ICLR 2025 Papers

Broadly intelligent agents should form task-specific abstractions that selectively expose the essential elements of a task, while abstracting away the complexity of the raw sensorimotor space. In this work, we present Neuro-Symbolic Predicates, a first-order abstraction language that combines the strengths of symbolic and neural knowledge representations. We outline an online algorithm for inventing such predicates and learning abstract world models. We compare our approach to hierarchical reinforcement learning, vision-language model planning, and symbolic predicate invention approaches, on both in- and out-of-distribution tasks across five simulated robotic domains. Results show that our approach offers better sample complexity, stronger out-of-distribution generalization, and improved interpretability.

Reward shaping is a reinforcement learning technique that addresses the sparse-reward problem by providing frequent, informative feedback. We propose an efficient self-adaptive reward-shaping mechanism that uses success rates derived from historical experiences as shaped rewards. The success rates are sampled from Beta distributions, which evolve from uncertainty to reliability as data accumulates. Initially, shaped rewards are stochastic to encourage exploration, gradually becoming more certain to promote exploitation and maintain a natural balance between exploration and exploitation. We apply Kernel Density Estimation (KDE) with Random Fourier Features (RFF) to derive Beta distributions, providing a computationally efficient solution for continuous and high-dimensional state spaces. Our method, validated on tasks with extremely sparse rewards, improves sample efficiency and convergence stability over relevant baselines.

Artificially intelligent agents are increasingly being integrated into human decision-making: from large language model (LLM) assistants to autonomous vehicles. These systems often optimize their individual objective, leading to conflicts, particularly in general-sum games where naive reinforcement learning agents empirically converge to Pareto-suboptimal Nash equilibria. To address this issue, opponent shaping has emerged as a paradigm for finding socially beneficial equilibria in general-sum games. In this work, we introduce Advantage Alignment, a family of algorithms derived from first principles that perform opponent shaping efficiently and intuitively. We achieve this by aligning the advantages of interacting agents, increasing the probability of mutually beneficial actions when their interaction has been positive. We prove that existing opponent shaping methods implicitly perform Advantage Alignment. Compared to these methods, Advantage Alignment simplifies the mathematical formulation of opponent shaping, reduces the computational burden and extends to continuous action domains. We demonstrate the effectiveness of our algorithms across a range of social dilemmas, achieving state-of-the-art cooperation and robustness against exploitation.

Reward models are critical in techniques like Reinforcement Learning from Human Feedback (RLHF) and Inference Scaling Laws, where they guide language model alignment and select optimal responses. Despite their importance, existing reward model benchmarks often evaluate models by asking them to distinguish between responses generated by models of varying power. However, this approach fails to assess reward models on subtle but critical content changes and variations in style, resulting in a low correlation with policy model performance. To this end, we introduce RM-Bench, a novel benchmark designed to evaluate reward models based on their sensitivity to subtle content differences and resistance to style biases. Extensive experiments demonstrate that RM-Bench strongly correlates with policy model performance, making it a reliable reference for selecting reward models to align language models effectively. We evaluate nearly 40 reward models on RM-Bench. Our results reveal that even state-of-the-art models achieve an average performance of only 46.6%, which falls short of random-level accuracy (50%) when faced with style bias interference. These findings highlight the significant room for improvement in current reward models.

Reinforcement Learning with Human Feedback (RLHF) has achieved great success in aligning large language models (LLMs) with human preferences. Prevalent RLHF approaches are reward-based, following the Bradley-Terry (BT) model assumption, which may not fully capture the complexity of human preferences. In this paper, we explore RLHF under a general preference framework and approach it from a game-theoretic perspective. Specifically, we formulate the problem as a two-player game and propose a novel online algorithm, iterative Nash policy optimization (INPO). The key idea is to let the policy play against itself via no- regret learning, thereby approximating the Nash policy. Unlike previous methods, INPO bypasses the need for estimating the expected win rate for individual responses, which typically incurs high computational or annotation costs. Instead, we introduce a new loss objective that is directly minimized over a preference dataset. We provide theoretical analysis for our approach and demonstrate its effectiveness through experiments on various representative benchmarks. With an LLaMA-3-8B-based SFT model, INPO achieves a 42.6% length-controlled win rate on AlpacaEval 2.0 and a 37.8% win rate on Arena-Hard, showing substantial improvement over the state-of-the-art online RLHF algorithms.

