes. We also show through Global Climate Models (GCM) that H$_2$-rich atmospheres of various compositions (100$\times$, 300$\times$, 1000$\times$solar metallicity) are ruled out to $>$10 $\sigma$. The GCM calculations predict that water clouds form below the transit photosphere, limiting their impact on transmission data. Our observations suggest that LHS 1140 b is either airless or, more likely, surrounded by an atmosphere with a high mean molecular weight. Our tentative evidence of an N$_2$-rich atmosphere provides strong motivation for future transmission spectroscopy observations of LHS 1140 b., Comment: 26 pages, 14 figures, 3 tables, accepted for publication in ApJL

We study the poset of hyperbolic structures on Thompson's group $F$ and its generalizations $F_n$ for $n \geq 2$. The global structure of this poset is as simple as one would expect, with the maximal non-elementary elements being two quasi-parabolic actions corresponding to well-known ascending HNN-extension expressions of $F_n$. However, the local structure turns out to be incredibly rich, in stark contrast with the situation for the $T$ and $V$ counterparts. We show that the subposet of quasi-parabolic hyperbolic structures consists of two isomorphic posets, each of which contains uncountably many subposets of \emph{lamplike} structures, which can be described combinatorially in terms of certain hyperbolic structures on related lamplighter groups. Moreover, each of these subposets, as well as intersections and complements thereof, is very large, in that it contains a copy of the power set of the natural numbers. We also prove that these uncountably many uncountable subposets are not the entire picture, indeed there exists a copy of the power set of the natural numbers consisting entirely of non-lamplike structures. We also prove that this entire vast array of hyperbolic structures on $F_n$ collapses as soon as one takes a natural semidirect product with $\mathbb{Z}/2\mathbb{Z}$. These results are all proved via a detailed analysis of confining subsets, and along the way we establish a number of fundamental results in the theory of confining subsets of groups., Comment: 60 pages, 4 figures

Multimodal Deep Learning enhances decision-making by integrating diverse information sources, such as texts, images, audio, and videos. To develop trustworthy multimodal approaches, it is essential to understand how uncertainty impacts these models. We propose LUMA, a unique benchmark dataset, featuring audio, image, and textual data from 50 classes, for learning from uncertain and multimodal data. It extends the well-known CIFAR 10/100 dataset with audio samples extracted from three audio corpora, and text data generated using the Gemma-7B Large Language Model (LLM). The LUMA dataset enables the controlled injection of varying types and degrees of uncertainty to achieve and tailor specific experiments and benchmarking initiatives. LUMA is also available as a Python package including the functions for generating multiple variants of the dataset with controlling the diversity of the data, the amount of noise for each modality, and adding out-of-distribution samples. A baseline pre-trained model is also provided alongside three uncertainty quantification methods: Monte-Carlo Dropout, Deep Ensemble, and Reliable Conflictive Multi-View Learning. This comprehensive dataset and its benchmarking tools are intended to promote and support the development, evaluation, and benchmarking of trustworthy and robust multimodal deep learning approaches. We anticipate that the LUMA dataset will help the ICLR community to design more trustworthy and robust machine learning approaches for safety critical applications.

Large Language Models (LLMs) are increasingly used for boosting organizational efficiency and automating tasks. While not originally designed for complex cognitive processes, recent efforts have further extended to employ LLMs in activities such as reasoning, planning, and decision-making. In business processes, such abilities could be invaluable for leveraging on the massive corpora LLMs have been trained on for gaining deep understanding of such processes. In this work, we plant the seeds for the development of a benchmark to assess the ability of LLMs to reason about causal and process perspectives of business operations. We refer to this view as Causally-augmented Business Processes (BP^C). The core of the benchmark comprises a set of BP^C related situations, a set of questions about these situations, and a set of deductive rules employed to systematically resolve the ground truth answers to these questions. Also with the power of LLMs, the seed is then instantiated into a larger-scale set of domain-specific situations and questions. Reasoning on BP^C is of crucial importance for process interventions and process improvement. Our benchmark, accessible at https://huggingface.co/datasets/ibm/BPC, can be used in one of two possible modalities: testing the performance of any target LLM and training an LLM to advance its capability to reason about BP^C., Comment: 12 pages, 1 figure

