Sequence to Sequence Reward Modeling: Improving RLHF by Language Feedback

Aligning the behavior of Large language models (LLMs) with human intentions and values remains a critical challenge. Reinforcement learning from human feedback (RLHF) aligns LLMs by training a reward model (RM) on human preferences and fine-tuning the LLMs to maximize RM feedback. Despite its effectiveness and popularity, RLHF is prone to biased local optimization. It means RM fails to provide feedback that accurately aligns with human preference, causing LLMs to explore unexpected generalizations, and failing to achieve alignment objectives. To mitigate this issue, we propose a novel \textit{sequence-to-sequence (seq2seq) reward modeling} method. Its key insight is that learning from language feedback rather than scalar feedback improves RLHF without additional annotations. We replaced the reward modeling target from binary maximum likelihood estimation (MLE) with sequence MLE. This method enables richer and fine-grained language feedback without additional annotations, models, or training stages. Our experiments demonstrated its effectiveness, specifically, reducing the refusal-to-response paradigm in single-turn safety dialogues and the long-response bias in text summarization tasks. We provide further analysis that seq2seq RM improves RLHF performance across 2B and 7B LLMs on 3 NLP tasks, achieving an average win rate of 76.9\%. We further show that seq2seq RM can still improve the performance of RLHF under out-of-distribution prompts.

ProgressGym: Alignment with a Millennium of Moral Progress

Frontier AI systems, including large language models (LLMs), hold increasing influence over the epistemology of human users. Such influence can reinforce prevailing societal values, potentially contributing to the lock-in of misguided moral beliefs and, consequently, the perpetuation of problematic moral practices on a broad scale. We introduce progress alignment as a technical solution to mitigate this imminent risk. Progress alignment algorithms learn to emulate the mechanics of human moral progress, thereby addressing the susceptibility of existing alignment methods to contemporary moral blindspots. To empower research in progress alignment, we introduce ProgressGym, an experimental framework allowing the learning of moral progress mechanics from history, in order to facilitate future progress in real-world moral decisions. Leveraging 9 centuries of historical text and 18 historical LLMs, ProgressGym enables codification of real-world progress alignment challenges into concrete benchmarks. Specifically, we introduce three core challenges: tracking evolving values (PG-Follow), preemptively anticipating moral progress (PG-Predict), and regulating the feedback loop between human and AI value shifts (PG-Coevolve). Alignment methods without a temporal dimension are inapplicable to these tasks. In response, we present lifelong and extrapolative algorithms as baseline methods of progress alignment, and build an open leaderboard soliciting novel algorithms and challenges. The framework and the leaderboard are available at https://github.com/PKU-Alignment/ProgressGym and https://huggingface.co/spaces/PKU-Alignment/ProgressGym-LeaderBoard respectively.

SafeSora: Towards Safety Alignment of Text2Video Generation via a Human Preference Dataset

To mitigate the risk of harmful outputs from large vision models (LVMs), we introduce the SafeSora dataset to promote research on aligning text-to-video generation with human values. This dataset encompasses human preferences in text-to-video generation tasks along two primary dimensions: helpfulness and harmlessness. To capture in-depth human preferences and facilitate structured reasoning by crowdworkers, we subdivide helpfulness into 4 sub-dimensions and harmlessness into 12 sub-categories, serving as the basis for pilot annotations. The SafeSora dataset includes 14,711 unique prompts, 57,333 unique videos generated by 4 distinct LVMs, and 51,691 pairs of preference annotations labeled by humans. We further demonstrate the utility of the SafeSora dataset through several applications, including training the text-video moderation model and aligning LVMs with human preference by fine-tuning a prompt augmentation module or the diffusion model. These applications highlight its potential as the foundation for text-to-video alignment research, such as human preference modeling and the development and validation of alignment algorithms.

