ICLR 2025 Papers

Human alignment in large language models (LLMs) is an active area of research. A recent groundbreaking work, direct preference optimization (DPO), has greatly simplified the process from past work in reinforcement learning from human feedback (RLHF) by bypassing the reward learning stage in RLHF. DPO, after training, provides an implicit reward model. In this work, we make a novel observation that this implicit reward model can by itself be used in a bootstrapping fashion to further align the LLM. Our approach is to use the rewards from a current LLM to construct a preference dataset, which is then used in subsequent DPO rounds. We incorporate two refinements to further improve our approach: 1) length-regularized reward shaping to make the preference dataset length-unbiased; 2) experience replay to enhance the quality of the preference dataset. Our approach, named self-alignment with DPO ImpliCit rEwards (DICE), shows great improvements in alignment. It achieves an increase of more than 8$\\%$ in lengthcontrolled win rate on AlpacaEval 2 for all the different base models that we tried, without relying on external feedback. Our code is available at https://github.com/sail-sg/dice.

Self-adaptive large language models (LLMs) aim to solve the challenges posed by traditional fine-tuning methods, which are often computationally intensive and static in their ability to handle diverse tasks. We introduce Transformer-Squared, a novel self-adaptation framework that adapts LLMs for unseen tasks in real-time by selectively adjusting only the singular components of their weight matrices. During inference, Transformer-Squared employs a two-pass mechanism: first, a dispatch system identifies the task properties, and then task-specific 'expert' vectors, trained using reinforcement learning, are dynamically mixed to obtain targeted behavior for the incoming prompt. Our method consistently outperforms ubiquitous approaches such as LoRA, with fewer parameters and greater efficiency. Furthermore, Transformer-Squared demonstrates versatility across different LLM architectures and modalities, including vision-language tasks. Transformer-Squared represents a significant leap forward, offering a scalable, efficient solution for enhancing the adaptability and task-specific performance of LLMs, paving the way for truly dynamic, self-organizing AI systems.

Offline Safe Reinforcement Learning (RL) seeks to address safety constraints by learning from static datasets and restricting exploration. However, these approaches heavily rely on the dataset and struggle to generalize to unseen scenarios safely. In this paper, we aim to improve safety during the deployment of vision-based robotic tasks through online fine-tuning an offline pretrained policy. To facilitate effective fine-tuning, we introduce model-based RL, which is known for its data efficiency. Specifically, our method employs in-sample optimization to improve offline training efficiency while incorporating reachability guidance to ensure safety. After obtaining an offline safe policy, a safe policy expansion approach is leveraged for online fine-tuning. The performance of our method is validated on simulation benchmarks with five vision-only tasks and through real-world robot deployment using limited data. It demonstrates that our approach significantly improves the generalization of offline policies to unseen safety-constrained scenarios. To the best of our knowledge, this is the first work to explore offline-to-online RL for safe generalization tasks. The videos are available at https://sunlighted.github.io/fosp_web/.

Hierarchical policies enable strong performance in many sequential decision-making problems, such as those with high-dimensional action spaces, those requiring long-horizon planning, and settings with sparse rewards. However, learning hierarchical policies from static offline datasets presents a significant challenge. Crucially, actions taken by higher-level policies may not be directly observable within hierarchical controllers, and the offline dataset might have been generated using a different policy structure, hindering the use of standard offline learning algorithms. In this work, we propose $\textit{OHIO}$: a framework for offline reinforcement learning (RL) of hierarchical policies. Our framework leverages knowledge of the policy structure to solve the $\textit{inverse problem}$, recovering the unobservable high-level actions that likely generated the observed data under our hierarchical policy. This approach constructs a dataset suitable for off-the-shelf offline training. We demonstrate our framework on robotic and network optimization problems and show that it substantially outperforms end-to-end RL methods and improves robustness. We investigate a variety of instantiations of our framework, both in direct deployment of policies trained offline and when online fine-tuning is performed. Code and data are available at https://ohio-offline-hierarchical-rl.github.io.

