NeurIPS 2024 Papers

Recent studies developed jailbreaking attacks, which construct jailbreaking prompts to "fool" LLMs into responding to harmful questions. Early-stage jailbreaking attacks require access to model internals or significant human efforts. More advanced attacks utilize genetic algorithms for automatic and black-box attacks. However, the random nature of genetic algorithms significantly limits the effectiveness of these attacks. In this paper, we propose RLbreaker, a black-box jailbreaking attack driven by deep reinforcement learning (DRL). We model jailbreaking as a search problem and design an RL agent to guide the search, which is more effective and has less randomness than stochastic search, such as genetic algorithms. Specifically, we design a customized DRL system for the jailbreaking problem, including a novel reward function and a customized proximal policy optimization (PPO) algorithm. Through extensive experiments, we demonstrate that RLbreaker is much more effective than existing jailbreaking attacks against six state-of-the-art (SOTA) LLMs. We also show that RLbreaker is robust against three SOTA defenses and its trained agents can transfer across different LLMs. We further validate the key design choices of RLbreaker via a comprehensive ablation study.

Deep reinforcement learning aims to learn deep neural network policies to solve large-scale decision-making problems. However, approximating policies using deep neural networks makes it difficult to interpret the learned decision-making process. To address this issue, prior works (Trivedi et al., 2021; Liu et al., 2023; Carvalho et al., 2024) proposed to use human-readable programs as policies to increase the interpretability of the decision-making pipeline. Nevertheless, programmatic policies generated by these methods struggle to effectively solve long and repetitive RL tasks and cannot generalize to even longer horizons during testing. To solve these problems, we propose the Hierarchical Programmatic Option framework (HIPO), which aims to solve long and repetitive RL problems with human-readable programs as options (low-level policies). Specifically, we propose a method that retrieves a set of effective, diverse, and compatible programs as options. Then, we learn a high-level policy to effectively reuse these programmatic options to solve reoccurring subtasks. Our proposed framework outperforms programmatic RL and deep RL baselines on various tasks. Ablation studies justify the effectiveness of our proposed search algorithm for retrieving a set of programmatic options.

Learning from AI feedback (LAIF) is a popular paradigm for improving the instruction-following abilities of powerful pre-trained language models. LAIF first performs supervised fine-tuning (SFT) using demonstrations from a teacher model and then further fine-tunes the model with reinforcement learning (RL) or direct preference optimization (DPO), using feedback from a critic model. While recent popular open-source models have demonstrated substantial improvements in performance from the RL step, in this paper we question whether the complexity of this RL step is truly warranted for AI feedback. We show that the improvements of the RL step are virtually entirely due to the widespread practice of using a weaker teacher model (e.g. GPT-3.5) for SFT data collection than the critic (e.g., GPT-4) used for AI feedback generation. Specifically, we show that simple supervised fine-tuning with GPT-4 as the teacher outperforms existing LAIF pipelines. More generally, we find that the gains from LAIF vary substantially across base model families, test-time evaluation protocols, and critic models. Finally, we provide a mechanistic explanation for when SFT may outperform the full two-step LAIF pipeline as well as suggestions for making LAIF maximally useful in practice.

Adding additional guidance to pretrained diffusion models has become an increasingly popular research area, with extensive applications in computer vision, reinforcement learning, and AI for science. Recently, several studies have proposed training-free loss-based guidance by using off-the-shelf networks pretrained on clean images. This approach enables zero-shot conditional generation for universal control formats, which appears to offer a free lunch in diffusion guidance. In this paper, we aim to develop a deeper understanding of training-free guidance, as well as overcome its limitations. We offer a theoretical analysis that supports training-free guidance from the perspective of optimization, distinguishing it from classifier-based (or classifier-free) guidance. To elucidate their drawbacks, we theoretically demonstrate that training-free guidance is more susceptible to misaligned gradients and exhibits slower convergence rates compared to classifier guidance. We then introduce a collection of techniques designed to overcome the limitations, accompanied by theoretical rationale and empirical evidence. Our experiments in image and motion generation confirm the efficacy of these techniques.

