Search

Language models (LMs) hallucinate. We inquire: Can we detect and mitigate hallucinations before they happen? This work answers this research question in the positive, by showing that the internal representations of LMs provide rich signals that can be used for this purpose. We introduce FactCheckMate, which preemptively detects hallucinations by learning a classifier that predicts whether the LM will hallucinate, based on the model's hidden states produced over the inputs, before decoding begins. If a hallucination is detected, FactCheckMate then intervenes, by adjusting the LM's hidden states such that the model will produce more factual outputs. FactCheckMate provides fresh insights that the inner workings of LMs can be revealed by their hidden states. Practically, both the detection and mitigation models in FactCheckMate are lightweight, adding little inference overhead; FactCheckMate proves a more efficient approach for mitigating hallucinations compared to many post-hoc alternatives. We evaluate FactCheckMate over LMs of different scales and model families (including Llama, Mistral, and Gemma), across a variety of QA datasets from different domains. Our results demonstrate the effectiveness of leveraging internal representations for early hallucination detection and mitigation, achieving over 70% preemptive detection accuracy. On average, outputs generated by LMs with intervention are 34.4% more factual compared to those without intervention. The average overhead difference in the inference time introduced by FactCheckMate is around 3.16 seconds.

We demonstrate that preference optimization methods can effectively enhance LLM safety. Applying various alignment techniques to the Falcon 11B model using safety datasets, we achieve a significant boost in global safety score (from $57.64\%$ to $99.90\%$) as measured by LlamaGuard 3 8B, competing with state-of-the-art models. On toxicity benchmarks, average scores in adversarial settings dropped from over $0.6$ to less than $0.07$. However, this safety improvement comes at the cost of reduced general capabilities, particularly in math, suggesting a trade-off. We identify noise contrastive alignment (Safe-NCA) as an optimal method for balancing safety and performance. Our study ultimately shows that alignment techniques can be sufficient for building safe and robust models.

Despite the growing importance of Arabic as a global language, there is a notable lack of language models pre-trained exclusively on Arabic data. This shortage has led to limited benchmarks available for assessing language model performance in Arabic. To address this gap, we introduce two novel benchmarks designed to evaluate models' mathematical reasoning and language understanding abilities in Arabic. These benchmarks are derived from a General Aptitude Test (GAT) called Qiyas exam, a standardized test widely used for university admissions in Saudi Arabia. For validation purposes, we assess the performance of ChatGPT-3.5-trubo and ChatGPT-4 on our benchmarks. Our findings reveal that these benchmarks pose a significant challenge, with ChatGPT-4 achieving an overall average accuracy of 64%, while ChatGPT-3.5-trubo achieved an overall accuracy of 49% across the various question types in the Qiyas benchmark. We believe the release of these benchmarks will pave the way for enhancing the mathematical reasoning and language understanding capabilities of future models tailored for the low-resource Arabic language.

Label errors are a common issue in machine learning datasets, particularly for tasks such as Named Entity Recognition. Such label errors might hurt model training, affect evaluation results, and lead to an inaccurate assessment of model performance. In this study, we dived deep into one of the widely adopted Arabic NER benchmark datasets (ANERcorp) and found a significant number of annotation errors, missing labels, and inconsistencies. Therefore, in this study, we conducted empirical research to understand these errors, correct them and propose a cleaner version of the dataset named CLEANANERCorp. CLEANANERCorp will serve the research community as a more accurate and consistent benchmark., Comment: Proceedings of the 6th Workshop on Open-Source Arabic Corpora and Processing Tools (OSACT) with Shared Tasks on Arabic LLMs Hallucination and Dialect to MSA Machine Translation @ LREC-COLING 2024

We investigate the spontaneous disintegration of magnons in two-dimensional ferromagnets and antiferromagnets dominated by long-range dipolar interactions. Analyzing kinematic constraints, we show that the unusual dispersion of dipolar ferromagnets in a uniform magnetic field precludes magnon-decay at all fields, in sharp contrast to short-range exchange-driven magnets. However, in a staggered magnetic field, used for state preparation in recent Rydberg array experiments, magnons can decay in both dipolar ferromagnets and antiferromagnets. Remarkably, such decays do not require a minimum threshold field, and happen over a nearly fixed fraction of the Brillouin Zone in the XY limit, highlighting the significant role played by dipolar interactions. In addition, topological transitions in the decay surfaces lead to singularities in the magnon spectrum. Regularizing such singular behavior via a self-consistent approach, we make predictions for dynamical spin correlations accessible to near-term quantum simulators and sensors., Comment: 5+6 pages, 4+2 figures

