Search

Answering reasoning-based complex questions over text and hybrid sources, including tables, is a challenging task. Recent advances in large language models (LLMs) have enabled in-context learning (ICL), allowing LLMs to acquire proficiency in a specific task using only a few demonstration samples (exemplars). A critical challenge in ICL is the selection of optimal exemplars, which can be either task-specific (static) or test-example-specific (dynamic). Static exemplars provide faster inference times and increased robustness across a distribution of test examples. In this paper, we propose an algorithm for static exemplar subset selection for complex reasoning tasks. We introduce EXPLORA, a novel exploration method designed to estimate the parameters of the scoring function, which evaluates exemplar subsets without incorporating confidence information. EXPLORA significantly reduces the number of LLM calls to ~11% of those required by state-of-the-art methods and achieves a substantial performance improvement of 12.24%. We open-source our code and data (https://github.com/kiranpurohit/EXPLORA).

Our work aims to build a model that performs dual tasks of image captioning and image generation while being trained on only one task. The central idea is to train an invertible model that learns a one-to-one mapping between the image and text embeddings. Once the invertible model is efficiently trained on one task, the image captioning, the same model can generate new images for a given text through the inversion process, with no additional training. This paper proposes a simple invertible neural network architecture for this problem and presents our current findings., Comment: Accepted as Tiny Paper at ICVGIP 2024 conference

Users can divulge sensitive information to proprietary LLM providers, raising significant privacy concerns. While open-source models, hosted locally on the user's machine, alleviate some concerns, models that users can host locally are often less capable than proprietary frontier models. Toward preserving user privacy while retaining the best quality, we propose Privacy-Conscious Delegation, a novel task for chaining API-based and local models. We utilize recent public collections of user-LLM interactions to construct a natural benchmark called PUPA, which contains personally identifiable information (PII). To study potential approaches, we devise PAPILLON, a multi-stage LLM pipeline that uses prompt optimization to address a simpler version of our task. Our best pipeline maintains high response quality for 85.5% of user queries while restricting privacy leakage to only 7.5%. We still leave a large margin to the generation quality of proprietary LLMs for future work. Our data and code will be available at https://github.com/siyan-sylvia-li/PAPILLON.

Deploying large language models (LLMs) locally on mobile devices is advantageous in scenarios where transmitting data to remote cloud servers is either undesirable due to privacy concerns or impractical due to network connection. Recent advancements (MLC, 2023a; Gerganov, 2023) have facilitated the local deployment of LLMs. However, local deployment also presents challenges, particularly in balancing quality (generative performance), latency, and throughput within the hardware constraints of mobile devices. In this paper, we introduce our lightweight, all-in-one automated benchmarking framework that allows users to evaluate LLMs on mobile devices. We provide a comprehensive benchmark of various popular LLMs with different quantization configurations (both weights and activations) across multiple mobile platforms with varying hardware capabilities. Unlike traditional benchmarks that assess full-scale models on high-end GPU clusters, we focus on evaluating resource efficiency (memory and power consumption) and harmful output for compressed models on mobile devices. Our key observations include i) differences in energy efficiency and throughput across mobile platforms; ii) the impact of quantization on memory usage, GPU execution time, and power consumption; and iii) accuracy and performance degradation of quantized models compared to their non-quantized counterparts; and iv) the frequency of hallucinations and toxic content generated by compressed LLMs on mobile devices., Comment: 10 pages

The main result of this article proves the nonvanishing of cuspidal cohomology for $GL(n)$ over a number field which is Galois over its maximal totally real subfield. The proof uses the internal structure of a strongly-pure weight that can possibly support cuspidal cohomology and the foundational work of Borel, Labesse, and Schwermer.

Many experimental platforms for quantum science depend on state control via laser fields. Frequently, however, the control fidelity is limited by optical phase noise. This is exacerbated in stabilized laser systems where high-frequency phase noise is an unavoidable consequence of feedback. Here we implement an optical feedforward technique to suppress laser phase noise in the STIRAP state transfer of ultracold RbCs molecules, across 114 THz, from a weakly bound Feshbach state to the rovibrational ground state. By performing over 100 state transfers on single molecules, we measure a significantly enhanced transfer efficiency of 98.7(1)% limited only by available laser intensity.

Backdoor data poisoning, inserted within instruction examples used to fine-tune a foundation Large Language Model (LLM) for downstream tasks (\textit{e.g.,} sentiment prediction), is a serious security concern due to the evasive nature of such attacks. The poisoning is usually in the form of a (seemingly innocuous) trigger word or phrase inserted into a very small fraction of the fine-tuning samples from a target class. Such backdoor attacks can: alter response sentiment, violate censorship, over-refuse (invoke censorship for legitimate queries), inject false content, or trigger nonsense responses (hallucinations). In this work we investigate the efficacy of instruction fine-tuning backdoor attacks as attack "hyperparameters" are varied under a variety of scenarios, considering: the trigger location in the poisoned examples; robustness to change in the trigger location, partial triggers, and synonym substitutions at test time; attack transfer from one (fine-tuning) domain to a related test domain; and clean-label vs. dirty-label poisoning. Based on our observations, we propose and evaluate two defenses against these attacks: i) a \textit{during-fine-tuning defense} based on word-frequency counts that assumes the (possibly poisoned) fine-tuning dataset is available and identifies the backdoor trigger tokens; and ii) a \textit{post-fine-tuning defense} based on downstream clean fine-tuning of the backdoored LLM with a small defense dataset. Finally, we provide a brief survey of related work on backdoor attacks and defenses., Comment: Under review