Generating ligand molecules for specific protein targets, known as structure-based drug design, is a fundamental problem in therapeutics development and biological discovery. Recently, target-aware generative models, especially diffusion models, have shown great promise in modeling protein-ligand interactions and generating candidate drugs. However, existing models primarily focus on learning the chemical distribution of all drug candidates, which lacks effective steerability on the chemical quality of model generations. In this paper, we propose a novel and general alignment framework to align pretrained target diffusion models with preferred functional properties, named AliDiff. AliDiff shifts the target-conditioned chemical distribution towards regions with higher binding affinity and structural rationality, specified by user-defined reward functions, via the preference optimization approach. To avoid the overfitting problem in common preference optimization objectives, we further develop an improved Exact Energy Preference Optimization method to yield an exact and efficient alignment of the diffusion models, and provide the closed-form expression for the converged distribution. Empirical studies on the CrossDocked2020 benchmark show that AliDiff can generate molecules with state-of-the-art binding energies with up to -7.07 Avg. Vina Score, while maintaining strong molecular properties. Code is available at https://github.com/MinkaiXu/AliDiff.

We present Preference Flow Matching (PFM), a new framework for preference alignment that streamlines the integration of preferences into an arbitrary class of pre-trained models. Existing alignment methods require fine-tuning pre-trained models, which presents challenges such as scalability, inefficiency, and the need for model modifications, especially with black-box APIs like GPT-4. In contrast, PFM utilizes flow matching techniques to directly learn from preference data, thereby reducing the dependency on extensive fine-tuning of pre-trained models. By leveraging flow-based models, PFM transforms less preferred data into preferred outcomes, and effectively aligns model outputs with human preferences without relying on explicit or implicit reward function estimation, thus avoiding common issues like overfitting in reward models. We provide theoretical insights that support our method’s alignment with standard preference alignment objectives. Experimental results indicate the practical effectiveness of our method, offering a new direction in aligning a pre-trained model to preference. Our code is available at https://github.com/jadehaus/preference-flow-matching.

One important property of DIstribution Correction Estimation (DICE) methods is that the solution is the optimal stationary distribution ratio between the optimized and data collection policy. In this work, we show that DICE-based methods can be viewed as a transformation from the behavior distribution to the optimal policy distribution. Based on this, we propose a novel approach, Diffusion-DICE, that directly performs this transformation using diffusion models. We find that the optimal policy's score function can be decomposed into two terms: the behavior policy's score function and the gradient of a guidance term which depends on the optimal distribution ratio. The first term can be obtained from a diffusion model trained on the dataset and we propose an in-sample learning objective to learn the second term. Due to the multi-modality contained in the optimal policy distribution, the transformation in Diffusion-DICE may guide towards those local-optimal modes. We thus generate a few candidate actions and carefully select from them to achieve global-optimum. Different from all other diffusion-based offline RL methods, the \textit{guide-then-select} paradigm in Diffusion-DICE only uses in-sample actions for training and brings minimal error exploitation in the value function. We use a didatic toycase example to show how previous diffusion-based methods fail to generate optimal actions due to leveraging these errors and how Diffusion-DICE successfully avoid that. We then conduct extensive experiments on benchmark datasets to show the strong performance of Diffusion-DICE.

