Search

The development of fast and strong microactuators that can be integrated in microdevices is an essential challenge due to a lack of appropriate driving principles. In this paper, a membrane actuator powered by internal combustion of hydrogen and oxygen in a chamber with a volume of 3.1 nanoliters is demonstrated. The combustion in such a small volume is possible only for an extremely high surface-to-volume (S/V) ratio on the order of 10^7 1/m. The fuel with this S/V is prepared electrochemically in a special regime that produces only nanobubbles. A cloud of nanobubbles merges, forming a microbubble, which explodes, increasing the volume 500 times in 10us. The actuator generates an instantaneous force up to 0.5N and is able to move a body 11,000 times more massive than itself. The natural response time of about 10ms is defined by the incubation time needed to produce an exploding bubble. The device demonstrates reliable cyclic actuation at a frequency of 1Hz restricted by the effect of electrolyte aging. After 40,000 explosions, no significant wear in the chamber is observed. Due to record-breaking acceleration and standard microfabrication techniques, the actuator can be used as a universal engine for various microdevices including micro-electro-mechanical systems, microfluidics, microrobotics, wearable and implantable devices., Comment: 28 pages,6 figures, 1 Video (ancillary), supplemental: 1 text file, 1 audio, 2 video

Model-based evaluation is at the heart of successful model development -- as a reward model for training, and as a replacement for human evaluation. To train such evaluators, the standard approach is to collect a large amount of human preference judgments over model responses, which is costly and the data becomes stale as models improve. In this work, we present an approach that aims to im-prove evaluators without human annotations, using synthetic training data only. Starting from unlabeled instructions, our iterative self-improvement scheme generates contrasting model outputs and trains an LLM-as-a-Judge to produce reasoning traces and final judgments, repeating this training at each new iteration using the improved predictions. Without any labeled preference data, our Self-Taught Evaluator can improve a strong LLM (Llama3-70B-Instruct) from 75.4 to 88.3 (88.7 with majority vote) on RewardBench. This outperforms commonly used LLM judges such as GPT-4 and matches the performance of the top-performing reward models trained with labeled examples.

Advanced AI assistants combine frontier LLMs and tool access to autonomously perform complex tasks on behalf of users. While the helpfulness of such assistants can increase dramatically with access to user information including emails and documents, this raises privacy concerns about assistants sharing inappropriate information with third parties without user supervision. To steer information-sharing assistants to behave in accordance with privacy expectations, we propose to operationalize contextual integrity (CI), a framework that equates privacy with the appropriate flow of information in a given context. In particular, we design and evaluate a number of strategies to steer assistants' information-sharing actions to be CI compliant. Our evaluation is based on a novel form filling benchmark composed of human annotations of common webform applications, and it reveals that prompting frontier LLMs to perform CI-based reasoning yields strong results.

This paper explores new methods for tracking learning growth using curriculum-embedded assessments. The key objective of the study was to examine new ways to capture students' performance data to help make ecologically-valid inferences about what learners know and can do long before benchmark or end-of-year assessments. Student learning outcomes are examined through the use of embedded assessments and learning growth mapping. A study of 450,934 students finds that continuous engagement with essential skills on an educational technology platform, BrainPOP, is strongly associated with statistically significant increases in learning gains in Literacy/English Language Arts, Math and Science. Educators might consider using embedded assessments to gain access to and provide early feedback for students. By identifying gaps in students' knowledge, teachers can offer timely support to their most vulnerable students.

Distributing long-distance entanglement is a fundamental goal that is necessary for a variety of tasks such as quantum communication, distributed quantum computing, and quantum metrology. Currently quantum repeater schemes typically aim to distribute one ebit at a time, the equivalent of one Bell pair's worth of entanglement. Here we present a method to distribute a macroscopic amount of entanglement across long-distances using a number of operations that scales only linearly with the chain length. The scheme involves ensembles of qubits and entangling them with an $S^z S^z$ interaction, which can be realized using atomic gas ensembles coupled by a shared optical mode. Using only local measurements on the intermediate ensembles, this leaves the ensembles at the ends of the chain entangled. We show that there are particular ``magic'' interaction times that allow for distribution of entanglement with perfect fidelity, with no degradation with chain length. The scheme is deterministic, such that with suitable local conditional unitary corrections, the same entangled state can always be prepared with good approximation.