Game environments offer a unique opportunity for training virtual agents due to their interactive nature, which provides diverse play traces and affect labels. Despite their potential, no reinforcement learning framework incorporates human affect models as part of their observation space or reward mechanism. To address this, we present the \emph{Affectively Framework}, a set of Open-AI Gym environments that integrate affect as part of the observation space. This paper introduces the framework and its three game environments and provides baseline experiments to validate its effectiveness and potential., Comment: 5 pages, 2 figures, 2 tables

Autonomously tailoring content to a set of predetermined affective patterns has long been considered the holy grail of affect-aware human-computer interaction at large. The experience-driven procedural content generation framework realises this vision by searching for content that elicits a certain experience pattern to a user. In this paper, we propose a novel reinforcement learning (RL) framework for generating affect-tailored content, and we test it in the domain of racing games. Specifically, the experience-driven RL (EDRL) framework is given a target arousal trace, and it then generates a racetrack that elicits the desired affective responses for a particular type of player. EDRL leverages a reward function that assesses the affective pattern of any generated racetrack from a corpus of arousal traces. Our findings suggest that EDRL can accurately generate affect-driven racing game levels according to a designer's style and outperforms search-based methods for personalised content generation. The method is not only directly applicable to game content generation tasks but also employable broadly to any domain that uses content for affective adaptation., Comment: 9 pages, 4 figures, 1 table

Parametric converters are parametric amplifiers that mix two spatially separate nondegenerate modes and are commonly used for amplifying and squeezing microwave signals in quantum computing and sensing. In Josephson parametric converters, the strong localized nonlinearity of the Josephson Junction limits the amplification and squeezing, as well as the dynamic range, in current devices. In contrast, a weak distributed nonlinearity can provide higher gain and dynamic range, when implemented as a kinetic inductance (KI) nanowire of a dirty superconductor, and has additional benefits such as resilience to magnetic field, higher-temperature operation, and simplified fabrication. Here, we propose, demonstrate, and analyze the performance of a KI parametric converter that relies on the weak distributed nonlinearity of a KI nanowire. The device utilizes three-wave mixing induced by a DC current bias. We demonstrate its operation as a nondegenerate parametric amplifier with high phase-sensitive gain, reaching two-mode amplification and deamplification of $\sim$30 dB for two resonances separated by 0.8 GHz, in excellent agreement with our theory of the device. We observe a dynamic range of -108~dBm at 30 dB gain. Our device can significantly broaden applications of quantum-limited signal processing devices including phase-preserving amplification and two-mode squeezing., Comment: 6 pages, 4 figures. Accepted for publication in Physical Review Applied

Data transfer using human-body communication (HBC) represents an actively explored alternative solution to address the challenges related to energy-efficiency, tissue absorption, and security of conventional wireless. Although the use of HBC for wearable-to-wearable communication has been well-explored, different configurations for the transmitter (Tx) and receiver (Rx) for implant-to-wearable HBC needs further studies. This paper substantiates the hypothesis that a fully implanted galvanic Tx is more efficient than a capacitive Tx for interaction with a wearable Rx. Given the practical limitations of implanting an ideal capacitive device, we choose a galvanic device with one electrode encapsulated to model the capacitive scenario. We analyze the lumped circuit model for in-body to out-of-body communication, and perform Circuit-based as well as Finite Element Method (FEM) simulations to explore how the encapsulation thickness affects the received signal levels. We demonstrate in-vivo experimental results on live Sprague Dawley rats to validate the hypothesis, and show that compared to the galvanic Tx, the channel loss will be $\approx$ 20 dB higher with each additional mm thickness of capacitive encapsulation, eventually going below the noise floor for ideal capacitive Tx.

