Search

This study addresses a critical gap in safety tuning practices for Large Language Models (LLMs) by identifying and tackling a refusal position bias within safety tuning data, which compromises the models' ability to appropriately refuse generating unsafe content. We introduce a novel approach, Decoupled Refusal Training (DeRTa), designed to empower LLMs to refuse compliance to harmful prompts at any response position, significantly enhancing their safety capabilities. DeRTa incorporates two novel components: (1) Maximum Likelihood Estimation (MLE) with Harmful Response Prefix, which trains models to recognize and avoid unsafe content by appending a segment of harmful response to the beginning of a safe response, and (2) Reinforced Transition Optimization (RTO), which equips models with the ability to transition from potential harm to safety refusal consistently throughout the harmful response sequence. Our empirical evaluation, conducted using LLaMA3 and Mistral model families across six attack scenarios, demonstrates that our method not only improves model safety without compromising performance but also surpasses well-known models such as GPT-4 in defending against attacks. Importantly, our approach successfully defends recent advanced attack methods (e.g., CodeAttack) that have jailbroken GPT-4 and LLaMA3-70B-Instruct. Our code and data can be found at https://github.com/RobustNLP/DeRTa.

Decision-making, a complicated task requiring various types of abilities, presents an excellent framework for assessing Large Language Models (LLMs). Our research investigates decision-making capabilities of LLMs through the lens of Game Theory. We focus specifically on games that support the simultaneous participation of more than two agents. We introduce GAMA($\gamma$)-Bench, which evaluates LLMs' Gaming Ability in Multi-Agent environments. $\gamma$-Bench includes eight classical multi-agent games and a scoring scheme specially designed to quantitatively assess LLMs' performance. Leveraging $\gamma$-Bench, we investigate LLMs' robustness, generalizability, and strategies for enhancement. Results reveal that while GPT-3.5 shows satisfying robustness, its generalizability is relatively limited. However, its performance can be improved through approaches such as Chain-of-Thought. Additionally, we evaluate twelve versions from six models, including GPT-3.5, GPT-4, Gemini, LLaMA-3.1, Mixtral, and Qwen-2. We find that Gemini-1.5-Pro outperforms other models with a score of $63.8$ out of $100$, followed by LLaMA-3.1-70B and GPT-4 with scores of $60.9$ and $60.5$, respectively. The code and experimental results are made publicly available via https://github.com/CUHK-ARISE/GAMABench., Comment: 11 pages of main text. 20 pages of appendices. 12 figures, 9 tables. Added models: Gemini-1.5-Pro, LLaMA-3.1-{7, 70, 405}B, Mixtral-8x{7, 22}B, Qwen-2-72B

The ability of Large Language Models (LLMs) to critique and refine their reasoning is crucial for their application in evaluation, feedback provision, and self-improvement. This paper introduces CriticBench, a comprehensive benchmark designed to assess LLMs' abilities to critique and rectify their reasoning across a variety of tasks. CriticBench encompasses five reasoning domains: mathematical, commonsense, symbolic, coding, and algorithmic. It compiles 15 datasets and incorporates responses from three LLM families. Utilizing CriticBench, we evaluate and dissect the performance of 17 LLMs in generation, critique, and correction reasoning, i.e., GQC reasoning. Our findings reveal: (1) a linear relationship in GQC capabilities, with critique-focused training markedly enhancing performance; (2) a task-dependent variation in correction effectiveness, with logic-oriented tasks being more amenable to correction; (3) GQC knowledge inconsistencies that decrease as model size increases; and (4) an intriguing inter-model critiquing dynamic, where stronger models are better at critiquing weaker ones, while weaker models can surprisingly surpass stronger ones in their self-critique. We hope these insights into the nuanced critique-correct reasoning of LLMs will foster further research in LLM critique and self-improvement., Comment: ACL 2024 Findings

Given a bounded domain $\Omega \subset {\mathbb R}^{n}$ with $n\ge2$, let $\phi $ is a Young function satisfying the doubling condition with the constant $K_\phi<2^{n}$. If $\Omega$ is a John domain, we show that $\Omega $ supports a $(\phi_{n}, \phi)$-Poincar\'e inequality. Conversely, assume additionally that $\Omega$ is simply connected domain when $n=2$ or a bounded domain which is quasiconformally equivalent to some uniform domain when $n\ge3$. If $\Omega$ supports a $(\phi_n, \phi)$-Poincar\'e inequality, we show that it is a John domain., Comment: arXiv admin note: substantial text overlap with arXiv:2305.04016