Filaments of cold gas ($T\leq 10^{4}$ K) are found in the inner regions of many cool-core clusters. These structures are thought to play a major role in the regulation of feedback from active galactic nuclei (AGN). We study the morphology of the filaments, their formation, and their impact on the propagation of the outflowing AGN jets. We present a set of GPU-accelerated 3D (magneto)hydrodynamical simulations of an idealized Perseus-like cluster using the performance portable code AthenaPK. We include radiative cooling, and a self-regulated AGN feedback model that redistributes accreted material through kinetic, thermal and magnetic feedback. We confirm that magnetic fields play an important role in both the formation and evolution of the cold material. These suppress the formation of massive cold disks and favor magnetically supported filaments over clumpy structures. Achieving resolutions of $25-50$ pc, we find that filaments are not monolithic as they contain numerous and complex magnetically--supported substructures. We find that the mass distribution of these clumps follows a power law for all investigated filaments, consistent with previous cloud-crushing simulations of individual clumps. Studying the evolution of individual filaments, we find that their formation pathways can be diverse. We find examples of filaments forming through a combination of cold gas uplifting and condensation, as well as systems of purely infalling clumps condensing out of the intracluster medium. The density contrast between the cold gas and the outflowing hot material leads to recurring deflections of the jets, favoring inflation of bubbles., Comment: 19 pages, 19 figures, submitted to Astronomy & Astrophysics

Large language models are first pre-trained on trillions of tokens and then instruction-tuned or aligned to specific preferences. While pre-training remains out of reach for most researchers due to the compute required, fine-tuning has become affordable thanks to parameter-efficient methods such as LoRA and QLoRA. Alignment is known to be sensitive to the many factors involved, including the quantity and quality of data, the alignment method, and the adapter rank. However, there has not yet been an extensive study of their effect on downstream performance. To address this gap, we conduct an in-depth investigation of the impact of popular choices for three crucial axes: (i) the alignment dataset (HH-RLHF and BeaverTails), (ii) the alignment technique (SFT and DPO), and (iii) the model (LLaMA-1, Vicuna-v1.3, Mistral-7b, and Mistral-7b-Instruct). Our extensive setup spanning over 300 experiments reveals consistent trends and unexpected findings. We observe how more informative data helps with preference alignment, cases where supervised fine-tuning outperforms preference optimization, and how aligning to a distinct preference boosts performance on downstream tasks. Through our in-depth analyses, we put forward key guidelines to help researchers perform more effective parameter-efficient LLM alignment., Comment: Accepted to ACL (Main) 2024

Reversible architectures have been shown to be capable of performing on par with their non-reversible architectures, being applied in deep learning for memory savings and generative modeling. In this work, we show how reversible architectures can solve challenges in parallelizing deep model training. We introduce PETRA, a novel alternative to backpropagation for parallelizing gradient computations. PETRA facilitates effective model parallelism by enabling stages (i.e., a set of layers) to compute independently on different devices, while only needing to communicate activations and gradients between each other. By decoupling the forward and backward passes and keeping a single updated version of the parameters, the need for weight stashing is also removed. We develop a custom autograd-like training framework for PETRA, and we demonstrate its effectiveness on CIFAR-10, ImageNet32, and ImageNet, achieving competitive accuracies comparable to backpropagation using ResNet-18, ResNet-34, and ResNet-50 models.

Training Large Language Models (LLMs) relies heavily on distributed implementations, employing multiple GPUs to compute stochastic gradients on model replicas in parallel. However, synchronizing gradients in data parallel settings induces a communication overhead increasing with the number of distributed workers, which can impede the efficiency gains of parallelization. To address this challenge, optimization algorithms reducing inter-worker communication have emerged, such as local optimization methods used in Federated Learning. While effective in minimizing communication overhead, these methods incur significant memory costs, hindering scalability: in addition to extra momentum variables, if communications are only allowed between multiple local optimization steps, then the optimizer's states cannot be sharded among workers. In response, we propose $\textbf{AC}$cumulate while $\textbf{CO}$mmunicate ($\texttt{ACCO}$), a memory-efficient optimization algorithm tailored for distributed training of LLMs. $\texttt{ACCO}$ allows to shard optimizer states across workers, overlaps gradient computations and communications to conceal communication costs, and accommodates heterogeneous hardware. Our method relies on a novel technique to mitigate the one-step delay inherent in parallel execution of gradient computations and communications, eliminating the need for warmup steps and aligning with the training dynamics of standard distributed optimization while converging faster in terms of wall-clock time. We demonstrate the effectiveness of $\texttt{ACCO}$ on several LLMs training and fine-tuning tasks.