Direct Preference Optimization (DPO) has recently emerged as a popular approach to improve reinforcement learning from human feedback (RLHF), leading to better techniques to fine-tune large language models (LLM). A weakness of DPO, however, lies in its lack of capability to characterize the diversity of human preferences. Inspired by Mallows' theory of preference ranking, we develop in this paper a new approach, the *MallowsPO*. A distinct feature of this approach is a *dispersion index*, which reflects the dispersion of human preference to prompts. We show that existing DPO models can be reduced to special cases of this dispersion index, thus unified with MallowsPO. More importantly, we demonstrate empirically how to use this dispersion index to enhance the performance of DPO in a broad array of benchmark tasks, from synthetic bandit selection to controllable generation and dialogues, while maintaining great generalization capabilities. MallowsPO is also compatible with other SOTA offline preference optimization methods, boosting nearly 2\% extra LC win rate when used as a plugin for fine-tuning Llama3-Instruct.

Reward inference (learning a reward model from human preferences) is a critical intermediate step in the Reinforcement Learning from Human Feedback (RLHF) pipeline for fine-tuning Large Language Models (LLMs). In practice, RLHF faces fundamental challenges such as distribution shift, reward model overfitting, and problem misspecification. An alternative approach is direct policy optimization without reward inference, such as Direct Preference Optimization (DPO), which provides a much simpler pipeline and has shown empirical success in LLM applications. However, DPO utilizes the closed-form expression between the optimal policy and the reward function, which is only suitable under the bandit setting or deterministic MDPs. This paper develops two RLHF algorithms without reward inference for general RL problems beyond bandits and deterministic MDPs, and general preference models beyond the Bradley-Terry model. The key idea is to estimate the local value function difference from human preferences and then approximate the policy gradient with a zeroth-order gradient approximator. For both algorithms, we establish polynomial convergence rates in terms of the number of policy gradient iterations, the number of trajectory samples, and human preference queries per iteration. Numerical experiments in stochastic environments validate the performance of our proposed algorithms, outperforming popular RLHF baselines such as DPO and PPO. Our paper shows there exist provably efficient methods to solve general RLHF problems without reward inference.

With the rapidly expanding landscape of large language models, aligning model generations with human values and preferences is becoming increasingly important. Popular alignment methods, such as Reinforcement Learning from Human Feedback, have shown significant success in guiding models with greater control. However, these methods require considerable computational resources, which is inefficient, and substantial collection of training data to accommodate the diverse and pluralistic nature of human preferences, which is impractical. These limitations significantly constrain the scope and efficacy of both task-specific and general preference alignment methods. In this work, we introduce On-the-fly Preference Alignment via Principle-Guided Decoding (OPAD) to directly align model outputs with human preferences during inference, eliminating the need for fine-tuning. Our approach involves first curating a surrogate solution to an otherwise infeasible optimization problem and then designing a principle-guided reward function based on this surrogate. The final decoding policy is derived by maximizing this customized reward, which exploits the discrepancy between the constrained policy and its unconstrained counterpart. OPAD directly modifies the model’s predictions during inference, ensuring principle adherence without incurring the computational overhead of retraining or fine-tuning. Experiments show that OPAD achieves competitive or superior performance in both general and personalized alignment tasks, demonstrating its efficiency and effectiveness compared to state-of-the-art baselines.

Reinforcement Learning from Human Feedback (RLHF) has shown great potential in fine-tuning Large Language Models (LLMs) to align with human preferences. Existing methods perform preference alignment from a fixed dataset, which can be limited in data coverage and the resulting reward model is hard to generalize in out-of-distribution responses. Thus, online RLHF is more desirable to empower the LLM to explore outside the support of the initial dataset by iteratively collecting the prompt-response pairs. In this paper, we study the fundamental problem in online RLHF, i.e., how to explore for LLM. We give a theoretical motivation in linear reward assumption to show that an optimistic reward with an upper confidence bound (UCB) term leads to a provably efficient RLHF policy. Then, we reformulate our objective to direct preference optimization with an exploration term, where the UCB-term can be converted to a count-based exploration bonus. We further propose a practical algorithm, named Count-based Online Preference Optimization (COPO), which leverages a simple coin-flip counting module to estimate the pseudo-count of a prompt-response pair in previously collected data. COPO encourages LLMs to balance exploration and preference optimization in an iterative manner, which enlarges the exploration space and the entire data coverage of iterative LLM policies. We conduct online RLHF experiments on Zephyr and Llama-3 models. The results on instruction-following and standard academic benchmarks show that COPO significantly increases performance.