In this article, we study local Sobolev-Poincar\'e imbedding domains. The main result reads as below. \begin{enumerate} \item for $1\leq p\leq n$, a bounded uniform domain is also a local Sobolev-Poincar\'e imbedding domain of order $p$; conversely a local Sobolev-Poincar\'e imbedding domain of order $p$ is locally linearly connected $(LLC)$. A uniform domain is $(LLC)$. Conversely, with some very weak connecting assumption, a $(LLC)$ domain is uniform. \item for $n

The automatic evaluation of LLM-based agent intelligence is critical in developing advanced LLM-based agents. Although considerable effort has been devoted to developing human-annotated evaluation datasets, such as AlpacaEval, existing techniques are costly, time-consuming, and lack adaptability. In this paper, inspired by the popular language game ``Who is Spy'', we propose to use the word guessing game to assess the intelligence performance of LLMs. Given a word, the LLM is asked to describe the word and determine its identity (spy or not) based on its and other players' descriptions. Ideally, an advanced agent should possess the ability to accurately describe a given word using an aggressive description while concurrently maximizing confusion in the conservative description, enhancing its participation in the game. To this end, we first develop DEEP to evaluate LLMs' expression and disguising abilities. DEEP requires LLM to describe a word in aggressive and conservative modes. We then introduce SpyGame, an interactive multi-agent framework designed to assess LLMs' intelligence through participation in a competitive language-based board game. Incorporating multi-agent interaction, SpyGame requires the target LLM to possess linguistic skills and strategic thinking, providing a more comprehensive evaluation of LLMs' human-like cognitive abilities and adaptability in complex communication situations. The proposed evaluation framework is very easy to implement. We collected words from multiple sources, domains, and languages and used the proposed evaluation framework to conduct experiments. Extensive experiments demonstrate that the proposed DEEP and SpyGame effectively evaluate the capabilities of various LLMs, capturing their ability to adapt to novel situations and engage in strategic communication., Comment: Work in progress

A new resolvent-based method is developed to predict the space-time properties of the flow field. To overcome the deterioration of the prediction accuracy with the increasing distance between the measurements and predictions in the Resolvent-Based Estimation (RBE), the newly proposed method utilizes the RBE to estimate the relative energy distribution near the wall rather than the absolute energy directly estimated from the measurements. Using this extra information from RBE, the new method modifies the energy distribution of the spatially uniform and uncorrelated forcing that drives the flow system by minimizing the norm of the cross-spectral density (CSD) tensor of the error matrix in the near-wall region in comparison with the RBE-estimated one, and therefore it is named as the Resolvent-informed White-noise-based Estimation (RWE) method. For validation, three time-resolved direct numerical simulation (DNS) datasets with the friction Reynolds numbers $Re_\tau = 180$, 550, and 950 are generated, with various locations of measurements ranging from the near-wall region ($y^+ = 40$) to the upper bound of the logarithmic region ($y/h \approx 0.2$) for the predictions. Besides the RWE, three existing methods, i.e., the RBE, the $\lambda$-model, and the White-noise-Based Estimation (WBE), are also included for the validation. The performance of the RBE and $\lambda$-model in predicting the energy spectra shows a strong dependence on the measurement locations. The newly proposed RWE shows a low sensitivity on $Re_{\tau}$ and the measurement locations, which may range from the near-wall region to the upper bound of the logarithmic region, and has a high accuracy in predicting the energy spectra., Comment: 42 pages, 24 figures, accepted by Journal of Fluid Mechanics

Inspired by the fact that the matrix formulated by nonlocal similar patches in a natural image is of low rank, the rank approximation issue have been extensively investigated over the past decades, among which weighted nuclear norm minimization (WNNM) and weighted Schatten $p$-norm minimization (WSNM) are two prevailing methods have shown great superiority in various image restoration (IR) problems. Due to the physical characteristic of color images, color image restoration (CIR) is often a much more difficult task than its grayscale image counterpart. However, when applied to CIR, the traditional WNNM/WSNM method only processes three color channels individually and fails to consider their cross-channel correlations. Very recently, a quaternion-based WNNM approach (QWNNM) has been developed to mitigate this issue, which is capable of representing the color image as a whole in the quaternion domain and preserving the inherent correlation among the three color channels. Despite its empirical success, unfortunately, the convergence behavior of QWNNM has not been strictly studied yet. In this paper, on the one side, we extend the WSNM into quaternion domain and correspondingly propose a novel quaternion-based WSNM model (QWSNM) for tackling the CIR problems. Extensive experiments on two representative CIR tasks, including color image denoising and deblurring, demonstrate that the proposed QWSNM method performs favorably against many state-of-the-art alternatives, in both quantitative and qualitative evaluations. On the other side, more importantly, we preliminarily provide a theoretical convergence analysis, that is, by modifying the quaternion alternating direction method of multipliers (QADMM) through a simple continuation strategy, we theoretically prove that both the solution sequences generated by the QWNNM and QWSNM have fixed-point convergence guarantees., Comment: 46 pages, 10 figures; references added

