Semi-Supervised Reward Modeling via Iterative Self-Training

Reward models (RM) capture the values and preferences of humans and play a central role in Reinforcement Learning with Human Feedback (RLHF) to align pretrained large language models (LLMs). Traditionally, training these models relies on extensive human-annotated preference data, which poses significant challenges in terms of scalability and cost. To overcome these limitations, we propose Semi-Supervised Reward Modeling (SSRM), an approach that enhances RM training using unlabeled data. Given an unlabeled dataset, SSRM involves three key iterative steps: pseudo-labeling unlabeled examples, selecting high-confidence examples through a confidence threshold, and supervised finetuning on the refined dataset. Across extensive experiments on various model configurations, we demonstrate that SSRM significantly improves reward models without incurring additional labeling costs. Notably, SSRM can achieve performance comparable to models trained entirely on labeled data of equivalent volumes. Overall, SSRM substantially reduces the dependency on large volumes of human-annotated data, thereby decreasing the overall cost and time involved in training effective reward models.

Interpretable Preferences via Multi-Objective Reward Modeling and Mixture-of-Experts

Reinforcement learning from human feedback (RLHF) has emerged as the primary method for aligning large language models (LLMs) with human preferences. The RLHF process typically starts by training a reward model (RM) using human preference data. Conventional RMs are trained on pairwise responses to the same user request, with relative ratings indicating which response humans prefer. The trained RM serves as a proxy for human preferences. However, due to the black-box nature of RMs, their outputs lack interpretability, as humans cannot intuitively understand why an RM thinks a response is good or not. As RMs act as human preference proxies, we believe they should be human-interpretable to ensure that their internal decision processes are consistent with human preferences and to prevent reward hacking in LLM alignment. To build RMs with interpretable preferences, we propose a two-stage approach: i) train an Absolute-Rating Multi-Objective Reward Model (ArmoRM) with multi-dimensional absolute-rating data, each dimension corresponding to a human-interpretable objective (e.g., honesty, verbosity, safety); ii) employ a Mixture-of-Experts (MoE) strategy with a gating network that automatically selects the most suitable reward objectives based on the context. We efficiently trained an ArmoRM with Llama-3 8B and a gating network consisting of a shallow MLP on top of the ArmoRM. Our trained model, ArmoRM-Llama3-8B, obtains state-of-the-art performance on RewardBench, a benchmark evaluating RMs for language modeling. Notably, the performance of our model surpasses the LLM-as-a-judge method with GPT-4 judges by a margin, and approaches the performance of the much larger Nemotron-4 340B reward model.

RLHF Workflow: From Reward Modeling to Online RLHF

We present the workflow of Online Iterative Reinforcement Learning from Human Feedback (RLHF) in this technical report, which is widely reported to outperform its offline counterpart by a large margin in the recent large language model (LLM) literature. However, existing open-source RLHF projects are still largely confined to the offline learning setting. In this technical report, we aim to fill in this gap and provide a detailed recipe that is easy to reproduce for online iterative RLHF. In particular, since online human feedback is usually infeasible for open-source communities with limited resources, we start by constructing preference models using a diverse set of open-source datasets and use the constructed proxy preference model to approximate human feedback. Then, we discuss the theoretical insights and algorithmic principles behind online iterative RLHF, followed by a detailed practical implementation. Our trained LLM, SFR-Iterative-DPO-LLaMA-3-8B-R, achieves impressive performance on LLM chatbot benchmarks, including AlpacaEval-2, Arena-Hard, and MT-Bench, as well as other academic benchmarks such as HumanEval and TruthfulQA. We have shown that supervised fine-tuning (SFT) and iterative RLHF can obtain state-of-the-art performance with fully open-source datasets. Further, we have made our models, curated datasets, and comprehensive step-by-step code guidebooks publicly available. Please refer to https://github.com/RLHFlow/RLHF-Reward-Modeling and https://github.com/RLHFlow/Online-RLHF for more detailed information.

Den-SOFT: Dense Space-Oriented Light Field DataseT for 6-DOF Immersive Experience

We have built a custom mobile multi-camera large-space dense light field capture system, which provides a series of high-quality and sufficiently dense light field images for various scenarios. Our aim is to contribute to the development of popular 3D scene reconstruction algorithms such as IBRnet, NeRF, and 3D Gaussian splitting. More importantly, the collected dataset, which is much denser than existing datasets, may also inspire space-oriented light field reconstruction, which is potentially different from object-centric 3D reconstruction, for immersive VR/AR experiences. We utilized a total of 40 GoPro 10 cameras, capturing images of 5k resolution. The number of photos captured for each scene is no less than 1000, and the average density (view number within a unit sphere) is 134.68. It is also worth noting that our system is capable of efficiently capturing large outdoor scenes. Addressing the current lack of large-space and dense light field datasets, we made efforts to include elements such as sky, reflections, lights and shadows that are of interest to researchers in the field of 3D reconstruction during the data capture process. Finally, we validated the effectiveness of our provided dataset on three popular algorithms and also integrated the reconstructed 3DGS results into the Unity engine, demonstrating the potential of utilizing our datasets to enhance the realism of virtual reality (VR) and create feasible interactive spaces. The dataset is available at our project website.

