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.

Llemma: An Open Language Model For Mathematics

We present Llemma, a large language model for mathematics. We continue pretraining Code Llama on the Proof-Pile-2, a mixture of scientific papers, web data containing mathematics, and mathematical code, yielding Llemma. On the MATH benchmark Llemma outperforms all known open base models, as well as the unreleased Minerva model suite on an equi-parameter basis. Moreover, Llemma is capable of tool use and formal theorem proving without any further finetuning. We openly release all artifacts, including 7 billion and 34 billion parameter models, the Proof-Pile-2, and code to replicate our experiments.

Confronting Reward Model Overoptimization with Constrained RLHF

Large language models are typically aligned with human preferences by optimizing $\textit{reward models}$ (RMs) fitted to human feedback. However, human preferences are multi-faceted, and it is increasingly common to derive reward from a composition of simpler reward models which each capture a different aspect of language quality. This itself presents a challenge, as it is difficult to appropriately weight these component RMs when combining them. Compounding this difficulty, because any RM is only a proxy for human evaluation, this process is vulnerable to $\textit{overoptimization}$, wherein past a certain point, accumulating higher reward is associated with worse human ratings. In this paper, we perform, to our knowledge, the first study on overoptimization in composite RMs, showing that correlation between component RMs has a significant effect on the locations of these points. We then introduce an approach to solve this issue using constrained reinforcement learning as a means of preventing the agent from exceeding each RM's threshold of usefulness. Our method addresses the problem of weighting component RMs by learning dynamic weights, naturally expressed by Lagrange multipliers. As a result, each RM stays within the range at which it is an effective proxy, improving evaluation performance. Finally, we introduce an adaptive method using gradient-free optimization to identify and optimize towards these points during a single run.

Game-Theoretic Robust Reinforcement Learning Handles Temporally-Coupled Perturbations

Robust reinforcement learning (RL) seeks to train policies that can perform well under environment perturbations or adversarial attacks. Existing approaches typically assume that the space of possible perturbations remains the same across timesteps. However, in many settings, the space of possible perturbations at a given timestep depends on past perturbations. We formally introduce temporally-coupled perturbations, presenting a novel challenge for existing robust RL methods. To tackle this challenge, we propose GRAD, a novel game-theoretic approach that treats the temporally-coupled robust RL problem as a partially-observable two-player zero-sum game. By finding an approximate equilibrium in this game, GRAD ensures the agent's robustness against temporally-coupled perturbations. Empirical experiments on a variety of continuous control tasks demonstrate that our proposed approach exhibits significant robustness advantages compared to baselines against both standard and temporally-coupled attacks, in both state and action spaces.

Policy Space Diversity for Non-Transitive Games

Policy-Space Response Oracles (PSRO) is an influential algorithm framework for approximating a Nash Equilibrium (NE) in multi-agent non-transitive games. Many previous studies have been trying to promote policy diversity in PSRO. A major weakness in existing diversity metrics is that a more diverse (according to their diversity metrics) population does not necessarily mean (as we proved in the paper) a better approximation to a NE. To alleviate this problem, we propose a new diversity metric, the improvement of which guarantees a better approximation to a NE. Meanwhile, we develop a practical and well-justified method to optimize our diversity metric using only state-action samples. By incorporating our diversity regularization into the best response solving in PSRO, we obtain a new PSRO variant, Policy Space Diversity PSRO (PSD-PSRO). We present the convergence property of PSD-PSRO. Empirically, extensive experiments on various games demonstrate that PSD-PSRO is more effective in producing significantly less exploitable policies than state-of-the-art PSRO variants.

Language Models can Solve Computer Tasks

Agents capable of carrying out general tasks on a computer can improve efficiency and productivity by automating repetitive tasks and assisting in complex problem-solving. Ideally, such agents should be able to solve new computer tasks presented to them through natural language commands. However, previous approaches to this problem require large amounts of expert demonstrations and task-specific reward functions, both of which are impractical for new tasks. In this work, we show that a pre-trained large language model (LLM) agent can execute computer tasks guided by natural language using a simple prompting scheme where the agent recursively criticizes and improves its output (RCI). The RCI approach significantly outperforms existing LLM methods for automating computer tasks and surpasses supervised learning (SL) and reinforcement learning (RL) approaches on the MiniWoB++ benchmark. RCI is competitive with the state-of-the-art SL+RL method, using only a handful of demonstrations per task rather than tens of thousands, and without a task-specific reward function. Furthermore, we demonstrate RCI prompting's effectiveness in enhancing LLMs' reasoning abilities on a suite of natural language reasoning tasks, outperforming chain of thought (CoT) prompting. We find that RCI combined with CoT performs better than either separately.

Ensemble Value Functions for Efficient Exploration in Multi-Agent Reinforcement Learning

Cooperative multi-agent reinforcement learning (MARL) requires agents to explore to learn to cooperate. Existing value-based MARL algorithms commonly rely on random exploration, such as $\epsilon$-greedy, which is inefficient in discovering multi-agent cooperation. Additionally, the environment in MARL appears non-stationary to any individual agent due to the simultaneous training of other agents, leading to highly variant and thus unstable optimisation signals. In this work, we propose ensemble value functions for multi-agent exploration (EMAX), a general framework to extend any value-based MARL algorithm. EMAX trains ensembles of value functions for each agent to address the key challenges of exploration and non-stationarity: (1) The uncertainty of value estimates across the ensemble is used in a UCB policy to guide the exploration of agents to parts of the environment which require cooperation. (2) Average value estimates across the ensemble serve as target values. These targets exhibit lower variance compared to commonly applied target networks and we show that they lead to more stable gradients during the optimisation. We instantiate three value-based MARL algorithms with EMAX, independent DQN, VDN and QMIX, and evaluate them in 21 tasks across four environments. Using ensembles of five value functions, EMAX improves sample efficiency and final evaluation returns of these algorithms by 54%, 55%, and 844%, respectively, averaged all 21 tasks.

ASP: Learn a Universal Neural Solver!

Applying machine learning to combinatorial optimization problems has the potential to improve both efficiency and accuracy. However, existing learning-based solvers often struggle with generalization when faced with changes in problem distributions and scales. In this paper, we propose a new approach called ASP: Adaptive Staircase Policy Space Response Oracle to address these generalization issues and learn a universal neural solver. ASP consists of two components: Distributional Exploration, which enhances the solver's ability to handle unknown distributions using Policy Space Response Oracles, and Persistent Scale Adaption, which improves scalability through curriculum learning. We have tested ASP on several challenging COPs, including the traveling salesman problem, the vehicle routing problem, and the prize collecting TSP, as well as the real-world instances from TSPLib and CVRPLib. Our results show that even with the same model size and weak training signal, ASP can help neural solvers explore and adapt to unseen distributions and varying scales, achieving superior performance. In particular, compared with the same neural solvers under a standard training pipeline, ASP produces a remarkable decrease in terms of the optimality gap with 90.9% and 47.43% on generated instances and real-world instances for TSP, and a decrease of 19% and 45.57% for CVRP.