We prove that the groups of orientation-preserving homeomorphisms and diffeomorphisms of $\mathbb{R}^n$ are boundedly acyclic, in all regularities. This is the first full computation of the bounded cohomology of a transformation group that is not compactly supported, and it implies that many characteristic classes of flat $\mathbb{R}^n$- and $S^n$-bundles are unbounded. We obtain the same result for the group of homeomorphisms of the disc that restrict to the identity on the boundary, and for the homeomorphism group of the non-compact Cantor set. In the appendix, Alexander Kupers proves a controlled version of the annulus theorem which we use to study the bounded cohomology of the homeomorphism group of the discs., Comment: 43 pages. Main paper by the first three authors, appendix by the fourth author. v2: minor changes

Hydro-Quebec (HQ) is a vertically integrated utility that produces, transmits, and distributes most of the electricity in the province of Quebec. The power grid it operates has a particular architecture created by large hydroelectric dams located far north and the extensive 735kV transmission grid that allows the generated power to reach the majority of the load located thousands of kilometers away in the southern region of Quebec. The specificity of the grid has led HQ to develop monitoring tools responsible for generating so-called stability limits. Those stability limits take into account several nonlinear phenomena such as angular stability, frequency stability, or voltage stability. Since generation planning and operation tools rely mostly on mixed integer linear programming formulation, HQ had to adapt its tools to integrate stability limits into them. This paper presents the challenges it faced, especially considering its reserve monitoring tool and unit commitment tool., Comment: 12 pages, 7 figures, 7 tables, 1 algorithm

Creating defects in graphene by hydrogenation, either to achieve hydrogen chemisorption or partial etching, is a way to open an electronic band gap in graphene. Understanding the range of stability conditions of partially etched or hydrogenated graphene is crucial for application, as processing conditions (e.g. temperature) and quality control (characterization) conditions may result in modifying the material through partial or full dehydrogenation, and subsequent alteration of its electronic properties. This work reports a study of various dehydrogenation conditions of hydrogenated or hydrogen-etched monolayer graphene (1LG), either free-standing or supported by an interferential (SiO2/Si) substrate, using incremental annealing under nitrogen atmosphere up to 400 {\textdegree}C. Materials were investigated by Raman spectroscopy. Indeed, it has been known since 2012 that the intensity ratio of two Raman bands activated by double resonance, D over D' (ID/ID') can be used to identify the type of defects in defective graphene. It is shown that hydrogenated 1LG, characterized by a large ID/ID' ratio (~9-15), is stable provided annealing remains below 300 {\textdegree}C. On the other hand, defective 1LG resulting from hydrogen etching remains stable up to 400 {\textdegree}C, whether the 1LG is hydrogenated on one side or both sides, while a modification in the type and proportions of defects is likely. Experimental conditions for the safe use of Raman spectroscopy, otherwise able to induce specimen overheating because of the laser energy and power, are also determined and discussed.

The 5.0 industry promotes collaborative robots (cobots). This research studies the impacts of cobot collaboration using an experimental setup. 120 participants realized a simple and a complex assembly task. 50% collaborated with another human (H/H) and 50% with a cobot (H/C). The workload and the acceptability of the cobotic collaboration were measured. Working with a cobot decreases the effect of the task complexity on the human workload and on the output quality. However, it increases the time completion and the number of gestures (while decreasing their frequency). The H/C couples have a higher chance of success but they take more time and more gestures to realize the task. The results of this research could help developers and stakeholders to understand the impacts of implementing a cobot in production chains.

The presence of kiloparsec-sized bubble structures in both sides of the Galactic plan suggests active phases of Sgr A$^\star$, the central supermassive black hole of the Milky-Way in the last 1-6 Myr. The contribution of such event on the cosmic-ray flux measured in the solar neighborhood is investigated with numerical simulations. We evaluate whether the population of high-energy charged particles emitted by the Galactic Center could be sufficient to significantly impact the CR flux measured in the solar neighborhood. We present a set of 3D magnetohydrodynamical simulations, following the anisotropic propagation of CR in a Milky - Way like Galaxy. Independent populations of cosmic-ray are followed through time, originating from two different sources types, namely Supernovae and the Galactic Center. To assess the evolution of the CR flux spectrum properties, we split these populations into two independent energy groups of 100 GeV and 10 TeV. We find that the anisotropic nature of cosmic-ray diffusion dramatically affects the amount of cosmic-ray energy received in the solar neighborhood. Typical timescale to observe measurable changes in the CR spectrum slope is of the order 10 Myr, largely surpassing estimated ages of the Fermi bubbles in the AGN jet-driven scenario. We conclude that a cosmic-ray outburst from the Galactic center in the last few Myr is unlikely to produce any observable feature in the local CR spectrum in the TeV regime within times consistent with current estimates of the Fermi bubbles ages., Comment: 9 pages, 10 figures, Accepted for publication in Astronomy & Astrophysics