Ramanujan showed that $\tau(p) \equiv p^{11}+1 \pmod{691}$, where $\tau(n)$ is the $n$-th Fourier coefficient of the unique normalized cusp form of weight $12$ and full level, and the prime $691$ appears in the numerator of $\zeta(12)/\pi^{12}$ for the Riemann zeta function $\zeta(s)$. Searching for such congruences, it is shown that the prime $67$ appears in the numerator of $L(6,\chi)/(\pi^6 \sqrt{5})$, where $\chi$ is the unique nontrivial quadratic Dirichlet character modulo $5$ and $L(s,\chi)$ its Dirichlet $L$-function, giving rise to a congruence $f_\chi \equiv E^\circ_{6, \chi} \pmod{67}$ between a cusp form $f_\chi$ and an Eisenstein series $E^\circ_{6, \chi}$ of weight $6$ on $\Gamma_0(5)$ with nebentypus character $\chi.$, Comment: To appear in Involve - a journal of Mathematics, this article is based on an undergraduate research project at Fordham University

Given a pure motive $M$ over $\mathbb{Q}$ with a multilinear algebraic structure $\mathsf{s}$ on $M$, and given a representation $V$ of the group respecting $\mathsf{s}$, we describe a functorial transfer $M^V$. We formulate a criterion that guarantees when the two periods of $M^V$ are equal. This has an implication for the critical values of the $L$-function attached to $M^V.$ The criterion is explicated in a variety of examples such as: tensor product motives and Rankin-Selberg $L$-functions; orthogonal motives and the standard $L$-function for even orthogonal groups; twisted tensor motives and Asai $L$-functions.

A number field is said to be a CM-number field if it is a totally imaginary quadratic extension of a totally real number field. We define a totally imaginary number field to be of CM-type if it contains a CM-subfield, and of TR-type if it does not contain a CM-subfield. For quartic totally imaginary number fields when ordered by discriminant, we show that about 69.95% are of TR-type and about 33.05% are of CM-type. For a sextic totally imaginary number field we classify its type in terms of its Galois group and possibly some additional information about the location of complex conjugation in the Galois group., Comment: 16 pages including an appendix with 3 figures

Polymer droplets subjected to a heated environment have significance in several fields ranging from spray drying and powder formation to surface coating. In the present work, we investigate the evaporation of a high viscoelastic modulus aqueous polymeric droplet in an acoustically levitated environment. Depending on the laser irradiation intensity, we observe nucleation of a bubble in the dilute regime of polymer concentration, contrary to the previously observed bubble nucleation in a semi-dilute entangled regime for low viscoelastic modulus polymer droplets. After the bubble nucleation, a quasi steady bubble growth occurs depending on the laser irradiation intensity and concentrations. Our scaling analysis reveals that bubble growth follows Plesset-Zwick criteria independent of the viscoelastic properties of the polymer solution. Further, we establish that the onset of bubble growth has an inverse nonlinear dependence on the laser irradiation intensity. At high concentrations and laser irradiation intensities, we report the expansion and collapse of polymer membrane without rupture, indicating the formation of an interfacial skin with significant strength. The droplet oscillations are primarily driven by the presence of multiple bubbles and, to some extent, by the rotational motion of the droplet. Finally, depending on the nature of bubble growth, different types of precipitate form contrary to the different modes of atomization observed in low viscoelastic modulus polymer droplets.

We report on the transport of ultracold cesium and rubidium atoms over $37.2\,$cm in under $25\,$ms using an optical conveyor belt formed by two counter-propagating beams with a controllable frequency difference that generate a movable optical lattice. By carefully selecting the waists and focus positions, we are able to use two static Gaussian beams for the transport, avoiding the need for a Bessel beam or vari-focus lenses. We characterize the transport efficiency for both species, including a comparison of different transport trajectories, gaining insight into the loss mechanisms and finding the minimum jerk trajectory to be optimum. Using the optimized parameters, we are able to transport up to $7 \times 10^6$ cesium or rubidium atoms with an efficiency up to $75\,$%. To demonstrate the viability of our transport scheme for experiments employing quantum gas microscopy, we produce Bose-Einstein condensates of either species after transport and present measurements of the simultaneous transport of both species., Comment: 15 pages, 10 figures

We present Self-Driven Strategy Learning ($\textit{sdsl}$), a lightweight online learning methodology for automated reasoning tasks that involve solving a set of related problems. $\textit{sdsl}$ does not require offline training, but instead automatically constructs a dataset while solving earlier problems. It fits a machine learning model to this data which is then used to adjust the solving strategy for later problems. We formally define the approach as a set of abstract transition rules. We describe a concrete instance of the sdsl calculus which uses conditional sampling for generating data and random forests as the underlying machine learning model. We implement the approach on top of the Kissat solver and show that the combination of Kissat+$\textit{sdsl}$ certifies larger bounds and finds more counter-examples than other state-of-the-art bounded model checking approaches on benchmarks obtained from the latest Hardware Model Checking Competition., Comment: Condensed version published at FMCAD'23