Currently, large language models are gaining popularity, their achievements are used in many areas, ranging from text translation to generating answers to queries. However, the main problem with these new machine learning algorithms is that training such models requires large computing resources that only large IT companies have. To avoid this problem, a number of methods (LoRA, quantization) have been proposed so that existing models can be effectively fine-tuned for specific tasks. In this paper, we propose an E2E (end to end) audio summarization model using these techniques. In addition, this paper examines the effectiveness of these approaches to the problem under consideration and draws conclusions about the applicability of these methods., Comment: 36 pages, Master's thesis, 14 figures

A classic result of Brooks, Smith, Stone and Tutte \cite{BSST40} associates to any finite planar network with distinguished source and sink vertices, a tiling of a rectangle by smaller subrectangles whose aspect ratios are given by the conductances of corresponding edges in the network. This tiling can be viewed as a discrete analogue of the uniformizing conformal map that maps a simply connected domain with four distinguished prime ends to a rectangle, so that the four prime ends are mapped to the four corners of the rectangle. We make this intuition precise by showing that if $\Omega$ is a simply connected domain with four distinguished prime ends $A,B,C,D$ in counterclockwise order and $(\Omega_{n})_{n\geq{1}}$ is a sequence of orthodiagonal maps with distinguished boundary vertices $A_{n}, B_{n}, C_{n}, D_{n}$ in counterclockwise order, that are finer and finer approximations of $\Omega$ with its distinguished boundary points $A,B,C,D$, then the corresponding "rectangle tiling maps" converge uniformly on compacts to the aforementioned conformal map on $\Omega$.

In recent work, Harman and Snowden introduced a notion of measure on a Fra\"iss\'e class $\mathfrak{F}$, and showed how such measures lead to interesting tensor categories. Constructing and classifying measures is a difficult problem, and so far only a handful of cases have been worked out. In this paper, we obtain some of the first general results on measures. Our main theorem states that if $\mathfrak{F}$ is distal (in the sense of Simon), and there are some bounds on automorphism groups, then $\mathfrak{F}$ admits only finitely many measures; moreover, we give an effective upper bound on their number. For example, if $\mathfrak{F}$ is the class of ``$s$-dimensional permutations'' (finite sets equipped with $s$ total orders), we show that the number of measures is bounded above by approximately $\exp(\exp(s^2 \log{s}))$., Comment: 23 pages

We consider the remote state restoring and perfect transfer of the zero-order coherence matrix (PTZ) in a spin system governed by the XXZ-Hamiltonian conserving the excitation number. The restoring tool is represented by several nonzero Larmor frequencies in the Hamiltonian. To simplify the analysis we use two approximating models including either step-wise or pulse-type time-dependence of the Larmor frequencies. Restoring in spin chains with up to 20 nodes is studied. Studying PTZ, we consider the zigzag and rectangular configurations and optimize the transfer of the 0-order coherence matrix using geometrical parameters of the communication line as well as the special unitary transformation of the extended receiver. Overall observation is that XXZ-chains require longer time for state transfer than XX-chains, which is confirmed by the analytical study of the evolution under the nearest-neighbor approximation. We demonstrate the exponential increase of the state-transfer time with the spin chain length., Comment: 22 pages, 13 figures. arXiv admin note: text overlap with arXiv:2401.09569