Traditionally, reinforcement learning (RL) agents learn to solve new tasks by updating their neural network parameters through interactions with the task environment. However, recent works demonstrate that some RL agents, after certain pretraining procedures, can learn to solve unseen new tasks without parameter updates, a phenomenon known as in-context reinforcement learning (ICRL). The empirical success of ICRL is widely attributed to the hypothesis that the forward pass of the pretrained agent neural network implements an RL algorithm. In this paper, we support this hypothesis by showing, both empirically and theoretically, that when a transformer is trained for policy evaluation tasks, it can discover and learn to implement temporal difference learning in its forward pass.

Reinforcement learning from human feedback (RLHF) is an effective method for aligning large language models (LLMs) with human values. However, reward over-optimization remains an open challenge leading to discrepancies between the performance of LLMs under the reward model and the true human objectives. A primary contributor to reward over-optimization is the extrapolation error that arises when the reward model evaluates out-of-distribution (OOD) responses. However, current methods still fail to prevent the increasing frequency of OOD response generation during the reinforcement learning (RL) process and are not effective at handling extrapolation errors from OOD responses. In this work, we propose the *Behavior-Supported Policy Optimization* (BSPO) method to mitigate the reward over-optimization issue. Specifically, we define *behavior policy* as the next token distribution of the reward training dataset to model the in-distribution (ID) region of the reward model. Building on this, we introduce the behavior-supported Bellman operator to regularize the value function, penalizing all OOD values without impacting the ID ones. Consequently, BSPO reduces the generation of OOD responses during the RL process, thereby avoiding overestimation caused by the reward model’s extrapolation errors. Theoretically, we prove that BSPO guarantees a monotonic improvement of the supported policy until convergence to the optimal behavior-supported policy. Empirical results from extensive experiments show that BSPO outperforms baselines in preventing reward over-optimization due to OOD evaluation and finding the optimal ID policy.

The advent of large language models (LLMs) has revolutionized the field of text generation, producing outputs that closely mimic human-like writing. Although academic and industrial institutions have developed detectors to prevent the malicious usage of LLM-generated texts, other research has doubt about the robustness of these systems. To stress test these detectors, we introduce a humanized proxy-attack (HUMPA) strategy that effortlessly compromises LLMs, causing them to produce outputs that align with human-written text and mislead detection systems. Our method attacks the source model by leveraging a reinforcement learning (RL) fine-tuned humanized small language model (SLM) in the decoding phase. Through an in-depth analysis, we demonstrate that our attack strategy is capable of generating responses that are indistinguishable to detectors, preventing them from differentiating between machine-generated and human-written text. We conduct systematic evaluations on extensive datasets using proxy-attacked open-source models, including Llama2-13B, Llama3-70B, and Mixtral-8x7B in both white- and black-box settings. Our findings show that the proxy-attack strategy effectively deceives the leading detectors, resulting in an average AUROC drop of 70.4% across multiple datasets, with a maximum drop of 95.0% on a single dataset. Furthermore, in cross-discipline scenarios, our strategy also bypasses these detectors, leading to a significant relative decrease of up to 90.9%, while in cross-language scenario, the drop reaches 91.3%. Despite our proxy-attack strategy successfully bypassing the detectors with such significant relative drops, we find that the generation quality of the attacked models remains preserved, even within a modest utility budget, when compared to the text produced by the original, unattacked source model.

