R3HF: Reward Redistribution for Enhancing Reinforcement Learning from Human Feedback

Reinforcement learning from human feedback (RLHF) provides a paradigm for aligning large language models (LLMs) with human preferences. This involves the initial training of a reward model based on pairwise human feedback. The reward model is subsequently utilized in reinforcement learning to assess the scores of each generated sentence as a whole, further guiding the optimization of LLMs. However, current approaches have a significant shortcoming: \emph{They allocate a single, sparse, and delayed reward to an entire sequence of output}. This may overlook some significant individual contributions of each token towards the desired outcome. To overcome this limitation, our paper proposes a novel reward redistribution method called R3HF, which facilitates a more fine-grained, token-level reward allocation. Specifically, our method treats the reward prediction task of the reward model as a regression problem. As a result, the redistributed rewards are computed by evaluating the specific contribution of each token to the reward model's output. This detailed approach improves the model's understanding of language nuances, leading to more precise enhancements in its performance. Our method is crafted to integrate seamlessly with most current techniques while incurring minimal computational costs. Through comprehensive experiments across diverse datasets and tasks, we have verified the effectiveness and superiority of our approach.

Which Demographics do LLMs Default to During Annotation?

Demographics and cultural background of annotators influence the labels they assign in text annotation -- for instance, an elderly woman might find it offensive to read a message addressed to a "bro", but a male teenager might find it appropriate. It is therefore important to acknowledge label variations to not under-represent members of a society. Two research directions developed out of this observation in the context of using large language models (LLM) for data annotations, namely (1) studying biases and inherent knowledge of LLMs and (2) injecting diversity in the output by manipulating the prompt with demographic information. We combine these two strands of research and ask the question to which demographics an LLM resorts to when no demographics is given. To answer this question, we evaluate which attributes of human annotators LLMs inherently mimic. Furthermore, we compare non-demographic conditioned prompts and placebo-conditioned prompts (e.g., "you are an annotator who lives in house number 5") to demographics-conditioned prompts ("You are a 45 year old man and an expert on politeness annotation. How do you rate {instance}"). We study these questions for politeness and offensiveness annotations on the POPQUORN data set, a corpus created in a controlled manner to investigate human label variations based on demographics which has not been used for LLM-based analyses so far. We observe notable influences related to gender, race, and age in demographic prompting, which contrasts with previous studies that found no such effects.

RMP-YOLO: A Robust Motion Predictor for Partially Observable Scenarios even if You Only Look Once

We introduce RMP-YOLO, a unified framework designed to provide robust motion predictions even with incomplete input data. Our key insight stems from the observation that complete and reliable historical trajectory data plays a pivotal role in ensuring accurate motion prediction. Therefore, we propose a new paradigm that prioritizes the reconstruction of intact historical trajectories before feeding them into the prediction modules. Our approach introduces a novel scene tokenization module to enhance the extraction and fusion of spatial and temporal features. Following this, our proposed recovery module reconstructs agents' incomplete historical trajectories by leveraging local map topology and interactions with nearby agents. The reconstructed, clean historical data is then integrated into the downstream prediction modules. Our framework is able to effectively handle missing data of varying lengths and remains robust against observation noise, while maintaining high prediction accuracy. Furthermore, our recovery module is compatible with existing prediction models, ensuring seamless integration. Extensive experiments validate the effectiveness of our approach, and deployment in real-world autonomous vehicles confirms its practical utility. In the 2024 Waymo Motion Prediction Competition, our method, RMP-YOLO, achieves state-of-the-art performance, securing third place.

Generative AI for Energy Harvesting Internet of Things Network: Fundamental, Applications, and Opportunities

Internet of Things (IoT) devices are typically powered by small-sized batteries with limited energy storage capacity, requiring regular replacement or recharging. To reduce costs and maintain connectivity in IoT networks, energy harvesting technologies are regarded as a promising solution. Notably, due to its robust analytical and generative capabilities, generative artificial intelligence (GenAI) has demonstrated significant potential in optimizing energy harvesting networks. Therefore, we discuss key applications of GenAI in improving energy harvesting wireless networks for IoT in this article. Specifically, we first review the key technologies of GenAI and the architecture of energy harvesting wireless networks. Then, we show how GenAI can address different problems to improve the performance of the energy harvesting wireless networks. Subsequently, we present a case study of unmanned aerial vehicle (UAV)-enabled data collection and energy transfer. The case study shows distinctively the necessity of energy harvesting technology and verify the effectiveness of GenAI-based methods. Finally, we discuss some important open directions.