PKU-SafeRLHF: A Safety Alignment Preference Dataset for Llama Family Models

In this work, we introduce the PKU-SafeRLHF dataset, designed to promote research on safety alignment in large language models (LLMs). As a sibling project to SafeRLHF and BeaverTails, we separate annotations of helpfulness and harmlessness for question-answering pairs, providing distinct perspectives on these coupled attributes. Overall, we provide 44.6k refined prompts and 265k question-answer pairs with safety meta-labels for 19 harm categories and three severity levels ranging from minor to severe, with answers generated by Llama-family models. Based on this, we collected 166.8k preference data, including dual-preference (helpfulness and harmlessness decoupled) and single-preference data (trade-off the helpfulness and harmlessness from scratch), respectively. Using the large-scale annotation data, we further train severity-sensitive moderation for the risk control of LLMs and safety-centric RLHF algorithms for the safety alignment of LLMs. We believe this dataset will be a valuable resource for the community, aiding in the safe deployment of LLMs.

Language Models Resist Alignment

Large language models (LLMs) may exhibit undesirable behaviors. Recent efforts have focused on aligning these models to prevent harmful generation. Despite these efforts, studies have shown that even a well-conducted alignment process can be easily circumvented, whether intentionally or accidentally. Do alignment fine-tuning have robust effects on models, or are merely superficial? In this work, we answer this question through both theoretical and empirical means. Empirically, we demonstrate the elasticity of post-alignment models, i.e., the tendency to revert to the behavior distribution formed during the pre-training phase upon further fine-tuning. Using compression theory, we formally derive that such fine-tuning process \textit{disproportionately} undermines alignment compared to pre-training, potentially by orders of magnitude. We conduct experimental validations to confirm the presence of elasticity across models of varying types and sizes. Specifically, we find that model performance declines rapidly before reverting to the pre-training distribution, after which the rate of decline drops significantly. We further reveal that elasticity positively correlates with increased model size and the expansion of pre-training data. Our discovery signifies the importance of taming the inherent elasticity of LLMs, thereby overcoming the resistance of LLMs to alignment finetuning.

Rethinking Information Structures in RLHF: Reward Generalization from a Graph Theory Perspective

There is a trilemma in reinforcement learning from human feedback (RLHF): the incompatibility between highly diverse contexts, low labeling cost, and reliable alignment performance. Here we aim to mitigate such incompatibility through the design of dataset information structures during reward modeling, and meanwhile propose new, generalizable methods of analysis that have wider applications, including potentially shedding light on goal misgeneralization. Specifically, we first reexamine the RLHF process and propose a theoretical framework portraying it as an autoencoding process over text distributions. Our framework formalizes the RLHF objective of ensuring distributional consistency between human preference and large language model (LLM) behavior. Based on this framework, we introduce a new method to model generalization in the reward modeling stage of RLHF, the induced Bayesian network (IBN). Drawing from random graph theory and causal analysis, it enables empirically grounded derivation of generalization error bounds, a key improvement over classical methods of generalization analysis. An insight from our analysis is the superiority of the tree-based information structure in reward modeling, compared to chain-based baselines in conventional RLHF methods. We derive that in complex contexts with limited data, the tree-based reward model (RM) induces up to $\Theta(\log n/\log\log n)$ times less variance than chain-based RM where $n$ is the dataset size. As validation, we demonstrate that on three NLP tasks, the tree-based RM achieves 65% win rate on average against chain-based baselines. Looking ahead, we hope to extend the IBN analysis to help understand the phenomenon of goal misgeneralization.

Aligner: Achieving Efficient Alignment through Weak-to-Strong Correction

Efforts to align Large Language Models (LLMs) are mainly conducted via Reinforcement Learning from Human Feedback (RLHF) methods. However, RLHF encounters major challenges including training reward models, actor-critic engineering, and importantly, it requires access to LLM parameters. Here we introduce Aligner, a new efficient alignment paradigm that bypasses the whole RLHF process by learning the correctional residuals between the aligned and the unaligned answers. Our Aligner offers several key advantages. Firstly, it is an autoregressive seq2seq model that is trained on the query-answer-correction dataset via supervised learning; this offers a parameter-efficient alignment solution with minimal resources. Secondly, the Aligner facilitates weak-to-strong generalization; finetuning large pretrained models by Aligner's supervisory signals demonstrates strong performance boost. Thirdly, Aligner functions as a model-agnostic plug-and-play module, allowing for its direct application on different open-source and API-based models. Remarkably, Aligner-7B improves 11 different LLMs by 21.9% in helpfulness and 23.8% in harmlessness on average (GPT-4 by 17.5% and 26.9%). When finetuning (strong) Llama2-70B with (weak) Aligner-13B's supervision, we can improve Llama2 by 8.2% in helpfulness and 61.6% in harmlessness. See our dataset and code at https://aligner2024.github.io