We explore the bound neutrons decay into invisible particles (e.g., $n\rightarrow 3 \nu$ or $nn \rightarrow 2 \nu$) in the JUNO liquid scintillator detector. The invisible decay includes two decay modes: $ n \rightarrow { inv} $ and $ nn \rightarrow { inv} $. The invisible decays of $s$-shell neutrons in $^{12}{\rm C}$ will leave a highly excited residual nucleus. Subsequently, some de-excitation modes of the excited residual nuclei can produce a time- and space-correlated triple coincidence signal in the JUNO detector. Based on a full Monte Carlo simulation informed with the latest available data, we estimate all backgrounds, including inverse beta decay events of the reactor antineutrino $\bar{\nu}_e$, natural radioactivity, cosmogenic isotopes and neutral current interactions of atmospheric neutrinos. Pulse shape discrimination and multivariate analysis techniques are employed to further suppress backgrounds. With two years of exposure, JUNO is expected to give an order of magnitude improvement compared to the current best limits. After 10 years of data taking, the JUNO expected sensitivities at a 90% confidence level are $\tau/B( n \rightarrow { inv} ) > 5.0 \times 10^{31} \, {\rm yr}$ and $\tau/B( nn \rightarrow { inv} ) > 1.4 \times 10^{32} \, {\rm yr}$., Comment: 28 pages, 7 figures, 4 tables

The performance of deep neural networks is enhanced by ensemble methods, which average the output of several models. However, this comes at an increased cost at inference. Weight averaging methods aim at balancing the generalization of ensembling and the inference speed of a single model by averaging the parameters of an ensemble of models. Yet, naive averaging results in poor performance as models converge to different loss basins, and aligning the models to improve the performance of the average is challenging. Alternatively, inspired by distributed training, methods like DART and PAPA have been proposed to train several models in parallel such that they will end up in the same basin, resulting in good averaging accuracy. However, these methods either compromise ensembling accuracy or demand significant communication between models during training. In this paper, we introduce WASH, a novel distributed method for training model ensembles for weight averaging that achieves state-of-the-art image classification accuracy. WASH maintains models within the same basin by randomly shuffling a small percentage of weights during training, resulting in diverse models and lower communication costs compared to standard parameter averaging methods.

The optimal model for a given task is often challenging to determine, requiring training multiple models from scratch which becomes prohibitive as dataset and model sizes grow. A more efficient alternative is to reuse smaller pre-trained models by expanding them, however, this is not widely adopted as how this impacts training dynamics remains poorly understood. While prior works have introduced statistics to measure these effects, they remain flawed. To rectify this, we offer a new approach for understanding and quantifying the impact of expansion through the lens of the loss landscape, which has been shown to contain a manifold of linearly connected minima. Building on this new perspective, we propose a metric to study the impact of expansion by estimating the size of the manifold. Experimental results show a clear relationship between gains in performance and manifold size, enabling the comparison of candidate models and presenting a first step towards expanding models more reliably based on geometric properties of the loss landscape.

The thermal conductivities of crystals and glasses vary strongly and with opposite trends upon heating, decreasing in crystals and increasing in glasses. Here, we show--both with first-principles predictions based on the Wigner transport equation and with thermoreflectance experiments--that the dominant transport mechanisms of crystals (particle-like propagation) and glasses (wave-like tunnelling) can coexist and compensate in materials with crystalline bond order and nearly glassy bond geometry. We demonstrate that ideal compensation emerges in a sample of silica in the form of tridymite, carved from a meteorite found in Steinbach (Germany) in 1724, and yields a "propagation-tunneling-invariant" (PTI) conductivity that is independent from temperature and intermediate between the opposite trends of $\alpha$-quartz crystal and silica glass. We show how such PTI conductivity occurs in the quantum regime below the Debye temperature, and can largely persist at high temperatures in a geometrically amorphous tridymite phase found in refractory bricks fired for years in furnaces for steel smelting. Last, we discuss implications to heat transfer in solids exposed to extreme temperature variations, ranging from planetary cooling to heating protocols to reduce the carbon footprint of industrial furnaces., Comment: 5+11 pages, 3+6 figures