Arithmetic Control of LLMs for Diverse User Preferences: Directional Preference Alignment with Multi-Objective Rewards

Fine-grained control over large language models (LLMs) remains a significant challenge, hindering their adaptability to diverse user needs. While Reinforcement Learning from Human Feedback (RLHF) shows promise in aligning LLMs, its reliance on scalar rewards often limits its ability to capture diverse user preferences in real-world applications. To address this limitation, we introduce the Directional Preference Alignment (DPA) framework. Unlike the scalar-reward RLHF, DPA incorporates multi-objective reward modeling to represent diverse preference profiles. Additionally, DPA models user preferences as directions (i.e., unit vectors) in the reward space to achieve user-dependent preference control. Our method involves training a multi-objective reward model and then fine-tuning the LLM with a preference-conditioned variant of Rejection Sampling Finetuning (RSF), an RLHF method adopted by Llama 2. This method enjoys a better performance trade-off across various reward objectives. In comparison with the scalar-reward RLHF, DPA offers users intuitive control over LLM generation: they can arithmetically specify their desired trade-offs (e.g., more helpfulness with less verbosity). We also validate the effectiveness of DPA with real-world alignment experiments on Mistral-7B. Our method provides straightforward arithmetic control over the trade-off between helpfulness and verbosity while maintaining competitive performance with strong baselines such as Direct Preference Optimization (DPO).

Enhancing Compositional Generalization via Compositional Feature Alignment

Real-world applications of machine learning models often confront data distribution shifts, wherein discrepancies exist between the training and test data distributions. In the common multi-domain multi-class setup, as the number of classes and domains scales up, it becomes infeasible to gather training data for every domain-class combination. This challenge naturally leads the quest for models with Compositional Generalization (CG) ability, where models can generalize to unseen domain-class combinations. To delve into the CG challenge, we develop CG-Bench, a suite of CG benchmarks derived from existing real-world image datasets, and observe that the prevalent pretraining-finetuning paradigm on foundational models, such as CLIP and DINOv2, struggles with the challenge. To address this challenge, we propose Compositional Feature Alignment (CFA), a simple two-stage finetuning technique that i) learns two orthogonal linear heads on a pretrained encoder with respect to class and domain labels, and ii) fine-tunes the encoder with the newly learned head frozen. We theoretically and empirically justify that CFA encourages compositional feature learning of pretrained models. We further conduct extensive experiments on CG-Bench for CLIP and DINOv2, two powerful pretrained vision foundation models. Experiment results show that CFA outperforms common finetuning techniques in compositional generalization, corroborating CFA's efficacy in compositional feature learning.

Invariant-Feature Subspace Recovery: A New Class of Provable Domain Generalization Algorithms

Domain generalization asks for models trained over a set of training environments to generalize well in unseen test environments. Recently, a series of algorithms such as Invariant Risk Minimization (IRM) have been proposed for domain generalization. However, Rosenfeld et al. (2021) shows that in a simple linear data model, even if non-convexity issues are ignored, IRM and its extensions cannot generalize to unseen environments with less than $d_s+1$ training environments, where $d_s$ is the dimension of the spurious-feature subspace. In this work, we propose Invariant-feature Subspace Recovery (ISR): a new class of algorithms to achieve provable domain generalization across the settings of classification and regression problems. First, in the binary classification setup of Rosenfeld et al. (2021), we show that our first algorithm, ISR-Mean, can identify the subspace spanned by invariant features from the first-order moments of the class-conditional distributions, and achieve provable domain generalization with $d_s+1$ training environments. Our second algorithm, ISR-Cov, further reduces the required number of training environments to $O(1)$ using the information of second-order moments. Notably, unlike IRM, our algorithms bypass non-convexity issues and enjoy global convergence guarantees. Next, we extend ISR-Mean to the more general setting of multi-class classification and propose ISR-Multiclass, which leverages class information and provably recovers the invariant-feature subspace with $\lceil d_s/k\rceil+1$ training environments for $k$-class classification. Finally, for regression problems, we propose ISR-Regression that can identify the invariant-feature subspace with $d_s+1$ training environments. Empirically, we demonstrate the superior performance of our ISRs on synthetic benchmarks. Further, ISR can be used as post-processing methods for feature extractors such as neural nets.