Over the last decade, there has been increasing interest in autonomous driving systems. Reinforcement Learning (RL) shows great promise for training autonomous driving controllers, being able to directly optimize a combination of criteria such as efficiency comfort, and stability. However, RL- based controllers typically offer no safety guarantees, making their readiness for real deployment questionable. In this paper, we propose SECRM-2D (the Safe, Efficient and Comfortable RL- based driving Model with Lane-Changing), an RL autonomous driving controller (both longitudinal and lateral) that balances optimization of efficiency and comfort and follows a fixed route, while being subject to hard analytic safety constraints. The aforementioned safety constraints are derived from the criterion that the follower vehicle must have sufficient headway to be able to avoid a crash if the leader vehicle brakes suddenly. We evaluate SECRM-2D against several learning and non-learning baselines in simulated test scenarios, including freeway driving, exiting, merging, and emergency braking. Our results confirm that representative previously-published RL AV controllers may crash in both training and testing, even if they are optimizing a safety objective. By contrast, our controller SECRM-2D is successful in avoiding crashes during both training and testing, improves over the baselines in measures of efficiency and comfort, and is more faithful in following the prescribed route. In addition, we achieve a good theoretical understanding of the longitudinal steady-state of a collection of SECRM-2D vehicles.

Place recognition is a challenging task in computer vision, crucial for enabling autonomous vehicles and robots to navigate previously visited environments. While significant progress has been made in learnable multimodal methods that combine onboard camera images and LiDAR point clouds, the full potential of these methods remains largely unexplored in localization applications. In this paper, we study the impact of leveraging a multi-camera setup and integrating diverse data sources for multimodal place recognition, incorporating explicit visual semantics and text descriptions. Our proposed method named MSSPlace utilizes images from multiple cameras, LiDAR point clouds, semantic segmentation masks, and text annotations to generate comprehensive place descriptors. We employ a late fusion approach to integrate these modalities, providing a unified representation. Through extensive experiments on the Oxford RobotCar and NCLT datasets, we systematically analyze the impact of each data source on the overall quality of place descriptors. Our experiments demonstrate that combining data from multiple sensors significantly improves place recognition model performance compared to single modality approaches and leads to state-of-the-art quality. We also show that separate usage of visual or textual semantics (which are more compact representations of sensory data) can achieve promising results in place recognition. The code for our method is publicly available: https://github.com/alexmelekhin/MSSPlace, Comment: This work has been submitted to the IEEE for possible publication

What do we want from machine intelligence? We envision machines that are not just tools for thought, but partners in thought: reasonable, insightful, knowledgeable, reliable, and trustworthy systems that think with us. Current artificial intelligence (AI) systems satisfy some of these criteria, some of the time. In this Perspective, we show how the science of collaborative cognition can be put to work to engineer systems that really can be called ``thought partners,'' systems built to meet our expectations and complement our limitations. We lay out several modes of collaborative thought in which humans and AI thought partners can engage and propose desiderata for human-compatible thought partnerships. Drawing on motifs from computational cognitive science, we motivate an alternative scaling path for the design of thought partners and ecosystems around their use through a Bayesian lens, whereby the partners we construct actively build and reason over models of the human and world.

Online marketing faces formidable challenges in managing and interpreting immense volumes of data necessary for competitor analysis, content research, and strategic branding. It is impossible to review hundreds to thousands of transient online content items by hand, and partial analysis often leads to suboptimal outcomes and poorly performing campaigns. We introduce an explainable AI framework SOMONITOR that aims to synergize human intuition with AI-based efficiency, helping marketers across all stages of the marketing funnel, from strategic planning to content creation and campaign execution. SOMONITOR incorporates a CTR prediction and ranking model for advertising content and uses large language models (LLMs) to process high-performing competitor content, identifying core content pillars such as target audiences, customer needs, and product features. These pillars are then organized into broader categories, including communication themes and targeted customer personas. By integrating these insights with data from the brand's own advertising campaigns, SOMONITOR constructs a narrative for addressing new customer personas and simultaneously generates detailed content briefs in the form of user stories that, as shown in the conducted case study, can be directly applied by marketing teams to streamline content production and campaign execution. The adoption of SOMONITOR in daily operations allows digital marketers to quickly parse through extensive datasets, offering actionable insights that significantly enhance campaign effectiveness and overall job satisfaction.