AI Alignment: A Comprehensive Survey

AI alignment aims to make AI systems behave in line with human intentions and values. As AI systems grow more capable, the potential large-scale risks associated with misaligned AI systems become salient. Hundreds of AI experts and public figures have expressed concerns about AI risks, arguing that "mitigating the risk of extinction from AI should be a global priority, alongside other societal-scale risks such as pandemics and nuclear war". To provide a comprehensive and up-to-date overview of the alignment field, in this survey paper, we delve into the core concepts, methodology, and practice of alignment. We identify the RICE principles as the key objectives of AI alignment: Robustness, Interpretability, Controllability, and Ethicality. Guided by these four principles, we outline the landscape of current alignment research and decompose them into two key components: forward alignment and backward alignment. The former aims to make AI systems aligned via alignment training, while the latter aims to gain evidence about the systems' alignment and govern them appropriately to avoid exacerbating misalignment risks. Forward alignment and backward alignment form a recurrent process where the alignment of AI systems from the forward process is verified in the backward process, meanwhile providing updated objectives for forward alignment in the next round. On forward alignment, we discuss learning from feedback and learning under distribution shift. On backward alignment, we discuss assurance techniques and governance practices that apply to every stage of AI systems' lifecycle. We also release and continually update the website (www.alignmentsurvey.com) which features tutorials, collections of papers, blog posts, and other resources.

Safety-Gymnasium: A Unified Safe Reinforcement Learning Benchmark

Artificial intelligence (AI) systems possess significant potential to drive societal progress. However, their deployment often faces obstacles due to substantial safety concerns. Safe reinforcement learning (SafeRL) emerges as a solution to optimize policies while simultaneously adhering to multiple constraints, thereby addressing the challenge of integrating reinforcement learning in safety-critical scenarios. In this paper, we present an environment suite called Safety-Gymnasium, which encompasses safety-critical tasks in both single and multi-agent scenarios, accepting vector and vision-only input. Additionally, we offer a library of algorithms named Safe Policy Optimization (SafePO), comprising 16 state-of-the-art SafeRL algorithms. This comprehensive library can serve as a validation tool for the research community. By introducing this benchmark, we aim to facilitate the evaluation and comparison of safety performance, thus fostering the development of reinforcement learning for safer, more reliable, and responsible real-world applications. The website of this project can be accessed at https://sites.google.com/view/safety-gymnasium.

Safe RLHF: Safe Reinforcement Learning from Human Feedback

With the development of large language models (LLMs), striking a balance between the performance and safety of AI systems has never been more critical. However, the inherent tension between the objectives of helpfulness and harmlessness presents a significant challenge during LLM training. To address this issue, we propose Safe Reinforcement Learning from Human Feedback (Safe RLHF), a novel algorithm for human value alignment. Safe RLHF explicitly decouples human preferences regarding helpfulness and harmlessness, effectively avoiding the crowdworkers' confusion about the tension and allowing us to train separate reward and cost models. We formalize the safety concern of LLMs as an optimization task of maximizing the reward function while satisfying specified cost constraints. Leveraging the Lagrangian method to solve this constrained problem, Safe RLHF dynamically adjusts the balance between the two objectives during fine-tuning. Through a three-round fine-tuning using Safe RLHF, we demonstrate a superior ability to mitigate harmful responses while enhancing model performance compared to existing value-aligned algorithms. Experimentally, we fine-tuned the Alpaca-7B using Safe RLHF and aligned it with collected human preferences, significantly improving its helpfulness and harmlessness according to human evaluations.