The Perfect Blend: Redefining RLHF with Mixture of Judges

Reinforcement learning from human feedback (RLHF) has become the leading approach for fine-tuning large language models (LLM). However, RLHF has limitations in multi-task learning (MTL) due to challenges of reward hacking and extreme multi-objective optimization (i.e., trade-off of multiple and/or sometimes conflicting objectives). Applying RLHF for MTL currently requires careful tuning of the weights for reward model and data combinations. This is often done via human intuition and does not generalize. In this work, we introduce a novel post-training paradigm which we called Constrained Generative Policy Optimization (CGPO). The core of CGPO is Mixture of Judges (MoJ) with cost-efficient constrained policy optimization with stratification, which can identify the perfect blend in RLHF in a principled manner. It shows strong empirical results with theoretical guarantees, does not require extensive hyper-parameter tuning, and is plug-and-play in common post-training pipelines. Together, this can detect and mitigate reward hacking behaviors while reaching a pareto-optimal point across an extremely large number of objectives. Our empirical evaluations demonstrate that CGPO significantly outperforms standard RLHF algorithms like PPO and DPO across various tasks including general chat, STEM questions, instruction following, and coding. Specifically, CGPO shows improvements of 7.4% in AlpacaEval-2 (general chat), 12.5% in Arena-Hard (STEM & reasoning), and consistent gains in other domains like math and coding. Notably, PPO, while commonly used, is prone to severe reward hacking in popular coding benchmarks, which CGPO successfully addresses. This breakthrough in RLHF not only tackles reward hacking and extreme multi-objective optimization challenges but also advances the state-of-the-art in aligning general-purpose LLMs for diverse applications.

Language-Queried Target Sound Extraction Without Parallel Training Data

Language-queried target sound extraction (TSE) aims to extract specific sounds from mixtures based on language queries. Traditional fully-supervised training schemes require extensively annotated parallel audio-text data, which are labor-intensive. We introduce a language-free training scheme, requiring only unlabelled audio clips for TSE model training by utilizing the multi-modal representation alignment nature of the contrastive language-audio pre-trained model (CLAP). In a vanilla language-free training stage, target audio is encoded using the pre-trained CLAP audio encoder to form a condition embedding for the TSE model, while during inference, user language queries are encoded by CLAP text encoder. This straightforward approach faces challenges due to the modality gap between training and inference queries and information leakage from direct exposure to target audio during training. To address this, we propose a retrieval-augmented strategy. Specifically, we create an embedding cache using audio captions generated by a large language model (LLM). During training, target audio embeddings retrieve text embeddings from this cache to use as condition embeddings, ensuring consistent modalities between training and inference and eliminating information leakage. Extensive experiment results show that our retrieval-augmented approach achieves consistent and notable performance improvements over existing state-of-the-art with better generalizability.

ECAFormer: Low-light Image Enhancement using Cross Attention

Low-light image enhancement (LLIE) is vital for autonomous driving. Despite the importance, existing LLIE methods often prioritize robustness in overall brightness adjustment, which can come at the expense of detail preservation. To overcome this limitation,we propose the Hierarchical Mutual Enhancement via Cross-Attention transformer (ECAFormer), a novel network that utilizes Dual Multi-head Self Attention (DMSA) to enhance both visual and semantic features across scales, significantly preserving details during the process. The cross-attention mechanism in ECAFormer not only improves upon traditional enhancement techniques but also excels in maintaining a balance between global brightness adjustment and local detail retention. Our extensive experimental validation on renowned low-illumination datasets, including SID and LOL, and additional tests on dark road scenarios. or performance over existing methods in terms of illumination enhancement and noise reduction, while also optimizing computational complexity and parameter count, further boosting SSIM and PSNR metrics. Our project is available at https://github.com/ruanyudi/ECAFormer.

Stochastic Online Optimization for Cyber-Physical and Robotic Systems

We propose a novel gradient-based online optimization framework for solving stochastic programming problems that frequently arise in the context of cyber-physical and robotic systems. Our problem formulation accommodates constraints that model the evolution of a cyber-physical system, which has, in general, a continuous state and action space, is nonlinear, and where the state is only partially observed. We also incorporate an approximate model of the dynamics as prior knowledge into the learning process and show that even rough estimates of the dynamics can significantly improve the convergence of our algorithms. Our online optimization framework encompasses both gradient descent and quasi-Newton methods, and we provide a unified convergence analysis of our algorithms in a non-convex setting. We also characterize the impact of modeling errors in the system dynamics on the convergence rate of the algorithms. Finally, we evaluate our algorithms in simulations of a flexible beam, a four-legged walking robot, and in real-world experiments with a ping-pong playing robot.

CLAPSep: Leveraging Contrastive Pre-trained Models for Multi-Modal Query-Conditioned Target Sound Extraction

Universal sound separation (USS) aims to extract arbitrary types of sounds from real-world sound recordings. Language-queried target sound extraction (TSE) is an effective approach to achieving USS. Such systems consist of two components: a query network that converts user queries into conditional embeddings, and a separation network that extracts the target sound based on conditional embeddings. Existing methods mainly suffer from two issues: firstly, they require training a randomly initialized model from scratch, lacking the utilization of pre-trained models, and substantial data and computational resources are needed to ensure model convergence; secondly, existing methods need to jointly train a query network and a separation network, which tends to lead to overfitting. To address these issues, we build the CLAPSep model based on contrastive language-audio pre-trained model (CLAP). We achieve this by using a pre-trained text encoder of CLAP as the query network and introducing pre-trained audio encoder weights of CLAP into the separation network to fully utilize the prior knowledge embedded in the pre-trained model to assist in target sound extraction tasks. Extensive experimental results demonstrate that the proposed method saves training resources while ensuring the model's performance and generalizability. Additionally, we explore the model's ability to comprehensively utilize language/audio multi-modal and positive/negative multi-valent user queries, enhancing system performance while providing diversified application modes.