We provide a general framework of estimates for convergence rates of random discrete model curves approaching Schramm Loewner Evolution (SLE) curves in the lattice size scaling limit. We show that a power-law convergence rate of an interface to an SLE curve can be derived from a power-law convergence rate for an appropriate martingale observable provided the discrete curve satisfies a specific bound on crossing events, the Kempannien-Smirnov condition, along with an estimate on the growth of the derivative of the SLE curve. We apply our framework to show that the exploration process for critical site percolation on hexagonal lattice converges to the SLE$_6$ curve with a power-law convergence rate., Comment: Corrected a misprint

We report on the wafer scale fabrication of single mode low-loss lithium niobate on insulator waveguides utilizing a chemically amplified resist and an optimized dry etching method. The fabricated single mode waveguides are free of residuals and re-deposition, with measured losses for straight waveguides around 2 dB/m (0.02 dB/cm). We present on a method offering advantages for large-scale production due to its cost-effectiveness, faster writing time, and simplified processes. This work holds promise for advancing integrated photonics and optical communication technologies.

An efficient implementation of the Toffoli gate is of conceptual importance for running various quantum algorithms, including Grover's search and Shor's integer factorization. However, direct realizations of the Toffoli gate require either a prohibitive growth of the number of two-qubit gates or using ancilla qubits, whereas both of these resources are limited in the current generation of noisy intermediate-scale quantum devices. Here we experimentally demonstrate a scalable improvement of the realization of the Toffoli gate using $^{171}$Yb$^{+}$ trapped-ion-based dual-type optic-microwave qutrits ($d=3$) for the cases of three-, four-qubit and five-qubit versions of the Toffoli gate. With the use of the Molmer-Sorensen gate as a basic two-particle operation, we compare the standard qubit decomposition with the qutrit approach, where upper levels are used as ancillas. The presented decomposition requires only global control of the ancilla levels, which simplifies experimental implementation of the proposed approach. We present an estimation of the scalable improvement of our approach in the case of multi-qubit gates. As we expect, by combining this approach with the leveraging qudits ($d\geq4$) as a set of qubits, our approach may lead to a more efficient realization of various quantum algorithms. With qutrit-based decomposition in Grover's search with three ions, we experimentally demonstrate the 10\% increase in the average algorithm performance., Comment: 8+4 pages, 4+5 figures

We introduce a cyclic quantum battery model, based on an interacting bipartite system, weakly coupled to a thermal bath. The working cycle of the battery consists of four strokes: system thermalization, disconnection of subsystems, ergotropy extraction, and reconnection. The thermal bath acts as a charger in the thermalization stroke, while ergotropy extraction is possible because the ensuing thermal state is no longer passive after the disconnection stroke. Focusing on the case of two interacting qubits, we show that phase coherence, in the presence of non-trivial correlations between the qubits, can be exploited to reach working regimes with efficiency higher than 50% while providing finite ergotropy. Our protocol is illustrated through a simple and feasible circuit model of a cyclic superconducting quantum battery. Furthermore, we simulate the considered cycle on superconducting IBM quantum machines. The good agreement between the theoretical and simulated results strongly suggests that our scheme for cyclic quantum batteries can be successfully realized in superconducting quantum hardware., Comment: 20 pages, 9 figures

Large language models (LLMs) can spend extra compute during inference to generate intermediate thoughts, which helps to produce better final responses. Since Chain-of-Thought (Wei et al., 2022), many such System 2 techniques have been proposed such as Rephrase and Respond (Deng et al., 2023a), System 2 Attention (Weston and Sukhbaatar, 2023) and Branch-Solve-Merge (Saha et al., 2023). In this work we investigate self-supervised methods to ``compile'' (distill) higher quality outputs from System 2 techniques back into LLM generations without intermediate reasoning token sequences, as this reasoning has been distilled into System 1. We show that several such techniques can be successfully distilled, resulting in improved results compared to the original System 1 performance, and with less inference cost than System 2. We posit that such System 2 distillation will be an important feature of future continually learning AI systems, enabling them to focus System 2 capabilities on the reasoning tasks that they cannot yet do well.