In response to the rapidly escalating costs of computing with large matrices and tensors caused by data movement, several lossy compression methods have been developed to significantly reduce data volumes. Unfortunately, all these methods require the data to be decompressed before further computations are done. In this work, we develop a lossy compressor that allows a dozen fairly fundamental operations directly on compressed data while offering good compression ratios and modest errors. We implement a new compressor PyBlaz based on the familiar GPU-powered PyTorch framework, and evaluate it on three non-trivial applications, choosing different number systems for internal representation. Our results demonstrate that the compressed-domain operations achieve good scalability with problem sizes while incurring errors well within acceptable limits. To our best knowledge, this is the first such lossy compressor that supports compressed-domain operations while achieving acceptable performance as well as error., Comment: An extended but earlier version of paper in https://dl.acm.org/doi/10.1145/3624062.3625122 published at the DRBSD Workshop in 2023

Recent astronomical observations revealed an increasing molecular complexity in the interstellar medium through the detection of a series of large cyclic carbon species. To correctly interpret these detections, a complex analysis is necessary that takes into account the non-local thermodynamic equilibrium (non-LTE) conditions of the emitting media (e.g. when energy level populations deviate from a Boltzman distribution). This requires proper state-to-state collisional data for the excitation and de-excitation processes of the molecular levels. Cyclopentadiene (c-C5H6), which was recently detected in multiple cold interstellar clouds, is extensively studied in many aspects due to its large importance for chemistry in general. At the same time, there are no collisional data available for this species, which are necessary for a more precise interpretation of the corresponding detections. In this work, we first provide an accurate 3D rigid-rotor interaction potential for the [c-C5H6 + He] complex from high-level of ab initio theories, which has been used to study their inelastic collision by the exact close coupling quantum scattering method. To the best of our knowledge, this is the first study where this method is systematically applied to treat the dynamics of molecular collisions involving more than ten atoms. We also analyse the collisional propensity rules and the differences in contrast to calculations, where the approximate coupled states scattering methods is used., Comment: Physical Chemistry Chemical Physics, 2024

We predict that the gapless $U(1)$ Dirac spin liquid naturally emerges in a two-dimensional array of quantum dipoles. In particular, we demonstrate that the dipolar XY model$\unicode{x2014}$realized in both Rydberg atom arrays and ultracold polar molecules$\unicode{x2014}$hosts a quantum spin liquid ground state on the kagome lattice. Large-scale density matrix renormalization group calculations indicate that this spin liquid exhibits signatures of gapless, linearly-dispersing spinons, consistent with the $U(1)$ Dirac spin liquid. We identify a route to adiabatic preparation via staggered on-site fields and demonstrate that this approach can prepare cold spin liquids within experimentally realistic time-scales. Finally, we propose a number of novel signatures of the Dirac spin liquid tailored to near-term quantum simulators, including termination-dependent edge modes and the Friedel response to a local perturbation., Comment: 5+13 pages, 4+10 figures

Program synthesis with language models (LMs) has unlocked a large set of reasoning abilities; code-tuned LMs have proven adept at generating programs that solve a wide variety of algorithmic symbolic manipulation tasks (e.g. word concatenation). However, not all reasoning tasks are easily expressible as code, e.g. tasks involving commonsense reasoning, moral decision-making, and sarcasm understanding. Our goal is to extend an LM's program synthesis skills to such tasks and evaluate the results via pseudo-programs, namely Python programs where some leaf function calls are left undefined. To that end, we propose, Code Generation and Emulated EXecution (CoGEX). CoGEX works by (1) training LMs to generate their own pseudo-programs, (2) teaching them to emulate their generated program's execution, including those leaf functions, allowing the LM's knowledge to fill in the execution gaps; and (3) using them to search over many programs to find an optimal one. To adapt the CoGEX model to a new task, we introduce a method for performing program search to find a single program whose pseudo-execution yields optimal performance when applied to all the instances of a given dataset. We show that our approach yields large improvements compared to standard in-context learning approaches on a battery of tasks, both algorithmic and soft reasoning. This result thus demonstrates that code synthesis can be applied to a much broader class of problems than previously considered. Our released dataset, fine-tuned models, and implementation can be found at \url{https://github.com/nweir127/CoGEX}., Comment: 16 pages, 10 figures

Self-correction has emerged as a promising solution to boost the reasoning performance of large language models (LLMs), where LLMs refine their solutions using self-generated critiques that pinpoint the errors. This work explores whether small (<= 13B) language models (LMs) have the ability of self-correction on reasoning tasks with minimal inputs from stronger LMs. We propose a novel pipeline that prompts smaller LMs to collect self-correction data that supports the training of self-refinement abilities. First, we leverage correct solutions to guide the model in critiquing their incorrect responses. Second, the generated critiques, after filtering, are used for supervised fine-tuning of the self-correcting reasoner through solution refinement. Our experimental results show improved self-correction abilities of two models on five datasets spanning math and commonsense reasoning, with notable performance gains when paired with a strong GPT-4-based verifier, though limitations are identified when using a weak self-verifier for determining when to correct., Comment: ACL Findings 2024 - Camera Ready