In this study, we establish a hybrid high-order smoothed particle hydrodynamics (SPH) framework (MLS-TENO-SPH) for compressible flows with discontinuities, which is able to achieve genuine high-order convergence in smooth regions and also capture discontinuities well in non-smooth regions. The framework can be either fully Lagrangian, Eulerian or realizing arbitary-Lagrangian-Eulerian (ALE) feature enforcing the isotropic particle distribution in specific cases. In the proposed framework, the computational domain is divided into smooth regions and non-smooth regions, and these two regions are determined by a strong scale separation strategy in the targeted essentially non-oscillatory (TENO) scheme. In smooth regions, the moving-least-square (MLS) approximation is used for evaluating high-order derivative operator, which is able to realize genuine high-order construction; in non-smooth regions, the new TENO scheme based on Vila's framework with several new improvements will be deployed to capture discontinuities and high-wavenumber flow scales with low numerical dissipation. The present MLS-TENO-SPH method is validated with a set of challenging cases based on the Eulerian, Lagrangian or ALE framework. Numerical results demonstrate that the MLS-TENO-SPH method features lower numerical dissipation and higher efficiency than the conventional method, and can restore genuine high-order accuracy in smooth regions. Overall, the proposed framework serves as a new exploration in high-order SPH methods, which are potential for compressible flow simulations with shockwaves., Comment: 36 pages, 15 figures, accepted by Journal of Computational Physics on June 1st, 2023

Modern large language models (LLMs) like ChatGPT have shown remarkable performance on general language tasks but still struggle on complex reasoning tasks, which drives the research on cognitive behaviors of LLMs to explore human-like problem-solving strategies. Along this direction, one representative strategy is self-reflection, which asks an LLM to refine the solution with the feedback generated by itself iteratively. However, our study shows that such reflection-style methods suffer from the Degeneration-of-Thought (DoT) problem: once the LLM has established confidence in its solutions, it is unable to generate novel thoughts later through reflection even if its initial stance is incorrect. To address the DoT problem, we propose a Multi-Agent Debate (MAD) framework, in which multiple agents express their arguments in the state of "tit for tat" and a judge manages the debate process to obtain a final solution. Clearly, our MAD framework encourages divergent thinking in LLMs which would be helpful for tasks that require deep levels of contemplation. Experiment results on two challenging datasets, commonsense machine translation and counter-intuitive arithmetic reasoning, demonstrate the effectiveness of our MAD framework. Extensive analyses suggest that the adaptive break of debate and the modest level of "tit for tat" state are required for MAD to obtain good performance. Moreover, we find that LLMs might not be a fair judge if different LLMs are used for agents. Code is available at https://github.com/Skytliang/Multi-Agents-Debate., Comment: Work in progress

Large language models (LLMs) have demonstrated impressive capabilities in general scenarios, exhibiting a level of aptitude that approaches, in some aspects even surpasses, human-level intelligence. Among their numerous skills, the translation abilities of LLMs have received considerable attention. In contrast to traditional machine translation that focuses solely on source-target mapping, LLM-based translation can potentially mimic the human translation process that takes many preparatory steps to ensure high-quality translation. This work aims to explore this possibility by proposing the MAPS framework, which stands for Multi-Aspect Prompting and Selection. Specifically, we enable LLMs to first analyze the given source text and extract three aspects of translation-related knowledge: keywords, topics and relevant demonstrations to guide the translation process. To filter out the noisy and unhelpful knowledge, we employ a selection mechanism based on quality estimation. Experiments suggest that MAPS brings significant and consistent improvements over text-davinci-003 and Alpaca on eight translation directions from the latest WMT22 test sets. Our further analysis shows that the extracted knowledge is critical in resolving up to 59% of hallucination mistakes in translation. Code is available at https://github.com/zwhe99/MAPS-mt., Comment: V2: add more experiments and case studies; polish writing

Let $\Omega$ be a bounded domain in $\mathbb{R}^n$ with $n\ge2$ and $s\in(0,1)$. Assume that $\phi : [0, \infty) \to [0, \infty)$ be a Young function obeying the doubling condition with the constant $K_\phi<2^{\frac{n}{s}}$. We demonstrate that $\Omega $ supports a $(\phi_\frac{n}{s}, \phi)$-Poincar\'e inequality if it is is a John domain. Alternately, assume further that $\Omega$ is a bounded domain that is quasiconformally equivalent to some uniform domain when $n\ge3$ or a simply connected domain when $n=2$. We demonstrate $\Omega$ is a John domain if a $(\phi_\frac{n}{s}, \phi)$-Poincar\'e inequality holds.