Reinforcement Learning (RL) controllers have demonstrated remarkable performance in complex robot control tasks. However, the presence of reality gap often leads to poor performance when deploying policies trained in simulation directly onto real robots. Previous sim-to-real algorithms like Domain Randomization (DR) requires domain-specific expertise and suffers from issues such as reduced control performance and high training costs. In this work, we introduce Evolutionary Adversarial Simulator Identification (EASI), a novel approach that combines Generative Adversarial Network (GAN) and Evolutionary Strategy (ES) to address sim-to-real challenges. Specifically, we consider the problem of sim-to-real as a search problem, where ES acts as a generator in adversarial competition with a neural network discriminator, aiming to find physical parameter distributions that make the state transitions between simulation and reality as similar as possible. The discriminator serves as the fitness function, guiding the evolution of the physical parameter distributions. EASI features simplicity, low cost, and high fidelity, enabling the construction of a more realistic simulator with minimal requirements for real-world data, thus aiding in transferring simulated-trained policies to the real world. We demonstrate the performance of EASI in both sim-to-sim and sim-to-real tasks, showing superior performance compared to existing sim-to-real algorithms.

We consider the robust routing problem with uncertain travel times under the min-max regret criterion, which represents an extended and robust version of the classic traveling salesman problem (TSP) and vehicle routing problem (VRP). The general budget uncertainty set is employed to capture the uncertainty, which provides the capability to control the conservatism of obtained solutions and covers the commonly used interval uncertainty set as a special case. The goal is to obtain a robust solution that minimizes the maximum deviation from the optimal routing time in the worst-case scenario. Given the significant advancements and broad applications of neural combinatorial optimization methods in recent years, we present our initial attempt to combine neural approaches for solving this problem. We propose a dual multi-head cross attention mechanism to extract problem features represented by the inputted uncertainty sets. To tackle the built-in maximization problem, we derive the regret value by invoking a pre-trained model, subsequently utilizing it as the reward during the model training. Our experimental results on the robust TSP and VRP demonstrate the efficacy of our neural combinatorial optimization method, showcasing its ability to efficiently handle the robust routing problem of various sizes within a shorter time compared with alternative heuristic approaches.

When humans need to learn a new skill, we can acquire knowledge through written books, including textbooks, tutorials, etc. However, current research for decision-making, like reinforcement learning (RL), has primarily required numerous real interactions with the target environment to learn a skill, while failing to utilize the existing knowledge already summarized in the text. The success of Large Language Models (LLMs) sheds light on utilizing such knowledge behind the books. In this paper, we discuss a new policy learning problem called Policy Learning from tutorial Books (PLfB) upon the shoulders of LLMs’ systems, which aims to leverage rich resources such as tutorial books to derive a policy network. Inspired by how humans learn from books, we solve the problem via a three-stage framework: Understanding, Rehearsing, and Introspecting (URI). In particular, it first rehearses decision-making trajectories based on the derived knowledge after understanding the books, then introspects in the imaginary dataset to distill a policy network. We build two benchmarks for PLfB~based on Tic-Tac-Toe and Football games. In experiment, URI's policy achieves at least 44% net win rate against GPT-based agents without any real data; In Football game, which is a complex scenario, URI's policy beat the built-in AIs with a 37% while using GPT-based agent can only achieve a 6\% winning rate. The project page: https://plfb-football.github.io.

Modern deep policy gradient methods achieve effective performance on simulated robotic tasks, but they all require large replay buffers or expensive batch updates, or both, making them incompatible for real systems with resource-limited computers. We show that these methods fail catastrophically when limited to small replay buffers or during *incremental learning*, where updates only use the most recent sample without batch updates or a replay buffer. We propose a novel incremental deep policy gradient method --- *Action Value Gradient (AVG)* and a set of normalization and scaling techniques to address the challenges of instability in incremental learning. On robotic simulation benchmarks, we show that AVG is the only incremental method that learns effectively, often achieving final performance comparable to batch policy gradient methods. This advancement enabled us to show for the first time effective deep reinforcement learning with real robots using only incremental updates, employing a robotic manipulator and a mobile robot.

