Abstract:Reinforcement Learning from Human Feedback (RLHF) has emerged as a critical technique for training large language models. However, reward hacking-a phenomenon where models exploit flaws in the reward model-remains a significant barrier to achieving robust and scalable intelligence through long-term training. Existing studies have proposed uncertain reward model to address reward hacking, however, they often lack systematic or theoretical foundations, failing to model the uncertainty intrinsically emerging from preference data. In this paper, we propose the Probabilistic Uncertain Reward Model (PURM), a natural generalization of the classical Bradley-Terry reward model. PURM learns reward distributions directly from preference data and quantifies per-sample uncertainty via the average overlap area between reward distributions. To mitigate reward hacking, we further introduce an uncertainty-aware penalty into Proximal Policy Optimization (PPO), which leverages the learned uncertainty to dynamically balance reward optimization and exploration. We propose a lightweight and easy-to-use implementation of PURM. Experiments demonstrate that PURM significantly delays the onset of reward hacking while improving final reward performance, outperforming baseline methods in both stability and effectiveness.
Abstract:High complexity in precoding design for frequency division duplex systems necessitates streamlined solutions. Guided by Synesthesia of Machines (SoM), this paper introduces a heterogeneous multi-vehicle, multi-modal sensing aided precoding scheme within a vertical federated learning (VFL) framework, which significantly minimizes pilot sequence length while optimizing the system's sum rate. We address the challenges posed by local data heterogeneity due to varying on-board sensor configurations through a meticulously designed VFL training procedure. To extract valuable channel features from multi-modal sensing, we employ three distinct data preprocessing methods that convert raw data into informative representations relevant for precoding. Additionally, we propose an online training strategy based on VFL framework, enabling the scheme to adapt dynamically to fluctuations in user numbers. Numerical results indicate that our approach, utilizing short pilot sequences, closely approximates the performance of traditional optimization methods with perfect channel state information.
Abstract:In this paper, a novel multi-modal intelligent channel model for sixth-generation (6G) multiple-unmanned aerial vehicle (multi-UAV)-to-multi-vehicle communications is proposed. To thoroughly explore the mapping relationship between the physical environment and the electromagnetic space in the complex multi-UAV-to-multi-vehicle scenario, two new parameters, i.e., terrestrial traffic density (TTD) and aerial traffic density (ATD), are developed and a new sensing-communication intelligent integrated dataset is constructed in suburban scenario under different TTD and ATD conditions. With the aid of sensing data, i.e., light detection and ranging (LiDAR) point clouds, the parameters of static scatterers, terrestrial dynamic scatterers, and aerial dynamic scatterers in the electromagnetic space, e.g., number, distance, angle, and power, are quantified under different TTD and ATD conditions in the physical environment. In the proposed model, the channel non-stationarity and consistency on the time and space domains and the channel non-stationarity on the frequency domain are simultaneously mimicked. The channel statistical properties, such as time-space-frequency correlation function (TSF-CF), time stationary interval (TSI), and Doppler power spectral density (DPSD), are derived and simulated. Simulation results match ray-tracing (RT) results well, which verifies the accuracy of the proposed multi-UAV-to-multi-vehicle channel model.
Abstract:Given the importance of datasets for sensing-communication integration research, a novel simulation platform for constructing communication and multi-modal sensory dataset is developed. The developed platform integrates three high-precision software, i.e., AirSim, WaveFarer, and Wireless InSite, and further achieves in-depth integration and precise alignment of them. Based on the developed platform, a new synthetic intelligent multi-modal sensing-communication dataset for Synesthesia of Machines (SoM), named SynthSoM, is proposed. The SynthSoM dataset contains various air-ground multi-link cooperative scenarios with comprehensive conditions, including multiple weather conditions, times of the day, intelligent agent densities, frequency bands, and antenna types. The SynthSoM dataset encompasses multiple data modalities, including radio-frequency (RF) channel large-scale and small-scale fading data, RF millimeter wave (mmWave) radar sensory data, and non-RF sensory data, e.g., RGB images, depth maps, and light detection and ranging (LiDAR) point clouds. The quality of SynthSoM dataset is validated via statistics-based qualitative inspection and evaluation metrics through machine learning (ML) via real-world measurements. The SynthSoM dataset is open-sourced and provides consistent data for cross-comparing SoM-related algorithms.
Abstract:This paper proposes a novel sixth-generation (6G) multi-modal intelligent vehicle-to-vehicle (V2V) channel model from light detection and ranging (LiDAR) point clouds based on Synesthesia of Machines (SoM). To explore the mapping relationship between physical environment and electromagnetic space, a new V2V high-fidelity mixed sensing-communication integration simulation dataset with different vehicular traffic densities (VTDs) is constructed. Based on the constructed dataset, a novel scatterer recognition (ScaR) algorithm utilizing neural network SegNet is developed to recognize scatterer spatial attributes from LiDAR point clouds via SoM. In the developed ScaR algorithm, the mapping relationship between LiDAR point clouds and scatterers is explored, where the distribution of scatterers is obtained in the form of grid maps. Furthermore, scatterers are distinguished into dynamic and static scatterers based on LiDAR point cloud features, where parameters, e.g., distance, angle, and number, related to scatterers are determined. Through ScaR, dynamic and static scatterers change with the variation of LiDAR point clouds over time, which precisely models channel non-stationarity and consistency under different VTDs. Some important channel statistical properties, such as time-frequency correlation function (TF-CF) and Doppler power spectral density (DPSD), are obtained. Simulation results match well with ray-tracing (RT)-based results, thus demonstrating the necessity of exploring the mapping relationship and the utility of the proposed model.
