Abstract: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.
Abstract: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.
Abstract: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.
Abstract:The existing fine-tuning paradigm for language models is brittle in knowledge editing scenarios, where the model must incorporate new information without extensive retraining. This brittleness often results in overfitting, reduced performance, and unnatural language generation. To address this, we propose Consistent In-Context Editing (ICE), a novel approach that leverages the model's in-context learning capability to tune toward a contextual distribution rather than a one-hot target. ICE introduces a straightforward optimization framework that includes both a target and a procedure, enhancing the robustness and effectiveness of gradient-based tuning methods. We provide analytical insights into ICE across four critical aspects of knowledge editing: accuracy, locality, generalization, and linguistic quality, showing its advantages. Experimental results across four datasets confirm the effectiveness of ICE and demonstrate its potential for continual editing, ensuring that updated information is incorporated while preserving the integrity of the model.
Abstract:Federated learning (FL) enables multiple parties to collaboratively fine-tune an large language model (LLM) without the need of direct data sharing. Ideally, by training on decentralized data that is aligned with human preferences and safety principles, federated instruction tuning can result in an LLM that could behave in a helpful and safe manner. In this paper, we for the first time reveal the vulnerability of safety alignment in FedIT by proposing a simple, stealthy, yet effective safety attack method. Specifically, the malicious clients could automatically generate attack data without involving manual efforts and attack the FedIT system by training their local LLMs on such attack data. Unfortunately, this proposed safety attack not only can compromise the safety alignment of LLM trained via FedIT, but also can not be effectively defended against by many existing FL defense methods. Targeting this, we further propose a post-hoc defense method, which could rely on a fully automated pipeline: generation of defense data and further fine-tuning of the LLM. Extensive experiments show that our safety attack method can significantly compromise the LLM's safety alignment (e.g., reduce safety rate by 70\%), which can not be effectively defended by existing defense methods (at most 4\% absolute improvement), while our safety defense method can significantly enhance the attacked LLM's safety alignment (at most 69\% absolute improvement).
Abstract:Despite the recent successes of multi-agent reinforcement learning (MARL) algorithms, efficiently adapting to co-players in mixed-motive environments remains a significant challenge. One feasible approach is to hierarchically model co-players' behavior based on inferring their characteristics. However, these methods often encounter difficulties in efficient reasoning and utilization of inferred information. To address these issues, we propose Hierarchical Opponent modeling and Planning (HOP), a novel multi-agent decision-making algorithm that enables few-shot adaptation to unseen policies in mixed-motive environments. HOP is hierarchically composed of two modules: an opponent modeling module that infers others' goals and learns corresponding goal-conditioned policies, and a planning module that employs Monte Carlo Tree Search (MCTS) to identify the best response. Our approach improves efficiency by updating beliefs about others' goals both across and within episodes and by using information from the opponent modeling module to guide planning. Experimental results demonstrate that in mixed-motive environments, HOP exhibits superior few-shot adaptation capabilities when interacting with various unseen agents, and excels in self-play scenarios. Furthermore, the emergence of social intelligence during our experiments underscores the potential of our approach in complex multi-agent environments.
Abstract: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.
Abstract:A popular approach for solving zero-sum games is to maintain populations of policies to approximate the Nash Equilibrium (NE). Previous studies have shown that Policy Space Response Oracle (PSRO) algorithm is an effective multi-agent reinforcement learning framework for solving such games. However, repeatedly training new policies from scratch to approximate Best Response (BR) to opponents' mixed policies at each iteration is both inefficient and costly. While some PSRO variants initialize a new policy by inheriting from past BR policies, this approach limits the exploration of new policies, especially against challenging opponents. To address this issue, we propose Fusion-PSRO, which employs policy fusion to initialize policies for better approximation to BR. By selecting high-quality base policies from meta-NE, policy fusion fuses the base policies into a new policy through model averaging. This approach allows the initialized policies to incorporate multiple expert policies, making it easier to handle difficult opponents compared to inheriting from past BR policies or initializing from scratch. Moreover, our method only modifies the policy initialization phase, allowing its application to nearly all PSRO variants without additional training overhead. Our experiments on non-transitive matrix games, Leduc Poker, and the more complex Liars Dice demonstrate that Fusion-PSRO enhances the performance of nearly all PSRO variants, achieving lower exploitability.
Abstract:With recent advancements in large language models (LLMs), alignment has emerged as an effective technique for keeping LLMs consensus with human intent. Current methods primarily involve direct training through Supervised Fine-tuning (SFT) or Reinforcement Learning from Human Feedback (RLHF), both of which require substantial computational resources and extensive ground truth data. This paper explores an efficient method for aligning black-box large models using smaller models, introducing a model-agnostic and lightweight Bayesian Persuasion Alignment framework. We formalize this problem as an optimization of the signaling strategy from the small model's perspective. In the persuasion process, the small model (Advisor) observes the information item (i.e., state) and persuades large models (Receiver) to elicit improved responses. The Receiver then generates a response based on the input, the signal from the Advisor, and its updated belief about the information item. Through training using our framework, we demonstrate that the Advisor can significantly enhance the performance of various Receivers across a range of tasks. We theoretically analyze our persuasion framework and provide an upper bound on the Advisor's regret, confirming its effectiveness in learning the optimal signaling strategy. Our Empirical results demonstrates that GPT-2 can significantly improve the performance of various models, achieving an average enhancement of 16.1% in mathematical reasoning ability and 13.7% in code generation. We hope our work can provide an initial step toward rethinking the alignment framework from the Bayesian Persuasion perspective.
Abstract:Traditional approaches in physics-based motion generation, centered around imitation learning and reward shaping, often struggle to adapt to new scenarios. To tackle this limitation, we propose AnySkill, a novel hierarchical method that learns physically plausible interactions following open-vocabulary instructions. Our approach begins by developing a set of atomic actions via a low-level controller trained via imitation learning. Upon receiving an open-vocabulary textual instruction, AnySkill employs a high-level policy that selects and integrates these atomic actions to maximize the CLIP similarity between the agent's rendered images and the text. An important feature of our method is the use of image-based rewards for the high-level policy, which allows the agent to learn interactions with objects without manual reward engineering. We demonstrate AnySkill's capability to generate realistic and natural motion sequences in response to unseen instructions of varying lengths, marking it the first method capable of open-vocabulary physical skill learning for interactive humanoid agents.