Multi-Agent Reinforcement Learning (MARL) methods have shown promise in enabling agents to learn a shared communication protocol from scratch and accomplish challenging team tasks. However, the learned language is usually not interpretable to humans or other agents not co-trained together, limiting its applicability in ad-hoc teamwork scenarios. In this work, we propose a novel computational pipeline that aligns the communication space between MARL agents with an embedding space of human natural language by grounding agent communications on synthetic data generated by embodied Large Language Models (LLMs) in interactive teamwork scenarios. Our results demonstrate that introducing language grounding not only maintains task performance but also accelerates the emergence of communication. Furthermore, the learned communication protocols exhibit zero-shot generalization capabilities in ad-hoc teamwork scenarios with unseen teammates and novel task states. This work presents a significant step toward enabling effective communication and collaboration between artificial agents and humans in real-world teamwork settings.

Federated reinforcement learning (RL) enables collaborative decision making of multiple distributed agents without sharing local data trajectories. In this work, we consider a multi-task setting, in which each agent has its own private reward function corresponding to different tasks, while sharing the same transition kernel of the environment. Focusing on infinite-horizon Markov decision processes, the goal is to learn a globally optimal policy that maximizes the sum of the discounted total rewards of all the agents in a decentralized manner, where each agent only communicates with its neighbors over some prescribed graph topology. We develop federated vanilla and entropy-regularized natural policy gradient (NPG) methods in the tabular setting under softmax parameterization, where gradient tracking is applied to estimate the global Q-function to mitigate the impact of imperfect information sharing. We establish non-asymptotic global convergence guarantees under exact policy evaluation, where the rates are nearly independent of the size of the state-action space and illuminate the impacts of network size and connectivity. To the best of our knowledge, this is the first time that global convergence is established for federated multi-task RL using policy optimization. We further go beyond the tabular setting by proposing a federated natural actor critic (NAC) method for multi-task RL with function approximation, and establish its finite-time sample complexity taking the errors of function approximation into account.

A central piece in enabling intelligent agentic behavior in foundation models is to make them capable of introspecting upon their behavior, reasoning, and correcting their mistakes as more computation or interaction is available. Even the strongest proprietary large language models (LLMs) do not quite exhibit the ability of continually improving their responses sequentially. In this paper, we develop $\textbf{RISE:}$ $\textbf{R}$ecursive $\textbf{I}$ntro$\textbf{S}$p$\textbf{E}$ction, an approach for fine-tuning LLMs to introduce this capability, despite prior work hypothesizing that this capability may not be possible to attain. Our approach prescribes an iterative fine-tuning procedure, which attempts to teach the model how to alter its response after having executed previously unsuccessful attempts to solve a hard test-time problem, with optionally additional environment feedback. RISE poses fine-tuning for a single-turn prompt as solving a multi-turn Markov decision process (MDP), where the initial state is the prompt. Inspired by principles in online imitation and offline reinforcement learning, we propose strategies for multi-turn data collection and training so as to imbue an LLM with the capability to recursively detect and correct its previous mistakes in subsequent iterations. Our experiments show that RISE enables Llama2, Llama3, and Mistral models to improve themselves with more turns on reasoning tasks, outperforming several single-turn strategies given an equal amount of inference-time computation. We also find that RISE scales well, often attaining larger benefits with more capable models, without disrupting one-turn abilities as a result of expressing more complex distributions.

Efficiently solving unseen tasks remains a challenge in reinforcement learning (RL), especially for long-horizon tasks composed of multiple subtasks. Pre-training policies from task-agnostic datasets has emerged as a promising approach, yet existing methods still necessitate substantial interactions via RL to learn new tasks. We introduce MGPO, a method that leverages the power of Transformer-based policies to model sequences of goals, enabling efficient online adaptation through prompt optimization. In its pre-training phase, MGPO utilizes hindsight multi-goal relabeling and behavior cloning. This combination equips the policy to model diverse long-horizon behaviors that align with varying goal sequences. During online adaptation, the goal sequence, conceptualized as a prompt, is optimized to improve task performance. We adopt a multi-armed bandit framework for this process, enhancing prompt selection based on the returns from online trajectories. Our experiments across various environments demonstrate that MGPO holds substantial advantages in sample efficiency, online adaptation performance, robustness, and interpretability compared with existing methods.