The efforts on the development, standardization and improvements to communication systems towards 5G Advanced and 6G are on track to provide benefits such as an unprecedented level of connectivity and performance, enabling a diverse range of vertical services. The full integration of non-terrestrial components into 6G plays a pivotal role in realizing this paradigm shift towards ubiquitous communication and global coverage. However, this integration into 6G brings forth a set of its own challenges, particularly in Radio Access Technologies (RATs). To this end, this paper comprehensively discusses those challenges at different levels of RATs and proposes the corresponding potential emerging advancements in the realm of 6G NTN. In particular, the focus is on advancing the prospective aspects of Radio Resource Management (RRM), spectral coexistence in terrestrial and non-terrestrial components and flexible waveform design solutions to combat the impediments. This discussion with a specific focus on emerging advancements in 6G NTN RATs is critical for shaping the next generation networks and potentially relevant in contributing the part in standardization in forthcoming releases, Comment: accepted in 2024 EuCNC and 6G Summit, Antwerp, Belgium, 3_6 June 2024

Large language models (LLMs) learn a vast amount of knowledge during pretraining, but they are often oblivious to the source(s) of such knowledge. We investigate the problem of intrinsic source citation, where LLMs are required to cite the pretraining source supporting a generated response. Intrinsic source citation can enhance LLM transparency, interpretability, and verifiability. To give LLMs such ability, we explore source-aware training -- a recipe that involves (i) training the LLM to associate unique source document identifiers with the knowledge in each document, followed by (ii) an instruction-tuning stage to teach the LLM to cite a supporting pretraining source when prompted. Source-aware training borrows from existing pretraining/fine-tuning frameworks and requires minimal changes to the model architecture or implementation. Through experiments on synthetic data, we demonstrate that our training recipe can enable faithful attribution to the pretraining data without a substantial impact on the model's perplexity compared to standard pretraining. Our findings also highlight the importance of pretraining data augmentation in achieving attribution. Code and data available here: \url{https://github.com/mukhal/intrinsic-source-citation}, Comment: COLM '24

Speech Emotion Recognition (SER) is crucial for enabling computers to understand the emotions conveyed in human communication. With recent advancements in Deep Learning (DL), the performance of SER models has significantly improved. However, designing an optimal DL architecture requires specialised knowledge and experimental assessments. Fortunately, Neural Architecture Search (NAS) provides a potential solution for automatically determining the best DL model. The Differentiable Architecture Search (DARTS) is a particularly efficient method for discovering optimal models. This study presents emoDARTS, a DARTS-optimised joint CNN and Sequential Neural Network (SeqNN: LSTM, RNN) architecture that enhances SER performance. The literature supports the selection of CNN and LSTM coupling to improve performance. While DARTS has previously been used to choose CNN and LSTM operations independently, our technique adds a novel mechanism for selecting CNN and SeqNN operations in conjunction using DARTS. Unlike earlier work, we do not impose limits on the layer order of the CNN. Instead, we let DARTS choose the best layer order inside the DARTS cell. We demonstrate that emoDARTS outperforms conventionally designed CNN-LSTM models and surpasses the best-reported SER results achieved through DARTS on CNN-LSTM by evaluating our approach on the IEMOCAP, MSP-IMPROV, and MSP-Podcast datasets., Comment: Submitted to IEEE Transactions on Affective Computing on February 19, 2024. arXiv admin note: text overlap with arXiv:2305.14402

The high cost of model training makes it increasingly desirable to develop techniques for unlearning. These techniques seek to remove the influence of a training example without having to retrain the model from scratch. Intuitively, once a model has unlearned, an adversary that interacts with the model should no longer be able to tell whether the unlearned example was included in the model's training set or not. In the privacy literature, this is known as membership inference. In this work, we discuss adaptations of Membership Inference Attacks (MIAs) to the setting of unlearning (leading to their "U-MIA" counterparts). We propose a categorization of existing U-MIAs into "population U-MIAs", where the same attacker is instantiated for all examples, and "per-example U-MIAs", where a dedicated attacker is instantiated for each example. We show that the latter category, wherein the attacker tailors its membership prediction to each example under attack, is significantly stronger. Indeed, our results show that the commonly used U-MIAs in the unlearning literature overestimate the privacy protection afforded by existing unlearning techniques on both vision and language models. Our investigation reveals a large variance in the vulnerability of different examples to per-example U-MIAs. In fact, several unlearning algorithms lead to a reduced vulnerability for some, but not all, examples that we wish to unlearn, at the expense of increasing it for other examples. Notably, we find that the privacy protection for the remaining training examples may worsen as a consequence of unlearning. We also discuss the fundamental difficulty of equally protecting all examples using existing unlearning schemes, due to the different rates at which examples are unlearned. We demonstrate that naive attempts at tailoring unlearning stopping criteria to different examples fail to alleviate these issues.