We introduce a new benchmark for reward models that quantifies their ability to produce strong language models through RLHF (Reinforcement Learning from Human Feedback). The gold-standard approach is to run a full RLHF training pipeline and directly probe downstream LLM performance. However, this process is prohibitively expensive. To address this, we build a predictive model of downstream LLM performance by evaluating the reward model on proxy tasks. These proxy tasks consist of a large-scale human preference and a verifiable correctness preference dataset, in which we measure 12 metrics across 12 domains. To investigate which reward model metrics are most correlated to gold-standard RLHF outcomes, we launch an end-to-end RLHF experiment on a large-scale crowd-sourced human preference platform to view real reward model downstream performance as ground truth. Ultimately, we compile our data and findings into Preference Proxy Evaluations (PPE), the first reward model benchmark explicitly linked to post-RLHF real-world human preference performance, which we opensource for public use and further development at https://github.com/lmarena/PPE.

Large Language Models (LLMs) have achieved remarkable success at tasks like summarization that involve a single turn of interaction. However, they can still struggle with multi-turn tasks like dialogue that require long-term planning. Previous works on multi-turn dialogue extend single-turn reinforcement learning from human feedback (RLHF) methods to the multi-turn setting by treating all prior dialogue turns as a long context. Such approaches suffer from covariate shift: the conversations in the training set have previous turns generated by some reference policy, which means that low training error may not necessarily correspond to good performance when the learner is actually in the conversation loop. In response, we introduce REgressing the RELative FUture (REFUEL), an efficient policy optimization approach designed to address multi-turn RLHF in LLMs. REFUEL employs a single model to estimate $Q$-values and trains on self-generated data, addressing the covariate shift issue. REFUEL frames the multi-turn RLHF problem as a sequence of regression tasks on iteratively collected datasets, enabling ease of implementation. Theoretically, we prove that REFUEL can match the performance of any policy covered by the training set. Empirically, we evaluate our algorithm by using Llama-3.1-70B-it to simulate a user in conversation with our model. REFUEL consistently outperforms state-of-the-art methods such as DPO and REBEL across various settings. Furthermore, despite having only 8 billion parameters, Llama-3-8B-it fine-tuned with REFUEL outperforms Llama-3.1-70B-it on long multi-turn dialogues. Implementation of REFUEL can be found at https://github.com/ZhaolinGao/REFUEL/, and models trained by REFUEL can be found at https://huggingface.co/Cornell-AGI.

Neural network training can be accelerated when a learnable update rule is used in lieu of classic adaptive optimizers (e.g. Adam). However, learnable update rules can be costly and unstable to train and use. Recently, Jang et al. (2023) proposed a simpler approach to accelerate training based on weight nowcaster networks (WNNs). In their approach, Adam is used for most of the optimization steps and periodically, only every few steps, a WNN nowcasts (predicts near future) parameters. We improve WNNs by proposing neuron interaction and nowcasting (NiNo) networks. In contrast to WNNs, NiNo leverages neuron connectivity and graph neural networks to more accurately nowcast parameters. We further show that in some networks, such as Transformers, modeling neuron connectivity accurately is challenging. We address this and other limitations, which allows NiNo to accelerate Adam training by up to 50% in vision and language tasks.

Embodiment co-design aims to optimize a robot's morphology and control policy simultaneously. While prior work has demonstrated its potential for generating environment-adaptive robots, this field still faces persistent challenges in optimization efficiency due to the (i) combinatorial nature of morphological search spaces and (ii) intricate dependencies between morphology and control. We prove that the ineffective morphology representation and unbalanced reward signals between the design and control stages are key obstacles to efficiency. To advance towards efficient embodiment co-design, we propose **BodyGen**, which utilizes (1) topology-aware self-attention for both design and control, enabling efficient morphology representation with lightweight model sizes; (2) a temporal credit assignment mechanism that ensures balanced reward signals for optimization. With our findings, BodyGen achieves an average **60.03%** performance improvement against state-of-the-art baselines. We provide codes and more results on the website: https://genesisorigin.github.io.