Generalization in visual reinforcement learning (VRL) aims to learn agents that can adapt to test environments with unseen visual distractions. Despite advances in robust representations learning, many methods do not take into account the essential downstream task of sequential decision-making. This leads to representations that lack critical long-term information, impairing decision-making abilities in test environments. To tackle this problem, we propose a novel robust action-value representation learning (ROUSER) under the information bottleneck (IB) framework. ROUSER learns robust representations to capture long-term information from the decision-making objective (i.e., action values). Specifically, ROUSER uses IB to encode robust representations by maximizing their mutual information with action values for long-term information, while minimizing mutual information with state-action pairs to discard irrelevant features. As action values are unknown, ROUSER proposes to decompose robust representations of state-action pairs into one-step rewards and robust representations of subsequent pairs. Thus, it can use known rewards to compute the loss for robust representation learning. Moreover, we show that ROUSER accurately estimates action values using learned robust representations, making it applicable to various VRL algorithms. Experiments demonstrate that ROUSER outperforms several state-of-the-art methods in eleven out of twelve tasks, across both unseen background and color distractions.

Self-play methods have demonstrated remarkable success in enhancing model capabilities across various domains. In the context of Reinforcement Learning from Human Feedback (RLHF), self-play not only boosts Large Language Model (LLM) performance but also overcomes the limitations of traditional Bradley-Terry (BT) model assumptions by finding the Nash equilibrium (NE) of a preference-based, two-player constant-sum game. However, existing methods either guarantee only average-iterate convergence, incurring high storage and inference costs, or converge to the NE of a regularized game, failing to accurately reflect true human preferences. In this paper, we introduce Magnetic Preference Optimization (MPO), a novel approach capable of achieving last-iterate convergence to the NE of the original game, effectively overcoming the limitations of existing methods. Building upon Magnetic Mirror Descent (MMD), MPO attains a linear convergence rate, making it particularly suitable for fine-tuning LLMs. To ensure our algorithm is both theoretically sound and practically viable, we present a simple yet effective implementation that adapts the theoretical insights to the RLHF setting. Empirical results demonstrate that MPO can significantly enhance the performance of LLMs, highlighting the potential of self-play methods in alignment.

Reinforcement Learning (RL) has achieved remarkable success in various continuous control tasks, such as robot manipulation and locomotion. Different to mainstream RL which makes decisions at individual steps, recent studies have incorporated action repetition into RL, achieving enhanced action persistence with improved sample efficiency and superior performance. However, existing methods treat all action dimensions as a whole during repetition, ignoring variations among them. This constraint leads to inflexibility in decisions, which reduces policy agility with inferior effectiveness. In this work, we propose a novel repetition framework called SDAR, which implements Spatially Decoupled Action Repetition through performing closed-loop act-or-repeat selection for each action dimension individually. SDAR achieves more flexible repetition strategies, leading to an improved balance between action persistence and diversity. Compared to existing repetition frameworks, SDAR is more sample efficient with higher policy performance and reduced action fluctuation. Experiments are conducted on various continuous control scenarios, demonstrating the effectiveness of spatially decoupled repetition design proposed in this work.

We present GameNGen, the first game engine powered entirely by a neural model that also enables real-time interaction with a complex environment over long trajectories at high quality. When trained on the classic game DOOM, GameNGen extracts gameplay and uses it to generate a playable environment that can interactively simulate new trajectories. GameNGen runs at 20 frames per second on a single TPU and remains stable over extended multi-minute play sessions. Next frame prediction achieves a PSNR of 29.4, comparable to lossy JPEG compression. Human raters are only slightly better than random chance at distinguishing short clips of the game from clips of the simulation, even after 5 minutes of auto-regressive generation. GameNGen is trained in two phases: (1) an RL-agent learns to play the game and the training sessions are recorded, and (2) a diffusion model is trained to produce the next frame, conditioned on the sequence of past frames and actions. Conditioning augmentations help ensure stable auto-regressive generation over long trajectories, and decoder fine-tuning improves the fidelity of visual details and text.