A detailed interpretation of the detected emission lines of environments in which propyne (or methyl acetylene, CH$_3$CCH) is observed requires a knowledge of its collisional rate coefficients with the most abundant species in the interstellar medium, He or H$_2$. We present the first three-dimensional potential energy surface (3D-PES) for the CH$_3$CCH-He molecular complex, study the dynamics of the collision, and report the first set of rate coefficients for temperatures up to 100 K for the collisional excitation of the lowest 60 ortho rotational levels and 60 para rotational levels of CH$_3$CCH by He atoms. We computed the 3D-PES with the explicitly correlated coupled-cluster with single-, double-, and perturbative triple-excitation method, in conjunction with the augmented correlation-consistent triple zeta basis set (CCSD(T)-F12a/aug-cc-pVTZ). The 3D-PES was fitted to an analytical function. Scattering computations of pure rotational (de-)excitation of CH$_3$CCH by collision with He atoms were performed and the state-to-state cross sections were computed using the close coupling method for total energies up to 100 cm$^{-1}$ and with the coupled states approximation at higher energy for both ortho and para symmetries of CH$_3$CCH. The PES obtained is caracterized by a large anisotropy and a potential well depth of 51.04 cm$^{-1}$. By thermally averaging the collisional cross sections, we determined quenching rate coefficients for kinetic temperatures up to 100 K. A strong even $\Delta j$ propensity rule at almost all collision energies exists for CH$_3$CCH-He complex. To evaluate the impact of rate coefficients in the analysis of observations, we carried out non-LTE radiative transfer computations of the excitation temperatures and we demonstrate that LTE conditions are typically not fulfilled for the propyne molecule., Comment: Accepted in A&A

Modern Standard Arabic (MSA) nominals present many morphological and lexical modeling challenges that have not been consistently addressed previously. This paper attempts to define the space of such challenges, and leverage a recently proposed morphological framework to build a comprehensive and extensible model for MSA nominals. Our model design addresses the nominals' intricate morphotactics, as well as their paradigmatic irregularities. Our implementation showcases enhanced accuracy and consistency compared to a commonly used MSA morphological analyzer and generator. We make our models publicly available., Comment: Findings of the Association for Computational Linguistics: EACL 2024

Microwave quantum information networks require reliable transmission of single photon propagating modes over lossy channels. In this article we propose a microwave noise-less linear amplifier (NLA) suitable to circumvent the losses incurred by a flying photon undergoing an amplitude damping channel (ADC). The proposed model is constructed by engineering a simple one-dimensional four node cluster state. Contrary to conventional NLAs based on quantum scissors (QS), single photon amplification is realized without the need for photon number resolving detectors (PNRDs). Entanglement between nodes comprising the device's cluster is achieved by means of a controlled phase gate (CPHASE). Furthermore, photon measurements are implemented by quantum non demolition detectors (QNDs), which are currently available as a part of circuit quantum electrodynamics (cQED) toolbox. We analyze the performance of our device practically by considering detection inefficiency and dark count probability. We further examine the potential usage of our device in low power quantum sensing applications and remote secret key generation (SKG). Specifically, we demonstrate the device's ability to prepare loss-free resources offline, and its capacity to overcome the repeater-less bound of SKG. We compare the performance of our device against a QS-NLA for the aforementioned applications, and highlight explicitly the operating conditions under which our device can outperform a QS-NLA. The proposed device is also suitable for applications in the optical domain.

We develop techniques for determining an explicit Berry-Esseen bound in the Kolmogorov distance for the normal approximation of a ratio of Gaussian functionals. We provide an upper bound in terms of the third and fourth cumulants, using some novel techniques and sharp estimates for cumulants. As applications, we study the rate of convergence of the distribution of discretized versions of minimum contrast and maximum likelihood estimators of the drift parameter of the Ornstein-Uhlenbeck process. Moreover, we derive upper bounds that are strictly sharper than those available in the literature., Comment: arXiv admin note: text overlap with arXiv:2211.11566