Reinforcement Learning from Human Feedback (RLHF) learns from the preference signal provided by a probabilistic preference model, which takes a prompt and two responses as input, and produces a score indicating the preference of one response against another. So far, the most popular RLHF paradigm is reward-based, which starts with an initial step of reward modeling, and the constructed reward is then used to provide a reward signal for the subsequent reward optimization stage. However, the existence of a reward function is a strong assumption and the reward-based RLHF is limited in expressivity and cannot capture the real-world complicated human preference. In this work, we provide theoretical insights for a recently proposed learning paradigm, Nash learning from human feedback (NLHF), which considered a general preference model and formulated the alignment process as a game between two competitive LLMs. The learning objective is to find a policy that consistently generates responses preferred over any competing policy while staying close to the initial model. The objective is defined as the Nash equilibrium (NE) of the KL-regularized preference model. We aim to make the first attempt to study the theoretical learnability of the KL-regularized NLHF by considering both offline and online settings. For the offline learning from a pre-collected dataset, we propose algorithms that are efficient under suitable coverage conditions of the dataset. For batch online learning from iterative interactions with a preference oracle, our proposed algorithm enjoys a finite sample guarantee under the structural condition of the underlying preference model. Our results connect the new NLHF paradigm with traditional RL theory, and validate the potential of reward-model-free learning under general preference.
相關內容
Large language models (LLMs) have shown remarkable progress in automated code generation. Yet, incorporating LLM-based code generation into real-life software projects poses challenges, as the generated code may contain errors in API usage, class, data structure, or missing project-specific information. As much of this project-specific context cannot fit into the prompts of LLMs, we must find ways to allow the model to explore the project-level code context. To this end, this paper puts forward a novel approach, termed ProCoder, which iteratively refines the project-level code context for precise code generation, guided by the compiler feedback. In particular, ProCoder first leverages compiler techniques to identify a mismatch between the generated code and the project's context. It then iteratively aligns and fixes the identified errors using information extracted from the code repository. We integrate ProCoder with two representative LLMs, i.e., GPT-3.5-Turbo and Code Llama (13B), and apply it to Python code generation. Experimental results show that ProCoder significantly improves the vanilla LLMs by over 80% in generating code dependent on project context, and consistently outperforms the existing retrieval-based code generation baselines.
This two-part paper studies a point-to-point resonant beam communication (RBCom) system, where two separately deployed retroreflectors are adopted to generate the resonant beam between the transmitter and the receiver, and analyzes the transmission rate of the considered system under both the quasi-static and mobile scenarios. Part I of this paper focuses on the quasi-static scenario where the locations of the transmitter and the receiver are relatively fixed. Specifically, we propose a new information-bearing scheme which adopts a synchronization-based amplitude modulation method to mitigate the echo interference caused by the reflected resonant beam. With this scheme, we show that the quasi-static RBCom channel is equivalent to a Markov channel and can be further simplified as an amplitude-constrained additive white Gaussian noise channel. Moreover, we develop an algorithm that jointly employs the bisection and exhaustive search to maximize its capacity upper and lower bounds. Finally, numerical results validate our analysis. Part II of this paper discusses the performance of the RBCom system under the mobile scenario.
Explainable Artificial Intelligence (XAI) strategies play a crucial part in increasing the understanding and trustworthiness of neural networks. Nonetheless, these techniques could potentially generate misleading explanations. Blinding attacks can drastically alter a machine learning algorithm's prediction and explanation, providing misleading information by adding visually unnoticeable artifacts into the input, while maintaining the model's accuracy. It poses a serious challenge in ensuring the reliability of XAI methods. To ensure the reliability of XAI methods poses a real challenge, we leverage statistical analysis to highlight the changes in CNN weights within a CNN following blinding attacks. We introduce a method specifically designed to limit the effectiveness of such attacks during the evaluation phase, avoiding the need for extra training. The method we suggest defences against most modern explanation-aware adversarial attacks, achieving an approximate decrease of ~99\% in the Attack Success Rate (ASR) and a ~91\% reduction in the Mean Square Error (MSE) between the original explanation and the defended (post-attack) explanation across three unique types of attacks.
Instruction-tuned Large Language Models (LLMs) have exhibited impressive language understanding and the capacity to generate responses that follow specific prompts. However, due to the computational demands associated with training these models, their applications often adopt a zero-shot setting. In this paper, we evaluate the zero-shot performance of two publicly accessible LLMs, ChatGPT and OpenAssistant, in the context of six Computational Social Science classification tasks, while also investigating the effects of various prompting strategies. Our experiments investigate the impact of prompt complexity, including the effect of incorporating label definitions into the prompt; use of synonyms for label names; and the influence of integrating past memories during foundation model training. The findings indicate that in a zero-shot setting, current LLMs are unable to match the performance of smaller, fine-tuned baseline transformer models (such as BERT-large). Additionally, we find that different prompting strategies can significantly affect classification accuracy, with variations in accuracy and F1 scores exceeding 10\%.