Multi-objective Aerial Collaborative Secure Communication Optimization via Generative Diffusion Model-enabled Deep Reinforcement Learning

Due to flexibility and low-cost, unmanned aerial vehicles (UAVs) are increasingly crucial for enhancing coverage and functionality of wireless networks. However, incorporating UAVs into next-generation wireless communication systems poses significant challenges, particularly in sustaining high-rate and long-range secure communications against eavesdropping attacks. In this work, we consider a UAV swarm-enabled secure surveillance network system, where a UAV swarm forms a virtual antenna array to transmit sensitive surveillance data to a remote base station (RBS) via collaborative beamforming (CB) so as to resist mobile eavesdroppers. Specifically, we formulate an aerial secure communication and energy efficiency multi-objective optimization problem (ASCEE-MOP) to maximize the secrecy rate of the system and to minimize the flight energy consumption of the UAV swarm. To address the non-convex, NP-hard and dynamic ASCEE-MOP, we propose a generative diffusion model-enabled twin delayed deep deterministic policy gradient (GDMTD3) method. Specifically, GDMTD3 leverages an innovative application of diffusion models to determine optimal excitation current weights and position decisions of UAVs. The diffusion models can better capture the complex dynamics and the trade-off of the ASCEE-MOP, thereby yielding promising solutions. Simulation results highlight the superior performance of the proposed approach compared with traditional deployment strategies and some other deep reinforcement learning (DRL) benchmarks. Moreover, performance analysis under various parameter settings of GDMTD3 and different numbers of UAVs verifies the robustness of the proposed approach.

GenBench: A Benchmarking Suite for Systematic Evaluation of Genomic Foundation Models

The Genomic Foundation Model (GFM) paradigm is expected to facilitate the extraction of generalizable representations from massive genomic data, thereby enabling their application across a spectrum of downstream applications. Despite advancements, a lack of evaluation framework makes it difficult to ensure equitable assessment due to experimental settings, model intricacy, benchmark datasets, and reproducibility challenges. In the absence of standardization, comparative analyses risk becoming biased and unreliable. To surmount this impasse, we introduce GenBench, a comprehensive benchmarking suite specifically tailored for evaluating the efficacy of Genomic Foundation Models. GenBench offers a modular and expandable framework that encapsulates a variety of state-of-the-art methodologies. Through systematic evaluations of datasets spanning diverse biological domains with a particular emphasis on both short-range and long-range genomic tasks, firstly including the three most important DNA tasks covering Coding Region, Non-Coding Region, Genome Structure, etc. Moreover, We provide a nuanced analysis of the interplay between model architecture and dataset characteristics on task-specific performance. Our findings reveal an interesting observation: independent of the number of parameters, the discernible difference in preference between the attention-based and convolution-based models on short- and long-range tasks may provide insights into the future design of GFM.

ADM: Accelerated Diffusion Model via Estimated Priors for Robust Motion Prediction under Uncertainties

Motion prediction is a challenging problem in autonomous driving as it demands the system to comprehend stochastic dynamics and the multi-modal nature of real-world agent interactions. Diffusion models have recently risen to prominence, and have proven particularly effective in pedestrian motion prediction tasks. However, the significant time consumption and sensitivity to noise have limited the real-time predictive capability of diffusion models. In response to these impediments, we propose a novel diffusion-based, acceleratable framework that adeptly predicts future trajectories of agents with enhanced resistance to noise. The core idea of our model is to learn a coarse-grained prior distribution of trajectory, which can skip a large number of denoise steps. This advancement not only boosts sampling efficiency but also maintains the fidelity of prediction accuracy. Our method meets the rigorous real-time operational standards essential for autonomous vehicles, enabling prompt trajectory generation that is vital for secure and efficient navigation. Through extensive experiments, our method speeds up the inference time to 136ms compared to standard diffusion model, and achieves significant improvement in multi-agent motion prediction on the Argoverse 1 motion forecasting dataset.