In environments with delayed observation, state augmentation by including actions within the delay window is adopted to retrieve Markovian property to enable reinforcement learning (RL). Whereas, state-of-the-art (SOTA) RL techniques with Temporal-Difference (TD) learning frameworks commonly suffer from learning inefficiency, due to the significant expansion of the augmented state space with the delay. To improve the learning efficiency without sacrificing performance, this work novelly introduces Variational Delayed Policy Optimization (VDPO), reforming delayed RL as a variational inference problem. This problem is further modelled as a two-step iterative optimization problem, where the first step is TD learning in the delay-free environment with a small state space, and the second step is behaviour cloning which can be addressed much more efficiently than TD learning. We not only provide a theoretical analysis of VDPO in terms of sample complexity and performance, but also empirically demonstrate that VDPO can achieve consistent performance with SOTA methods, with a significant enhancement of sample efficiency (approximately 50\% less amount of samples) in the MuJoCo benchmark.

Policy network parameter sharing is a commonly used technique in advanced deep multi-agent reinforcement learning (MARL) algorithms to improve learning efficiency by reducing the number of policy parameters and sharing experiences among agents. Nevertheless, agents that share the policy parameters tend to learn similar behaviors. To encourage multi-agent diversity, prior works typically maximize the mutual information between trajectories and agent identities using variational inference. However, this category of methods easily leads to inefficient exploration due to limited trajectory visitations. To resolve this limitation, inspired by the learning of pre-trained models, in this paper, we propose a novel Contrastive Trajectory Representation (CTR) method based on learning distinguishable trajectory representations to encourage multi-agent diversity. Specifically, CTR maps the trajectory of an agent into a latent trajectory representation space by an encoder and an autoregressive model. To achieve the distinguishability among trajectory representations of different agents, we introduce contrastive learning to maximize the mutual information between the trajectory representations and learnable identity representations of different agents. We implement CTR on top of QMIX and evaluate its performance in various cooperative multi-agent tasks. The empirical results demonstrate that our proposed CTR yields significant performance improvement over the state-of-the-art methods.

To solve control problems via model-based reasoning or planning, an agent needs to know how its actions affect the state of the world. The actions an agent has at its disposal often change the state of the environment in systematic ways. However, existing techniques for world modelling do not guarantee that the effect of actions are represented in such systematic ways. We introduce the Parsimonious Latent Space Model (PLSM), a world model that regularizes the latent dynamics to make the effect of the agent's actions more predictable. Our approach minimizes the mutual information between latent states and the change that an action produces in the agent's latent state, in turn minimizing the dependence the state has on the dynamics. This makes the world model softly state-invariant. We combine PLSM with different model classes used for i) future latent state prediction, ii) planning, and iii) model-free reinforcement learning. We find that our regularization improves accuracy, generalization, and performance in downstream tasks, highlighting the importance of systematic treatment of actions in world models.

The empirical success of distributional reinforcement learning (RL) highly relies on the choice of distribution divergence equipped with an appropriate distribution representation. In this paper, we propose \textit{Sinkhorn distributional RL (SinkhornDRL)}, which leverages Sinkhorn divergence—a regularized Wasserstein loss—to minimize the difference between current and target Bellman return distributions. Theoretically, we prove the contraction properties of SinkhornDRL, aligning with the interpolation nature of Sinkhorn divergence between Wasserstein distance and Maximum Mean Discrepancy (MMD). The introduced SinkhornDRL enriches the family of distributional RL algorithms, contributing to interpreting the algorithm behaviors compared with existing approaches by our investigation into their relationships. Empirically, we show that SinkhornDRL consistently outperforms or matches existing algorithms on the Atari games suite and particularly stands out in the multi-dimensional reward setting. \thanks{Code is available in \url{https://github.com/datake/SinkhornDistRL}.}.