Large language models (LLMs) like ChatGPT have exhibited remarkable abilities on a wide range of natural language processing~(NLP) tasks, including various machine translation abilities accomplished during chat. However, these models are only accessible through restricted APIs, which creates barriers to new research and advancements in the field. Therefore, we propose ParroT, a framework to enhance and regulate the translation abilities during chat based on open-source LLMs (e.g., LLaMA), human-written translation and feedback data. Specifically, ParroT reformulates translation data into the instruction-following style, and introduces a "$\mathbf{Hint}$" field for incorporating extra requirements to regulate the translation process. Accordingly, we propose three instruction types for finetuning ParroT models, including translation instruction, contrastive instruction, and error-guided instruction. Experiments on Flores subsets and WMT22 test sets suggest that translation instruction improves the translation performance of vanilla LLMs significantly while error-guided instruction can lead to further improvement, which demonstrates the importance of learning from low-quality translations annotated by humans. We also demonstrate the potential of automatic evaluation tools in providing quality information of translations, when constructing error-guided instructions for directions that lack human annotation data. Please refer to our Github project for more implementation details: https://github.com/wxjiao/ParroT, Comment: 12 pages; EMNLP 2023 (Findings)

In this paper, a new five-point targeted essentially non-oscillatory (TENO) scheme with adaptive dissipation is proposed. With the standard TENO weighting strategy, the cut-off parameter $C_T$ determines the nonlinear numerical dissipation of the resultant TENO scheme. Moreover, according to the dissipation-adaptive TENO5-A scheme, the choice of the cut-off parameter $C_T$ highly depends on the effective scale sensor. However, the scale sensor in TENO5-A can only roughly detect the discontinuity locations instead of evaluating the local flow wavenumber as desired. In this work, a new five-point scale sensor, which can estimate the local flow wavenumber accurately, is proposed to further improve the performance of TENO5-A. In combination with a hyperbolic tangent function, the new scale sensor is deployed to the TENO5-A framework for adapting the cut-off parameter $C_T$, i.e., the local nonlinear dissipation, according to the local flow wavenumber. Overall, sufficient numerical dissipation is generated to capture discontinuities, whereas a minimum amount of dissipation is delivered for better resolving the smooth flows. A set of benchmark cases is simulated to demonstrate the performance of the new TENO5-A scheme., Comment: 40 pages, 11 figures, accepted to be published by Communications in Computational Physics

This paper proposes a new metamaterial structure consisting of a periodically porous elastomer with pore coatings. This design enables us to engender finite deformation by a contactless load. As a case study, we apply thermal load to the pore coating and carry out a finite element analysis to probe instabilities and the associated phononic properties. It turns out that a novel buckling mode, preserving the nature of surface wrinkling in tubular structures, can be induced under a plane-strain setup, and a smaller size of the unit cell is attained compared to the counterpart of traditional buckled profile in soft porous elastomers. In particular, this buckling pattern is able to produce several bandgaps in different frequency ranges as the macroscopic mean strain increases. We further introduce a metallic core as local resonator, and the updated metamaterial allows a low-frequency bandgap, the bandgap width of which can be estimated by a simplified theoretical model. As more free parameters are involved in the structure, we perform a detailed parametric study to elucidate the influences of the modulus ratio between coating and matrix, the porosity, the core radius, and the macroscopic mean strain on the buckling initiation and the evolution of bandgap. Remarkably, a stiffer surface coating is prone to enhance the stability of the structure, which is contrary to existing results in film/substrate bilayers. It is expected that the current study could shed light on new insight into pattern formation and wave manipulation in porous elastomers., Comment: 16 pages, 19 figures

Higher neutron current densities can provide convenience for neutron experiments. Using neutron optical focusing elements, large flux beams transported to sample can be achieved. As one kind of focusing elements, polycapillary lens is very suitable for neutron absorption experiments such as PGAA and NDP technology. At present, a Neutron Physics and Application Spectrometer was in construction in CSNS, which is the first pulsed neutron source in China. To provide some suggestions and ideas for the following design of enhanced PGAA or NDP instrument with polycapillary lens in CSNS, a first neutron focusing experiment using polycapillary lens in CSNS was conducted. For 0.5-12.6 polychromatic beam, a focal spot with FWHM of 800 was obtained. As the value of wavelength increased, the beam size, transmission efficiency and gain increased. For cold neutron, the gain maintained in a level of 7., Comment: 12 pages, 13 figures, 12 references