Generative modeling aims at producing new datapoints whose statistical properties resemble the ones in a training dataset. In recent years, there has been a burst of machine learning techniques and settings that can achieve this goal with remarkable performances. In most of these settings, one uses the training dataset in conjunction with noise, which is added as a source of statistical variability and is essential for the generative task. Here, we explore the idea of using internal chaotic dynamics in high-dimensional chaotic systems as a way to generate new datapoints from a training dataset. We show that simple learning rules can achieve this goal within a set of vanilla architectures and characterize the quality of the generated datapoints through standard accuracy measures.

We present THERADIA WoZ, an ecological corpus designed for audiovisual research on affect in healthcare. Two groups of senior individuals, consisting of 52 healthy participants and 9 individuals with Mild Cognitive Impairment (MCI), performed Computerised Cognitive Training (CCT) exercises while receiving support from a virtual assistant, tele-operated by a human in the role of a Wizard-of-Oz (WoZ). The audiovisual expressions produced by the participants were fully transcribed, and partially annotated based on dimensions derived from recent models of the appraisal theories, including novelty, intrinsic pleasantness, goal conduciveness, and coping. Additionally, the annotations included 23 affective labels drew from the literature of achievement affects. We present the protocols used for the data collection, transcription, and annotation, along with a detailed analysis of the annotated dimensions and labels. Baseline methods and results for their automatic prediction are also presented. The corpus aims to serve as a valuable resource for researchers in affective computing, and is made available to both industry and academia.

The effects of sintering temperature (Ts) and chemical substitution on the structural and magnetic properties of manganite compounds La1-xSrxMn1-yFeyO3 (0.025 <= x <= 0.7; y= 0.01,0.15) are explored in a search to optimize their magnetocaloric properties around room temperature. A ferromagnetic (FM) to paramagnetic (PM) phase transition is observed at a Curie temperature Tc that can be controlled to approach room temperature by Sr and Fe substitution, but also by adjusting the sintering temperature Ts. Accordingly, the magnetic entropy change (-DSM) quantifying the magnetocaloric effect (MCE) presents a peak at or close to Tc that shifts and broadens with both Sr and Fe doping and is further tuned with sintering temperature. Altogether, we show that it is possible to adjust the strength and dominance of the ferromagnetic coupling in these ceramics, but also using disorder as a tool to broaden and adjust the temperature range with significant magnetic entropy change., Comment: 61 pages, 14 figures, 3 tables. Supp. mat. : 3 figures, 1 table. Submitted to J. Magn. Magn. Mater

This work is a numerical study of a transitional shock wave boundary layer interaction (SWBLI). The main goal is to improve our understanding of the well known low-frequency SWBLI unsteadiness and especially the suspected role of triadic interactions in the underlying physical mechanism. To this end, a Direct Numerical Simulation (DNS) is performed using high-order finite volume scheme equipped with a suitable shock capture procedure. The resulting database is then extensively post-processed in order to extract the main dynamical features of the interaction zone (involved characteristic frequencies, characteristics of the vortical structures, etc.). The dynamical organisation and space-time evolution of the flow at dominant frequencies are then further characterised by mean of a Spectral Proper Orthogonal Decomposition (SPOD) analysis. In order to study the role of triadic interactions occurring in the interaction region, a BiSpectral Mode Decomposition (BSMD) analysis is applied to the data base. It allows us to extract the significant triadic interactions, their location and the resulting physical spatial modes. Strong triadic interactions are detected in the downstream part of the separation bubble whose role on the low-frequency unsteadiness is characterised. All the results of the various analyses are then discussed and integrated to formulate a possible mechanism fuelling low-frequency SWBLI unsteadiness., Comment: Under consideration for publication in Journal of Fluid Mechanics