Large language models (LLMs) bring unprecedented flexibility in defining and executing complex, creative natural language generation (NLG) tasks. Yet, this flexibility brings new challenges, as it introduces new degrees of freedom in formulating the task inputs and instructions and in evaluating model performance. To facilitate the exploration of creative NLG tasks, we propose a three-component research framework that consists of systematic input manipulation, reference data, and output measurement. We use this framework to explore citation text generation -- a popular scholarly NLP task that lacks consensus on the task definition and evaluation metric and has not yet been tackled within the LLM paradigm. Our results highlight the importance of systematically investigating both task instruction and input configuration when prompting LLMs, and reveal non-trivial relationships between different evaluation metrics used for citation text generation. Additional human generation and human evaluation experiments provide new qualitative insights into the task to guide future research in citation text generation. We make our code and data publicly available., Comment: Accepted to ACL 2024 (Main)

Large language models (LLMs), known for their exceptional reasoning capabilities, generalizability, and fluency across diverse domains, present a promising avenue for enhancing speech-related tasks. In this paper, we focus on integrating decoder-only LLMs to the task of speech-to-text translation (S2TT). We propose a decoder-only architecture that enables the LLM to directly consume the encoded speech representation and generate the text translation. Additionally, we investigate the effects of different parameter-efficient fine-tuning techniques and task formulation. Our model achieves state-of-the-art performance on CoVoST 2 and FLEURS among models trained without proprietary data. We also conduct analyses to validate the design choices of our proposed model and bring insights to the integration of LLMs to S2TT., Comment: Accepted to Interspeech 2024

The roadmap is organized into several thematic sections, outlining current computing challenges, discussing the neuromorphic computing approach, analyzing mature and currently utilized technologies, providing an overview of emerging technologies, addressing material challenges, exploring novel computing concepts, and finally examining the maturity level of emerging technologies while determining the next essential steps for their advancement., Comment: 90 pages, 22 figures, roadmap, neuromorphic

Vulnerability of Frontier language models to misuse and jailbreaks has prompted the development of safety measures like filters and alignment training in an effort to ensure safety through robustness to adversarially crafted prompts. We assert that robustness is fundamentally insufficient for ensuring safety goals, and current defenses and evaluation methods fail to account for risks of dual-intent queries and their composition for malicious goals. To quantify these risks, we introduce a new safety evaluation framework based on impermissible information leakage of model outputs and demonstrate how our proposed question-decomposition attack can extract dangerous knowledge from a censored LLM more effectively than traditional jailbreaking. Underlying our proposed evaluation method is a novel information-theoretic threat model of inferential adversaries, distinguished from security adversaries, such as jailbreaks, in that success is measured by inferring impermissible knowledge from victim outputs as opposed to forcing explicitly impermissible outputs from the victim. Through our information-theoretic framework, we show that to ensure safety against inferential adversaries, defense mechanisms must ensure information censorship, bounding the leakage of impermissible information. However, we prove that such defenses inevitably incur a safety-utility trade-off.

Generalization and robustness to input variation are core desiderata of machine learning research. Language varies along several axes, most importantly, language instance (e.g. French) and domain (e.g. news). While adapting NLP models to new languages within a single domain, or to new domains within a single language, is widely studied, research in joint adaptation is hampered by the lack of evaluation datasets. This prevents the transfer of NLP systems from well-resourced languages and domains to non-dominant language-domain combinations. To address this gap, we introduce M2QA, a multi-domain multilingual question answering benchmark. M2QA includes 13,500 SQuAD 2.0-style question-answer instances in German, Turkish, and Chinese for the domains of product reviews, news, and creative writing. We use M2QA to explore cross-lingual cross-domain performance of fine-tuned models and state-of-the-art LLMs and investigate modular approaches to domain and language adaptation. We witness 1) considerable performance variations across domain-language combinations within model classes and 2) considerable performance drops between source and target language-domain combinations across all model sizes. We demonstrate that M2QA is far from solved, and new methods to effectively transfer both linguistic and domain-specific information are necessary. We make M2QA publicly available at https://github.com/UKPLab/m2qa.