SAM-CLIP: Merging Vision Foundation Models towards Semantic and Spatial Understanding

The landscape of publicly available vision foundation models (VFMs), such as CLIP and Segment Anything Model (SAM), is expanding rapidly. VFMs are endowed with distinct capabilities stemming from their pre-training objectives. For instance, CLIP excels in semantic understanding, while SAM specializes in spatial understanding for segmentation. In this work, we introduce a simple recipe to efficiently merge VFMs into a unified model that assimilates their expertise. Our proposed method integrates multi-task learning, continual learning techniques, and teacher-student distillation. This strategy entails significantly less computational cost compared to traditional multi-task training from scratch. Additionally, it only demands a small fraction of the pre-training datasets that were initially used to train individual models. By applying our method to SAM and CLIP, we derive SAM-CLIP: a unified model that amalgamates the strengths of SAM and CLIP into a single backbone, making it apt for edge device applications. We show that SAM-CLIP learns richer visual representations, equipped with both localization and semantic features, suitable for a broad range of vision tasks. SAM-CLIP obtains improved performance on several head probing tasks when compared with SAM and CLIP. We further show that SAM-CLIP not only retains the foundational strengths of its precursor models but also introduces synergistic functionalities, most notably in zero-shot semantic segmentation, where SAM-CLIP establishes new state-of-the-art results on 5 benchmarks. It outperforms previous models that are specifically designed for this task by a large margin, including +6.8% and +5.9% mean IoU improvement on Pascal-VOC and COCO-Stuff datasets, respectively.

Gradual Domain Adaptation: Theory and Algorithms

Unsupervised domain adaptation (UDA) adapts a model from a labeled source domain to an unlabeled target domain in a one-off way. Though widely applied, UDA faces a great challenge whenever the distribution shift between the source and the target is large. Gradual domain adaptation (GDA) mitigates this limitation by using intermediate domains to gradually adapt from the source to the target domain. In this work, we first theoretically analyze gradual self-training, a popular GDA algorithm, and provide a significantly improved generalization bound compared with Kumar et al. (2020). Our theoretical analysis leads to an interesting insight: to minimize the generalization error on the target domain, the sequence of intermediate domains should be placed uniformly along the Wasserstein geodesic between the source and target domains. The insight is particularly useful under the situation where intermediate domains are missing or scarce, which is often the case in real-world applications. Based on the insight, we propose $\textbf{G}$enerative Gradual D$\textbf{O}$main $\textbf{A}$daptation with Optimal $\textbf{T}$ransport (GOAT), an algorithmic framework that can generate intermediate domains in a data-dependent way. More concretely, we first generate intermediate domains along the Wasserstein geodesic between two given consecutive domains in a feature space, then apply gradual self-training to adapt the source-trained classifier to the target along the sequence of intermediate domains. Empirically, we demonstrate that our GOAT framework can improve the performance of standard GDA when the given intermediate domains are scarce, significantly broadening the real-world application scenarios of GDA. Our code is available at https://github.com/yifei-he/GOAT.

Speciality vs Generality: An Empirical Study on Catastrophic Forgetting in Fine-tuning Foundation Models

Foundation models, including Vision Language Models (VLMs) and Large Language Models (LLMs), possess the $generality$ to handle diverse distributions and tasks, which stems from their extensive pre-training datasets. The fine-tuning of foundation models is a common practice to enhance task performance or align the model's behavior with human expectations, allowing them to gain $speciality$. However, the small datasets used for fine-tuning may not adequately cover the diverse distributions and tasks encountered during pre-training. Consequently, the pursuit of speciality during fine-tuning can lead to a loss of {generality} in the model, which is related to catastrophic forgetting (CF) in deep learning. In this study, we demonstrate this phenomenon in both VLMs and LLMs. For instance, fine-tuning VLMs like CLIP on ImageNet results in a loss of generality in handling diverse distributions, and fine-tuning LLMs like Galactica in the medical domain leads to a loss in following instructions and common sense. To address the trade-off between the speciality and generality, we investigate multiple regularization methods from continual learning, the weight averaging method (Wise-FT) from out-of-distributional (OOD) generalization, which interpolates parameters between pre-trained and fine-tuned models, and parameter-efficient fine-tuning methods like Low-Rank Adaptation (LoRA). Our findings show that both continual learning and Wise-ft methods effectively mitigate the loss of generality, with Wise-FT exhibiting the strongest performance in balancing speciality and generality.