Conventional beamforming with fixed-position antenna (FPA) arrays has a fundamental trade-off between maximizing the signal power (array gain) over a desired direction and simultaneously minimizing the interference power over undesired directions. To overcome this limitation, this letter investigates the movable antenna (MA) array enhanced beamforming by exploiting the new degree of freedom (DoF) via antenna position optimization, in addition to the design of antenna weights. We show that by jointly optimizing the antenna positions vector (APV) and antenna weights vector (AWV) of a linear MA array, the full array gain can be achieved over the desired direction while null steering can be realized over all undesired directions, under certain numbers of MAs and null-steering directions. The optimal solutions for AWV and APV are derived in closed form, which reveal that the optimal AWV for MA arrays requires only the signal phase adjustment with a fixed amplitude. Numerical results validate our analytical solutions for MA array beamforming and show their superior performance to the conventional beamforming techniques with FPA arrays.
We explore a spectral initialization method that plays a central role in contemporary research on signal estimation in nonconvex scenarios. In a noiseless phase retrieval framework, we precisely analyze the method's performance in the high-dimensional limit when sensing vectors follow a multivariate Gaussian distribution for two rotationally invariant models of the covariance matrix C. In the first model C is a projector on a lower dimensional space while in the second it is a Wishart matrix. Our analytical results extend the well-established case when C is the identity matrix. Our examination shows that the introduction of biased spatial directions leads to a substantial improvement in the spectral method's effectiveness, particularly when the number of measurements is less than the signal's dimension. This extension also consistently reveals a phase transition phenomenon dependent on the ratio between sample size and signal dimension. Surprisingly, both of these models share the same threshold value.
Amidst the rapid evolution of LLMs, the significance of evaluation in comprehending and propelling these models forward is increasingly paramount. Evaluations have revealed that factors such as scaling, training types, architectures and other factors profoundly impact the performance of LLMs. However, the extent and nature of these impacts continue to be subjects of debate because most assessments have been restricted to a limited number of models and data points. Clarifying the effects of these factors on performance scores can be more effectively achieved through a statistical lens. Our study embarks on a thorough re-examination of these LLMs, targeting the inadequacies in current evaluation methods. With the advent of a uniform evaluation framework, our research leverages an expansive dataset of evaluation results, introducing a comprehensive statistical methodology. This includes the application of ANOVA, Tukey HSD tests, GAMM, and clustering technique, offering a robust and transparent approach to deciphering LLM performance data. Contrary to prevailing findings, our results challenge assumptions about emergent abilities and the influence of given training types and architectures in LLMs. These findings furnish new perspectives on the characteristics, intrinsic nature, and developmental trajectories of LLMs. By providing straightforward and reliable methods to scrutinize and reassess LLM performance data, this study contributes a nuanced perspective on LLM efficiency and potentials.
We consider a nonparametric regression model with continuous endogenous independent variables when only discrete instruments are available that are independent of the error term. While this framework is very relevant for applied research, its implementation is cumbersome, as the regression function becomes the solution to a nonlinear integral equation. We propose a simple iterative procedure to estimate such models and showcase some of its asymptotic properties. In a simulation experiment, we discuss the details of its implementation in the case when the instrumental variable is binary. We conclude with an empirical application in which we examine the effect of pollution on house prices in a short panel of U.S. counties.
We introduce RoDia, the first dataset for Romanian dialect identification from speech. The RoDia dataset includes a varied compilation of speech samples from five distinct regions of Romania, covering both urban and rural environments, totaling 2 hours of manually annotated speech data. Along with our dataset, we introduce a set of competitive models to be used as baselines for future research. The top scoring model achieves a macro F1 score of 59.83% and a micro F1 score of 62.08%, indicating that the task is challenging. We thus believe that RoDia is a valuable resource that will stimulate research aiming to address the challenges of Romanian dialect identification. We release our dataset at //github.com/codrut2/RoDia.
The advancement of Large Language Models (LLMs) has significantly boosted performance in natural language processing (NLP) tasks. However, the deployment of high-performance LLMs incurs substantial costs, primarily due to the increased number of parameters aimed at enhancing model performance. This has made the use of state-of-the-art LLMs more expensive for end-users. AI service providers, such as OpenAI and Anthropic, often offer multiple versions of LLMs with varying prices and performance. However, end-users still face challenges in choosing the appropriate LLM for their tasks that balance result quality with cost. We introduce SMART, Scaling Models Adaptively for Reduced Token Fees, a novel LLM framework designed to minimize the inference costs of NLP tasks while ensuring sufficient result quality. It enables users to specify an accuracy constraint in terms of the equivalence of outputs to those of the most powerful LLM. SMART then generates results that deviate from the outputs of this LLM only with a probability below a user-defined threshold. SMART employs a profiling phase that evaluates the performance of multiple LLMs to identify those that meet the user-defined accuracy level. SMART optimizes the tradeoff between profiling overheads and the anticipated cost savings resulting from profiling. Moreover, our approach significantly reduces inference costs by strategically leveraging a mix of LLMs. Our experiments on three real-world datasets show that, based on OpenAI models, SMART achieves significant cost savings, up to 25.6x in comparison to GPT-4.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191