Spurious patterns refer to a mathematical association between two or more variables in a dataset that are not causally related. However, this notion of spuriousness, which is usually introduced due to sampling biases in the dataset, has classically lacked a formal definition. To address this gap, this work presents the first information-theoretic formalization of spuriousness in a dataset (given a split of spurious and core features) using a mathematical framework called Partial Information Decomposition (PID). Specifically, we disentangle the joint information content that the spurious and core features share about another target variable (e.g., the prediction label) into distinct components, namely unique, redundant, and synergistic information. We propose the use of unique information, with roots in Blackwell Sufficiency, as a novel metric to formally quantify dataset spuriousness and derive its desirable properties. We empirically demonstrate how higher unique information in the spurious features in a dataset could lead a model into choosing the spurious features over the core features for inference, often having low worst-group-accuracy. We also propose a novel autoencoder-based estimator for computing unique information that is able to handle high-dimensional image data. Finally, we also show how this unique information in the spurious feature is reduced across several dataset-based spurious-pattern-mitigation techniques such as data reweighting and varying levels of background mixing, demonstrating a novel tradeoff between unique information (spuriousness) and worst-group-accuracy., Comment: Accepted at ICML 2024 Workshop on Data-centric Machine Learning Research (DMLR): Datasets for Foundation Models

Exact unlearning was first introduced as a privacy mechanism that allowed a user to retract their data from machine learning models on request. Shortly after, inexact schemes were proposed to mitigate the impractical costs associated with exact unlearning. More recently unlearning is often discussed as an approach for removal of impermissible knowledge i.e. knowledge that the model should not possess such as unlicensed copyrighted, inaccurate, or malicious information. The promise is that if the model does not have a certain malicious capability, then it cannot be used for the associated malicious purpose. In this paper we revisit the paradigm in which unlearning is used for in Large Language Models (LLMs) and highlight an underlying inconsistency arising from in-context learning. Unlearning can be an effective control mechanism for the training phase, yet it does not prevent the model from performing an impermissible act during inference. We introduce a concept of ununlearning, where unlearned knowledge gets reintroduced in-context, effectively rendering the model capable of behaving as if it knows the forgotten knowledge. As a result, we argue that content filtering for impermissible knowledge will be required and even exact unlearning schemes are not enough for effective content regulation. We discuss feasibility of ununlearning for modern LLMs and examine broader implications.

Describing dynamics of a quantum system coupled to a complex many-body environment is a ubiquitous problem in quantum science. General non-Markovian environments are characterized by their influence matrix~(IM) -- a multi-time tensor arising from repeated interactions between the system and environment. While complexity of the most generic IM grows exponentially with the evolution time, recent works argued that for many instances of physical many-body environments, the IM is significantly less complex. This is thanks to area-law scaling of temporal entanglement, which quantifies the correlations between the past and the future states of the system. However, efficient classical algorithms for computing IM are only available for non-interacting environments or certain interacting 1D environments. Here, we study a learning algorithm for reconstructing IMs of large many-body environments simulated on a quantum processor. This hybrid algorithm involves experimentally collecting quantum measurement results of auxiliary qubits which are repeatedly coupled to the many-body environment, followed by a classical machine-learning construction of a matrix-product (MPS) representation of the IM. Using the example of 1D spin-chain environments, with a classically generated training dataset, we demonstrate that the algorithm allows scalable reconstruction of IMs for long evolution times. The reconstructed IM can be used to efficiently model quantum transport through an impurity, including cases with multiple leads and time-dependent controls. These results indicate the feasibility of characterizing long-time dynamics of complex environments using a limited number of measurements, under the assumption of a moderate temporal entanglement., Comment: 15 pages, 10 figures

Aligned instruction following models can better fulfill user requests than their unaligned counterparts. However, it has been shown that there is a length bias in evaluation of such models, and that training algorithms tend to exploit this bias by learning longer responses. In this work we show how to train models that can be controlled at inference time with instructions containing desired length constraints. Such models are superior in length instructed evaluations, outperforming standard instruction following models such as GPT4, Llama 3 and Mixtral., Comment: 13 pages