Despite a plethora of anomaly detection models developed over the years, their ability to generalize to unseen anomalies remains an issue, particularly in critical systems. This paper aims to address this challenge by introducing Swift Hydra, a new framework for training an anomaly detection method based on generative AI and reinforcement learning (RL). Through featuring an RL policy that operates on the latent variables of a generative model, the framework synthesizes novel and diverse anomaly samples that are capable of bypassing a detection model. These generated synthetic samples are, in turn, used to augment the detection model, further improving its ability to handle challenging anomalies. Swift Hydra also incorporates Mamba models structured as a Mixture of Experts (MoE) to enable scalable adaptation of the number of Mamba experts based on data complexity, effectively capturing diverse feature distributions without increasing the model’s inference time. Empirical evaluations on ADBench benchmark demonstrate that Swift Hydra outperforms other state-of-the-art anomaly detection models while maintaining a relatively short inference time. From these results, our research highlights a new and auspicious paradigm of integrating RL and generative AI for advancing anomaly detection.

Policy optimization methods benefit from a simple and tractable policy parametrization, usually the Gaussian for continuous action spaces. In this paper, we consider a broader policy family that remains tractable: the $q$-exponential family. This family of policies is flexible, allowing the specification of both heavy-tailed policies ($q>1$) and light-tailed policies ($q<1$). This paper examines the interplay between $q$-exponential policies for several actor-critic algorithms conducted on both online and offline problems. We find that heavy-tailed policies are more effective in general and can consistently improve on Gaussian. In particular, we find the Student's t-distribution to be more stable than the Gaussian across settings and that a heavy-tailed $q$-Gaussian for Tsallis Advantage Weighted Actor-Critic consistently performs well in offline benchmark problems. In summary, we find that the Student's t policy a strong candidate for drop-in replacement to the Gaussian. Our code is available at \url{https://github.com/lingweizhu/qexp}.

Direct Preference Optimization (DPO) has gained attention as an efficient alternative to reinforcement learning from human feedback (RLHF) for aligning large language models (LLMs) with human preferences. Despite its advantages, DPO suffers from a length bias, generating responses longer than those from the reference model. Existing solutions like SimPO and SamPO address this issue but uniformly treat the contribution of rewards across sequences, overlooking temporal dynamics. To this end, we propose an enhanced preference optimization method that incorporates a temporal decay factor controlled by a gamma parameter. This dynamic weighting mechanism adjusts the influence of each reward based on its position in the sequence, prioritizing earlier tokens that are more critical for alignment. By adaptively focusing on more relevant feedback, our approach mitigates overfitting to less pertinent data and remains responsive to evolving human preferences. Experimental results on several benchmarks show that our approach consistently outperforms vanilla DPO by 5.9-8.8 points on AlpacaEval 2 and 3.3-9.7 points on Arena-Hard across different model architectures and sizes. Furthermore, additional experiments on mathematical and reasoning benchmarks (MMLU, GSM8K, and MATH) confirm that our method enhances performance without compromising general capabilities. Our codebase would be available at \url{https://github.com/LotuSrc/D2PO}.

The growing deployment of reinforcement learning from human feedback (RLHF) calls for a deeper theoretical investigation of its underlying models. The prevalent models of RLHF do not account for neuroscience-backed, partially-observed "internal states'' that can affect human feedback, nor do they accommodate intermediate feedback during an interaction. Both of these can be instrumental in speeding up learning and improving alignment. To address these limitations, we model RLHF as reinforcement learning with partially observed reward-states (PORRL). We accommodate two kinds of feedback &mdash; cardinal and dueling feedback. We first demonstrate that PORRL subsumes a wide class of RL problems, including traditional RL, RLHF, and reward machines. For cardinal feedback, we present two model-based methods (POR-UCRL, POR-UCBVI). We give both cardinal regret and sample complexity guarantees for the methods, showing that they improve over naive history-summarization. We then discuss the benefits of a model-free method like GOLF with naive history-summarization in settings with recursive internal states and dense intermediate feedback. For this purpose, we define a new history aware version of the Bellman-eluder dimension and give a new guarantee for GOLF in our setting, which can be exponentially sharper in illustrative examples. For dueling feedback, we show that a naive reduction to cardinal feedback fails to achieve sublinear dueling regret. We then present the first explicit reduction that converts guarantees for cardinal regret to dueling regret. In both feedback settings, we show that our models and guarantees generalize and extend existing ones.