Abstract:Low-rank adaptation (LoRA) reduces the computational and memory demands of fine-tuning large language models (LLMs) by approximating updates with low-rank matrices. However, low-rank approximation in two-dimensional space fails to capture high-dimensional structures within the target matrix. Recently, tensor decomposition methods have been explored for fine-tuning LLMs, leveraging their ability to extract structured information. Yet, these approaches primarily rely on random initialization, and the impact of initialization on tensor adaptation remains underexplored. In this paper, we reveal that random initialization significantly diverges from the validation loss achieved by full fine-tuning. To address this, we propose Weight-Decomposed Tensor Adaptation (DoTA), which leverages the Matrix Product Operator (MPO) decomposition of pre-trained weights for effective initialization in fine-tuning LLMs. Additionally, we introduce QDoTA, a quantized version of DoTA designed for 4-bit quantization. Experiments on commonsense and arithmetic reasoning tasks show that DoTA outperforms random initialization methods with fewer parameters. QDoTA further reduces memory consumption and achieves comparable performance to DoTA on commonsense reasoning tasks. We will release our code to support future research.
Abstract:Unstructured text data annotation and analysis are fundamental to management research, often relying on human annotators through crowdsourcing platforms. While Large Language Models (LLMs) promise to provide a cost-effective and efficient alternative to human annotation, there lacks a systematic workflow that evaluate when LLMs are suitable or how to proceed with LLM-based text annotation in a reproducible manner. This paper addresses this methodological gap by introducing the ``SILICON" (\textbf{S}ystematic \textbf{I}nference with \textbf{L}LMs for \textbf{I}nformation \textbf{C}lassificati\textbf{o}n and \textbf{N}otation) workflow. The workflow integrates established principles of human annotation with systematic prompt optimization and model selection, addressing challenges such as developing robust annotation guidelines, establishing high-quality human baselines, optimizing prompts, and ensuring reproducibility across LLMs. We validate the SILICON workflow through seven case studies covering common management research tasks, including business proposal evaluation, dialog intent and breakdown analysis, review attribute detection. Our findings highlight the importance of validating annotation guideline agreement, the superiority of expert-developed human baselines over crowdsourced ones, the iterative nature of prompt optimization, and the necessity of testing multiple LLMs. Notably, we propose a regression-based methodology to empirically compare LLM outputs across prompts and models. Our workflow advances management research by establishing reproducible processes for LLM-based annotation that maintain scientific rigor. We provide practical guidance for researchers to effectively navigate the evolving landscape of generative AI tools effectively while maintaining transparency and reproducibility.
Abstract:The potential benefits of integrated sensing and communication (ISAC) are anticipated to play a significant role in future sub-terahertz (sub-THz) systems. However, the beam squint effect is pronounced in sub-THz systems, expanding coverage areas while severely degrading communication performance. Existing hybrid precoding designs struggle to balance both functionalities in the presence of beam squint, limiting the performance gain achievable through ISAC. To address this challenge, we propose two squint-aware hybrid precoding schemes for sub-THz systems that proactively regulate the correlation between communication and sensing channels, leveraging the inherent degrees of freedom in the hardware to enhance integrated gain. We introduce a squint-aware optimization-based hybrid precoding algorithm (SA-Opt) and develop an unsupervised learning-assisted complex-valued squint-aware network (CSP-Net) to reduce complexity, tailoring its architecture to the specific data and task characteristics. The effectiveness of the proposed schemes is demonstrated through simulations.
Abstract:Channel prediction permits to acquire channel state information (CSI) without signaling overhead. However, almost all existing channel prediction methods necessitate the deployment of a dedicated model to accommodate a specific configuration. Leveraging the powerful modeling and multi-task learning capabilities of foundation models, we propose the first space-time-frequency (STF) wireless foundation model (WiFo) to address time-frequency channel prediction tasks in a one-for-all manner. Specifically, WiFo is initially pre-trained over massive and extensive diverse CSI datasets. Then, the model will be instantly used for channel prediction under various CSI configurations without any fine-tuning. We propose a masked autoencoder (MAE)-based network structure for WiFo to handle heterogeneous STF CSI data, and design several mask reconstruction tasks for self-supervised pre-training to capture the inherent 3D variations of CSI. To fully unleash its predictive power, we build a large-scale heterogeneous simulated CSI dataset consisting of 160K CSI samples for pre-training. Simulations validate its superior unified learning performance across multiple datasets and demonstrate its state-of-the-art (SOTA) zero-shot generalization performance via comparisons with other full-shot baselines.
Abstract:In this paper, we propose a novel dependency-aware task scheduling strategy for dynamic unmanned aerial vehicle-assisted connected autonomous vehicles (CAVs). Specifically, different computation tasks of CAVs consisting of multiple dependency subtasks are judiciously assigned to nearby CAVs or the base station for promptly completing tasks. Therefore, we formulate a joint scheduling priority and subtask assignment optimization problem with the objective of minimizing the average task completion time. The problem aims at improving the long-term system performance, which is reformulated as a Markov decision process. To solve the problem, we further propose a diffusion-based reinforcement learning algorithm, named Synthetic DDQN based Subtasks Scheduling, which can make adaptive task scheduling decision in real time. A diffusion model-based synthetic experience replay is integrated into the reinforcement learning framework, which can generate sufficient synthetic data in experience replay buffer, thereby significantly accelerating convergence and improving sample efficiency. Simulation results demonstrate the effectiveness of the proposed algorithm on reducing task completion time, comparing to benchmark schemes.