The availability of powerful open-source large language models (LLMs) opens exciting use-cases, such as automated personal assistants that adapt to the user's unique data and demands. Two key desiderata for such assistants are personalization-in the sense that the assistant should reflect the user's own style-and privacy-in the sense that users may prefer to always store their personal data locally, on their own computing device. We present a new design for such an automated assistant, for the specific use case of personal assistant for email generation, which we call Panza. Specifically, Panza can be both trained and inferenced locally on commodity hardware, and is personalized to the user's writing style. Panza's personalization features are based on a new technique called data playback, which allows us to fine-tune an LLM to better reflect a user's writing style using limited data. We show that, by combining efficient fine-tuning and inference methods, Panza can be executed entirely locally using limited resources-specifically, it can be executed within the same resources as a free Google Colab instance. Finally, our key methodological contribution is a careful study of evaluation metrics, and of how different choices of system components (e.g. the use of Retrieval-Augmented Generation or different fine-tuning approaches) impact the system's performance., Comment: Panza is available at https://github.com/IST-DASLab/PanzaMail

Language models are transforming the ways that their users engage with the world. Despite impressive capabilities, over-consumption of language model outputs risks propagating unchecked errors in the short-term and damaging human capabilities for critical thinking in the long-term. How can we develop scaffolding around language models to curate more appropriate use? We propose selective frictions for language model experiences, inspired by behavioral science interventions, to dampen misuse. Frictions involve small modifications to a user's experience, e.g., the addition of a button impeding model access and reminding a user of their expertise relative to the model. Through a user study with real humans, we observe shifts in user behavior from the imposition of a friction over LLMs in the context of a multi-topic question-answering task as a representative task that people may use LLMs for, e.g., in education and information retrieval. We find that frictions modulate over-reliance by driving down users' click rates while minimally affecting accuracy for those topics. Yet, frictions may have unintended effects. We find marked differences in users' click behaviors even on topics where frictions were not provisioned. Our contributions motivate further study of human-AI behavioral interaction to inform more effective and appropriate LLM use., Comment: NeurIPS Workshop on Behavioral ML; non-archival

In order for AI systems to communicate effectively with people, they must understand how we make decisions. However, people's decisions are not always rational, so the implicit internal models of human decision-making in Large Language Models (LLMs) must account for this. Previous empirical evidence seems to suggest that these implicit models are accurate -- LLMs offer believable proxies of human behavior, acting how we expect humans would in everyday interactions. However, by comparing LLM behavior and predictions to a large dataset of human decisions, we find that this is actually not the case: when both simulating and predicting people's choices, a suite of cutting-edge LLMs (GPT-4o & 4-Turbo, Llama-3-8B & 70B, Claude 3 Opus) assume that people are more rational than we really are. Specifically, these models deviate from human behavior and align more closely with a classic model of rational choice -- expected value theory. Interestingly, people also tend to assume that other people are rational when interpreting their behavior. As a consequence, when we compare the inferences that LLMs and people draw from the decisions of others using another psychological dataset, we find that these inferences are highly correlated. Thus, the implicit decision-making models of LLMs appear to be aligned with the human expectation that other people will act rationally, rather than with how people actually act.

One of the most striking findings in modern research on large language models (LLMs) is that scaling up compute during training leads to better results. However, less attention has been given to the benefits of scaling compute during inference. This survey focuses on these inference-time approaches. We explore three areas under a unified mathematical formalism: token-level generation algorithms, meta-generation algorithms, and efficient generation. Token-level generation algorithms, often called decoding algorithms, operate by sampling a single token at a time or constructing a token-level search space and then selecting an output. These methods typically assume access to a language model's logits, next-token distributions, or probability scores. Meta-generation algorithms work on partial or full sequences, incorporating domain knowledge, enabling backtracking, and integrating external information. Efficient generation methods aim to reduce token costs and improve the speed of generation. Our survey unifies perspectives from three research communities: traditional natural language processing, modern LLMs, and machine learning systems.

