The alignment of large language models (LLMs) is crucial for generating helpful and harmless content. Existing approaches leverage preference-based human feedback data to learn the reward function and align the LLM with the feedback data. However, these approaches focus on modeling the reward difference between the chosen and rejected demonstrations, rather than directly modeling the true reward from each demonstration. Moreover, these approaches assume that the reward is only obtained at the end of the sentence, which overlooks the modeling of intermediate rewards. These issues lead to insufficient use of training signals in the feedback data, limiting the representation and generalization ability of the reward and potentially resulting in reward hacking. In this paper, we formulate LLM alignment as a Bayesian Inverse Reinforcement Learning (BIRL) problem and propose a novel training objective, Approximated Variational Alignment (AVA), to perform LLM alignment through Approximated Variational Reward Imitation Learning (AVRIL). The BIRL formulation facilitates intermediate reward modeling and direct reward modeling on each single demonstration, which enhances the utilization of training signals in the feedback data. Experiments show that AVA outperforms existing LLM alignment approaches in reward modeling, RL fine-tuning, and direct optimization.

The goal of Bayesian inverse reinforcement learning (IRL) is recovering a posterior distribution over reward functions using a set of demonstrations from an expert optimizing for a reward unknown to the learner. The resulting posterior over rewards can then be used to synthesize an apprentice policy that performs well on the same or a similar task. A key challenge in Bayesian IRL is bridging the computational gap between the hypothesis space of possible rewards and the likelihood, often defined in terms of Q values: vanilla Bayesian IRL needs to solve the costly forward planning problem - going from rewards to the Q values - at every step of the algorithm, which may need to be done thousands of times. We propose to solve this by a simple change: instead of focusing on primarily sampling in the space of rewards, we can focus on primarily working in the space of Q-values, since the computation required to go from Q-values to reward is radically cheaper. Furthermore, this reversion of the computation makes it easy to compute the gradient allowing efficient sampling using Hamiltonian Monte Carlo. We propose ValueWalk - a new Markov chain Monte Carlo method based on this insight - and illustrate its advantages on several tasks.

Inverse reinforcement learning (IRL) is the problem of inferring a reward function from expert behavior. There are several approaches to IRL, but most are designed to learn a Markovian reward. However, a reward function might be non-Markovian, depending on more than just the current state, such as a reward machine (RM). Although there has been recent work on inferring RMs, it assumes access to the reward signal, absent in IRL. We propose a Bayesian IRL (BIRL) framework for inferring RMs directly from expert behavior, requiring significant changes to the standard framework. We define a new reward space, adapt the expert demonstration to include history, show how to compute the reward posterior, and propose a novel modification to simulated annealing to maximize this posterior. We demonstrate that our method performs well when optimizing according to its inferred reward and compares favorably to an existing method that learns exclusively binary non-Markovian rewards.

We present a nonparametric Bayesian approach to inverse reinforcement learning (IRL) for multiple reward functions. Most previous IRL algorithms assume that the behaviour data is obtained from an agent who is optimizing a single reward function, but this assumption is hard to guarantee in practice. Our approach is based on integrating the Dirichlet process mixture model into Bayesian IRL. We provide an efficient Metropolis-Hastings sampling algorithm utilizing the gradient of the posterior to estimate the underlying reward functions, and demonstrate that our approach outperforms previous ones via experiments on a number of problem domains.

Optimizing for humans' latent preferences remains a grand challenge in route recommendation. Prior research has provided increasingly general techniques based on inverse reinforcement learning (IRL), yet no approach has been successfully scaled to world-sized routing problems with hundreds of millions of states and demonstration trajectories. In this paper, we provide methods for scaling IRL using graph compression, spatial parallelization, and problem initialization based on dominant eigenvectors. We revisit classic algorithms and study them in a large-scale setting, and make the key observation that there exists a trade-off between the use of cheap, deterministic planners and expensive yet robust stochastic policies. We leverage this insight in Receding Horizon Inverse Planning (RHIP), a new generalization of classic IRL algorithms that provides fine-grained control over performance trade-offs via its planning horizon. Our contributions culminate in a policy that achieves a 16-24% improvement in global route quality, and to the best of our knowledge, represents the largest instance of IRL in a real-world setting to date. Benchmark results show critical benefits to more sustainable modes of transportation, where factors beyond journey time play a substantial role. We conclude by conducting an ablation study of key components, presenting negative results from alternative eigenvalue solvers, and identifying opportunities to further improve scalability via IRL-specific batching strategies.

The difficulty in inverse reinforcement learning (IRL) arises in choosing the best reward function since there are typically an infinite number of reward functions that yield the given behaviour data as optimal. Using a Bayesian framework, we address this challenge by using the maximum a posteriori (MAP) estimation for the reward function, and show that most of the previous IRL algorithms can be modeled into our framework. We also present a gradient method for the MAP estimation based on the (sub)differentiability of the posterior distribution. We show the effectiveness of our approach by comparing the performance of the proposed method to those of the previous algorithms.

The difficulty in inverse reinforcement learning (IRL) arises in choosing the best reward function since there are typically an infinite number of reward functions that yield the given behaviour data as optimal. Using a Bayesian framework, we address this challenge by using the maximum a posteriori (MAP) estimation for the reward function, and show that most of the previous IRL algorithms can be modeled into our framework. We also present a gradient method for the MAP estimation based on the (sub)differentiability of the posterior distribution. We show the effectiveness of our approach by comparing the performance of the proposed method to those of the previous algorithms. Papers published at the Neural Information Processing Systems Conference.

We present a nonparametric Bayesian approach to inverse reinforcement learning (IRL) for multiple reward functions. Most previous IRL algorithms assume that the behaviour data is obtained from an agent who is optimizing a single reward function, but this assumption is hard to be met in practice. Our approach is based on integrating the Dirichlet process mixture model into Bayesian IRL. We provide an efficient Metropolis-Hastings sampling algorithm utilizing the gradient of the posterior to estimate the underlying reward functions, and demonstrate that our approach outperforms the previous ones via experiments on a number of problem domains.