Poisoning Attacks against Recommender Systems: A Survey

Modern recommender systems (RS) have seen substantial success, yet they remain vulnerable to malicious activities, notably poisoning attacks. These attacks involve injecting malicious data into the training datasets of RS, thereby compromising their integrity and manipulating recommendation outcomes for gaining illicit profits. This survey paper provides a systematic and up-to-date review of the research landscape on Poisoning Attacks against Recommendation (PAR). A novel and comprehensive taxonomy is proposed, categorizing existing PAR methodologies into three distinct categories: Component-Specific, Goal-Driven, and Capability Probing. For each category, we discuss its mechanism in detail, along with associated methods. Furthermore, this paper highlights potential future research avenues in this domain. Additionally, to facilitate and benchmark the empirical comparison of PAR, we introduce an open-source library, ARLib, which encompasses a comprehensive collection of PAR models and common datasets. The library is released at https://github.com/CoderWZW/ARLib.

Extending Whisper with prompt tuning to target-speaker ASR

Target-speaker automatic speech recognition (ASR) aims to transcribe the desired speech of a target speaker from multi-talker overlapped utterances. Most of the existing target-speaker ASR (TS-ASR) methods involve either training from scratch or fully fine-tuning a pre-trained model, leading to significant training costs and becoming inapplicable to large foundation models. This work leverages prompt tuning, a parameter-efficient fine-tuning approach, to extend Whisper, a large-scale single-talker ASR model, to TS-ASR. Experimental results show that prompt tuning can achieve performance comparable to state-of-the-art full fine-tuning approaches while only requiring about 1% of task-specific model parameters. Notably, the original Whisper's features, such as inverse text normalization and timestamp prediction, are retained in target-speaker ASR, keeping the generated transcriptions natural and informative.

Adversarial Driving Behavior Generation Incorporating Human Risk Cognition for Autonomous Vehicle Evaluation

Autonomous vehicle (AV) evaluation has been the subject of increased interest in recent years both in industry and in academia. This paper focuses on the development of a novel framework for generating adversarial driving behavior of background vehicle interfering against the AV to expose effective and rational risky events. Specifically, the adversarial behavior is learned by a reinforcement learning (RL) approach incorporated with the cumulative prospect theory (CPT) which allows representation of human risk cognition. Then, the extended version of deep deterministic policy gradient (DDPG) technique is proposed for training the adversarial policy while ensuring training stability as the CPT action-value function is leveraged. A comparative case study regarding the cut-in scenario is conducted on a high fidelity Hardware-in-the-Loop (HiL) platform and the results demonstrate the adversarial effectiveness to infer the weakness of the tested AV.

Effective Long-Context Scaling of Foundation Models

We present a series of long-context LLMs that support effective context windows of up to 32,768 tokens. Our model series are built through continual pretraining from Llama 2 with longer training sequences and on a dataset where long texts are upsampled. We perform extensive evaluation on language modeling, synthetic context probing tasks, and a wide range of research benchmarks. On research benchmarks, our models achieve consistent improvements on most regular tasks and significant improvements on long-context tasks over Llama 2. Notably, with a cost-effective instruction tuning procedure that does not require human-annotated long instruction data, the 70B variant can already surpass gpt-3.5-turbo-16k's overall performance on a suite of long-context tasks. Alongside these results, we provide an in-depth analysis on the individual components of our method. We delve into Llama's position encodings and discuss its limitation in modeling long dependencies. We also examine the impact of various design choices in the pretraining process, including the data mix and the training curriculum of sequence lengths -- our ablation experiments suggest that having abundant long texts in the pretrain dataset is not the key to achieving strong performance, and we empirically verify that long context continual pretraining is more efficient and similarly effective compared to pretraining from scratch with long sequences.

Data-Efficient Online Learning of Ball Placement in Robot Table Tennis

We present an implementation of an online optimization algorithm for hitting a predefined target when returning ping-pong balls with a table tennis robot. The online algorithm optimizes over so-called interception policies, which define the manner in which the robot arm intercepts the ball. In our case, these are composed of the state of the robot arm (position and velocity) at interception time. Gradient information is provided to the optimization algorithm via the mapping from the interception policy to the landing point of the ball on the table, which is approximated with a black-box and a grey-box approach. Our algorithm is applied to a robotic arm with four degrees of freedom that is driven by pneumatic artificial muscles. As a result, the robot arm is able to return the ball onto any predefined target on the table after about 2-5 iterations. We highlight the robustness of our approach by showing rapid convergence with both the black-box and the grey-box gradients. In addition, the small number of iterations required to reach close proximity to the target also underlines the sample efficiency. A demonstration video can be found here: https://youtu.be/VC3KJoCss0k.