We derive a linear programming bound on the maximum cardinality of error-correcting codes in the sum-rank metric. Based on computational experiments on relatively small instances, we observe that the obtained bounds outperform all previously known bounds.

A circulant graph is a Cayley graph of a finite cyclic group. The Weisfeiler-Leman-dimension of a circulant graph $X$ with respect to the class of all circulant graphs is the smallest positive integer~$m$ such that the $m$-dimensional Weisfeiler-Leman algorithm correctly tests the isomorphism between $X$ and any other circulant graph. It is proved that for a circulant graph of order $n$ this dimension is less than or equal to $\Omega(n)+3$, where $\Omega(n)$ is the number of prime divisors of~$n$., Comment: 21 pages

Here we present the results of benchmarking of a quantum processor based on trapped $^{171}$Yb$^{+}$ ions by performing basic quantum machine learning algorithms. Specifically, we carry out a supervised binary classification of small-scale digits images, which are intentionally chosen so that they can be classified with 100% accuracy, using a quantum-enhanced Support Vector Machine algorithm with up to four qubits. In our work, we specifically consider different types of quantum encodings of the dataset and different levels of transpilation optimizations for the corresponding quantum circuits. For each quantum encoding, we obtain a classifier that is of 100% accuracy on both training and test sets, which demonstrates that the quantum processor can correctly solve the basic classification task considered. As we expect, with the increase of the capabilities quantum processors, they can become a useful tool for machine learning., Comment: 8 pages, 1 figure

Reinforcement learning with human feedback (RLHF) has become the dominant method to align large models to user preferences. Unlike fine-tuning, for which there are many studies regarding training data memorization, it is not clear how memorization is affected by or introduced in the RLHF alignment process. Understanding this relationship is important as real user data may be collected and used to align large models; if user data is memorized during RLHF and later regurgitated, this could raise privacy concerns. In addition to RLHF, other methods such as Direct Preference Optimization (DPO) and $\Psi$PO have gained popularity for learning directly from human preferences, removing the need for optimizing intermediary reward models with reinforcement learning. In this work, we analyze how training data memorization can surface and propagate through each phase of RLHF and direct preference learning. We focus our study on code completion models, as code completion is one of the most popular use cases for large language models. We find that RLHF significantly decreases the chance that data used for reward modeling and reinforcement learning is memorized in comparison to directly fine-tuning on this data, but that examples already memorized during the fine-tuning stage of RLHF, will, in the majority of cases, remain memorized after RLHF. In contrast, we find that aligning by learning directly from human preference data via a special case of $\Psi$PO, Identity Preference Optimization (IPO), increases the likelihood that training data is regurgitated compared to RLHF. Our work suggests that RLHF, as opposed to direct preference learning, is a safer way to mitigate the risk of regurgitating sensitive preference data when aligning large language models. We find our conclusions are robust across multiple code completion datasets, tasks, and model scales.

Deep neural networks, costly to train and rich in intellectual property value, are increasingly threatened by model extraction attacks that compromise their confidentiality. Previous attacks have succeeded in reverse-engineering model parameters up to a precision of float64 for models trained on random data with at most three hidden layers using cryptanalytical techniques. However, the process was identified to be very time consuming and not feasible for larger and deeper models trained on standard benchmarks. Our study evaluates the feasibility of parameter extraction methods of Carlini et al. [1] further enhanced by Canales-Mart\'inez et al. [2] for models trained on standard benchmarks. We introduce a unified codebase that integrates previous methods and reveal that computational tools can significantly influence performance. We develop further optimisations to the end-to-end attack and improve the efficiency of extracting weight signs by up to 14.8 times compared to former methods through the identification of easier and harder to extract neurons. Contrary to prior assumptions, we identify extraction of weights, not extraction of weight signs, as the critical bottleneck. With our improvements, a 16,721 parameter model with 2 hidden layers trained on MNIST is extracted within only 98 minutes compared to at least 150 minutes previously. Finally, addressing methodological deficiencies observed in previous studies, we propose new ways of robust benchmarking for future model extraction attacks.