Graphene Hall effect magnetic field sensors hold great promise for the development of ultra-sensitive magnetometers. Their performance is frequently analysed using the two-channel model where electron and hole conductivities are simply added. Unfortunately, this model is unable to capture all the features of the sensor, particularly the bias current dependence of the magnetic field sensitivity. Here we present an advanced model that provides an in-depth understanding of how graphene Hall sensors operate, and demonstrate its ability to quantitatively assess their performance. First, we report the fabrication of sensors with different qualities of graphene, with the best devices achieving magnetic field sensitivities as high as 5000 ohms/T, outperforming the best silicon and narrow-gap semiconductor-based sensors. Then, we examine their performance in detail using the proposed numerical model, which combines Boltzmann formalism, with distinct Fermi levels for electrons and holes, and a new method for the introduction of substrate-induced electron-hole puddles. Importantly, the dependences of magnetic field sensitivity on bias current, disorder, substrate and Hall bar geometry are quantitatively reproduced for the first time. In addition, the model emphasizes that the performance of devices with widths of the order of the charge carrier diffusion length, is significantly affected by the bias current due to the occurrence of large and non-symmetric carrier accumulation and depletion areas near the edges of the Hall bar. The formation of these areas induces a transverse diffusion particle flux capable of counterbalancing the particle flux induced by the Lorentz force when the Hall electric field cancels out in the ambipolar regime. Finally, we discuss how sensor performance can be enhanced by Fermi velocity engineering, paving the way for future ultra-sensitive graphene Hall effect sensors., Comment: 29 pages including supplemental material, 12 figures in the main text and 7 figures in the supplemental material

Training large deep learning models requires parallelization techniques to scale. In existing methods such as Data Parallelism or ZeRO-DP, micro-batches of data are processed in parallel, which creates two drawbacks: the total memory required to store the model's activations peaks at the end of the forward pass, and gradients must be simultaneously averaged at the end of the backpropagation step. We propose Cyclic Data Parallelism, a novel paradigm shifting the execution of the micro-batches from simultaneous to sequential, with a uniform delay. At the cost of a slight gradient delay, the total memory taken by activations is constant, and the gradient communications are balanced during the training step. With Model Parallelism, our technique reduces the number of GPUs needed, by sharing GPUs across micro-batches. Within the ZeRO-DP framework, our technique allows communication of the model states with point-to-point operations rather than a collective broadcast operation. We illustrate the strength of our approach on the CIFAR-10 and ImageNet datasets.

In order to characterize the global circulation of the subsurface ocean of Jovian and Saturnian moons, we analyze the properties of 21 three-dimensional simulations of Boussinesq thermal convection in a rapidly rotating spherical shell. Flow is driven by an adverse temperature contrast imposed across the domain, and is subjected to no-slip boundary conditions. We cover a region of parameter space previously unexplored by global simulations, both in terms of rapid rotation and vigor of convective forcing, closer to, yet still admittedly far from, the conditions appropriate for the subsurface ocean of Ganymede, Europa, Enceladus, and Titan. Our most extreme simulations exhibit a dynamic global circulation that combines powerful east-west zonal jets, planetary waves, and vortices. A spectral analysis of the kinetic energy distribution performed in cylindrical geometry reveals a high degree of anisotropy of the simulated flows. Specifically, the axisymmetric zonal energy spectra follow a steep $-5$ slope in wavenumber space, with the energy amplitude exclusively controlled by the rotation rate. In contrast, the non-axisymmetric residual spectra display a gentle $-5/3$ slope, with the energy amplitude controlled by the thermal buoyancy input power. This spectral behavior conforms with the theory of zonostrophic turbulence and allows us to propose tentative extrapolations of these findings to the more extreme conditions of icy satellites. By assuming that kinetic energy dissipates via Ekman friction we predict an upper bound for the zonal velocity ranging from a few centimeters per second for Enceladus to about one meter per second for Ganymede, with residual velocities smaller than the zonal velocity by an order of magnitude on each moon. These predictions yield typical jets size approaching the ocean depth of Titan, Ganymede and Europa and $10$ to $40\%$ of the ocean depth on Enceladus., Comment: 24 pages, 12 figures, 3 tables, accepted for publication in Icarus

The recent discovery of inertial waves on the surface of the Sun offers new possibilities to learn about the solar interior. These waves are long-lived with a period on the order of the Sun rotation period ($\sim$27 days) and are sensitive to parameters deep inside the Sun. They are excited by turbulent convection, leading to a passive imaging problem. In this work, we present the forward and inverse problem of reconstructing viscosity and differential rotation on the Sun from cross-covariance observations of these inertial waves., Comment: proceedings paper