Diffusion models have been widely adopted in image and language generation and are now being applied to reinforcement learning. However, the application of diffusion models in offline cooperative Multi-Agent Reinforcement Learning (MARL) remains limited. Although existing studies explore this direction, they suffer from scalability or poor cooperation issues due to the lack of design principles for diffusion-based MARL. The Individual-Global-Max (IGM) principle is a popular design principle for cooperative MARL. By satisfying this principle, MARL algorithms achieve remarkable performance with good scalability. In this work, we extend the IGM principle to the Individual-Global-identically-Distributed (IGD) principle. This principle stipulates that the generated outcome of a multi-agent diffusion model should be identically distributed as the collective outcomes from multiple individual-agent diffusion models. We propose DoF, a diffusion factorization framework for Offline MARL. It uses noise factorization function to factorize a centralized diffusion model into multiple diffusion models. We theoretically show that the noise factorization functions satisfy the IGD principle. Furthermore, DoF uses data factorization function to model the complex relationship among data generated by multiple diffusion models. Through extensive experiments, we demonstrate the effectiveness of DoF. The source code is available at [https://github.com/xmu-rl-3dv/DoF](https://github.com/xmu-rl-3dv/DoF).

Designing synthetically accessible molecules and recommending analogs to unsynthesizable molecules are important problems for accelerating molecular discovery. We reconceptualize both problems using ideas from program synthesis. Drawing inspiration from syntax-guided synthesis approaches, we decouple the syntactic skeleton from the semantics of a synthetic tree to create a bilevel framework for reasoning about the combinatorial space of synthesis pathways. Given a molecule we aim to generate analogs for, we iteratively refine its skeletal characteristics via Markov Chain Monte Carlo simulations over the space of syntactic skeletons. Given a black-box oracle to optimize, we formulate a joint design space over syntactic templates and molecular descriptors and introduce evolutionary algorithms that optimize both syntactic and semantic dimensions synergistically. Our key insight is that once the syntactic skeleton is set, we can amortize over the search complexity of deriving the program's semantics by training policies to fully utilize the fixed horizon Markov Decision Process imposed by the syntactic template. We demonstrate performance advantages of our bilevel framework for synthesizable analog generation and synthesizable molecule design. Notably, our approach offers the user explicit control over the resources required to perform synthesis and biases the design space towards simpler solutions, making it particularly promising for autonomous synthesis platforms. Supporting code is at https://github.com/shiningsunnyday/SynthesisNet.

RAG systems consist of multiple modules to work together. However, these modules are usually separately trained. We argue that a system like RAG that incorporates multiple modules should be jointly optimized to achieve optimal performance. To demonstrate this, we design a specific pipeline called SmartRAG that includes a policy network and a retriever. The policy network can serve as 1) a decision maker that decides when to retrieve, 2) a query rewriter to generate a query most suited to the retriever and 3) an answer generator that produces the final response with/without the observations. We then propose to jointly optimize the whole system using a reinforcement learning algorithm, with the reward designed to encourage the system to achieve the highest performance with minimal retrieval cost. When jointly optimized, each module can be aware of how other modules are working and thus find the best way to work together as a complete system. Empirical results demonstrate that the jointly optimized system can achieve better performance than separately optimized counterparts.