Two-Way Aerial Secure Communications via Distributed Collaborative Beamforming under Eavesdropper Collusion

Unmanned aerial vehicles (UAVs)-enabled aerial communication provides a flexible, reliable, and cost-effective solution for a range of wireless applications. However, due to the high line-of-sight (LoS) probability, aerial communications between UAVs are vulnerable to eavesdropping attacks, particularly when multiple eavesdroppers collude. In this work, we aim to introduce distributed collaborative beamforming (DCB) into UAV swarms and handle the eavesdropper collusion by controlling the corresponding signal distributions. Specifically, we consider a two-way DCB-enabled aerial communication between two UAV swarms and construct these swarms as two UAV virtual antenna arrays. Then, we minimize the two-way known secrecy capacity and the maximum sidelobe level to avoid information leakage from the known and unknown eavesdroppers, respectively. Simultaneously, we also minimize the energy consumption of UAVs for constructing virtual antenna arrays. Due to the conflicting relationships between secure performance and energy efficiency, we consider these objectives as a multi-objective optimization problem. Following this, we propose an enhanced multi-objective swarm intelligence algorithm via the characterized properties of the problem. Simulation results show that our proposed algorithm can obtain a set of informative solutions and outperform other state-of-the-art baseline algorithms. Experimental tests demonstrate that our method can be deployed in limited computing power platforms of UAVs and is beneficial for saving computational resources.

UAV-enabled Collaborative Beamforming via Multi-Agent Deep Reinforcement Learning

In this paper, we investigate an unmanned aerial vehicle (UAV)-assistant air-to-ground communication system, where multiple UAVs form a UAV-enabled virtual antenna array (UVAA) to communicate with remote base stations by utilizing collaborative beamforming. To improve the work efficiency of the UVAA, we formulate a UAV-enabled collaborative beamforming multi-objective optimization problem (UCBMOP) to simultaneously maximize the transmission rate of the UVAA and minimize the energy consumption of all UAVs by optimizing the positions and excitation current weights of all UAVs. This problem is challenging because these two optimization objectives conflict with each other, and they are non-concave to the optimization variables. Moreover, the system is dynamic, and the cooperation among UAVs is complex, making traditional methods take much time to compute the optimization solution for a single task. In addition, as the task changes, the previously obtained solution will become obsolete and invalid. To handle these issues, we leverage the multi-agent deep reinforcement learning (MADRL) to address the UCBMOP. Specifically, we use the heterogeneous-agent trust region policy optimization (HATRPO) as the basic framework, and then propose an improved HATRPO algorithm, namely HATRPO-UCB, where three techniques are introduced to enhance the performance. Simulation results demonstrate that the proposed algorithm can learn a better strategy compared with other methods. Moreover, extensive experiments also demonstrate the effectiveness of the proposed techniques.

Collaborative Ground-Space Communications via Evolutionary Multi-objective Deep Reinforcement Learning

In this paper, we propose a distributed collaborative beamforming (DCB)-based uplink communication paradigm for enabling ground-space direct communications. Specifically, DCB treats the terminals that are unable to establish efficient direct connections with the low Earth orbit (LEO) satellites as distributed antennas, forming a virtual antenna array to enhance the terminal-to-satellite uplink achievable rates and durations. However, such systems need multiple trade-off policies that variously balance the terminal-satellite uplink achievable rate, energy consumption of terminals, and satellite switching frequency to satisfy the scenario requirement changes. Thus, we perform a multi-objective optimization analysis and formulate a long-term optimization problem. To address availability in different terminal cluster scales, we reformulate this problem into an action space-reduced and universal multi-objective Markov decision process. Then, we propose an evolutionary multi-objective deep reinforcement learning algorithm to obtain the desirable policies, in which the low-value actions are masked to speed up the training process. As such, the applicability of a one-time trained model can cover more changing terminal-satellite uplink scenarios. Simulation results show that the proposed algorithm outmatches various baselines, and draw some useful insights. Specifically, it is found that DCB enables terminals that cannot reach the uplink achievable threshold to achieve efficient direct uplink transmission, which thus reveals that DCB is an effective solution for enabling direct ground-space communications. Moreover, it reveals that the proposed algorithm achieves multiple policies favoring different objectives and achieving near-optimal uplink achievable rates with low switching frequency.