One core capability of large language models~(LLMs) is to follow natural language instructions. However, the issue of automatically constructing high-quality training data to enhance the complex instruction-following abilities of LLMs without manual annotation remains unresolved. In this paper, we introduce AutoIF, the first scalable and reliable method for automatically generating instruction-following training data. AutoIF transforms the validation of instruction-following data quality into code verification, requiring LLMs to generate instructions, the corresponding code to verify the correctness of the instruction responses, and unit test samples to cross-validate the code's correctness. Then, execution feedback-based rejection sampling can generate data for Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF) training. AutoIF achieves significant improvements across three training algorithms, SFT, Offline DPO, and Online DPO, when applied to the advanced open-source LLMs, Qwen2 and LLaMA3, in self-alignment and strong-to-weak distillation settings. Using two widely-used and three challenging general instruction-following benchmarks, we demonstrate that AutoIF significantly improves LLM performance across a wide range of natural instruction constraints. Notably, AutoIF is the first to surpass 90\% accuracy in IFEval’s loose instruction accuracy, without compromising general, math and coding capabilities. Further analysis of quality, scaling, combination, and data efficiency highlights AutoIF's strong generalization and alignment potential. Our code are available at https://github.com/QwenLM/AutoIF

Collaboration is a key challenge in distributed multi-agent reinforcement learning (MARL) environments. Learning frameworks for these decentralized systems must weigh the benefits of explicit player coordination against the communication overhead and computational cost of sharing local observations and environmental data. Quantum computing has sparked a potential synergy between quantum entanglement and cooperation in multi-agent environments, which could enable more efficient distributed collaboration with minimal information sharing. This relationship is largely unexplored, however, as current state-of-the-art quantum MARL (QMARL) implementations rely on classical information sharing rather than entanglement over a quantum channel as a coordination medium. In contrast, in this paper, a novel framework dubbed entangled QMARL (eQMARL) is proposed. The proposed eQMARL is a distributed actor-critic framework that facilitates cooperation over a quantum channel and eliminates local observation sharing via a quantum entangled split critic. Introducing a quantum critic uniquely spread across the agents allows coupling of local observation encoders through entangled input qubits over a quantum channel, which requires no explicit sharing of local observations and reduces classical communication overhead. Further, agent policies are tuned through joint observation-value function estimation via joint quantum measurements, thereby reducing the centralized computational burden. Experimental results show that eQMARL with $\Psi^{+}$ entanglement converges to a cooperative strategy up to $17.8\\%$ faster and with a higher overall score compared to split classical and fully centralized classical and quantum baselines. The results also show that eQMARL achieves this performance with a constant factor of $25$-times fewer centralized parameters compared to the split classical baseline.

Designing effective reward functions is crucial to training reinforcement learning (RL) algorithms. However, this design is non-trivial, even for domain experts, due to the subjective nature of certain tasks that are hard to quantify explicitly. In recent works, large language models (LLMs) have been used for reward generation from natural language task descriptions, leveraging their extensive instruction tuning and commonsense understanding of human behavior. In this work, we hypothesize that LLMs, guided by human feedback, can be used to formulate reward functions that reflect human implicit knowledge. We study this in three challenging settings -- autonomous driving, humanoid locomotion, and dexterous manipulation -- wherein notions of ``good" behavior are tacit and hard to quantify. To this end, we introduce REvolve, a truly evolutionary framework that uses LLMs for reward design in RL. REvolve generates and refines reward functions by utilizing human feedback to guide the evolution process, effectively translating implicit human knowledge into explicit reward functions for training (deep) RL agents. Experimentally, we demonstrate that agents trained on REvolve-designed rewards outperform other state-of-the-art baselines.

Black-box optimization aims to find the optima through building a model close to the black-box objective function based on function value evaluation. However, in many real-world tasks, such as the design of molecular formulas and mechanical structures, it is perilous, costly, or even infeasible to evaluate the objective function value of an actively sampled solution. In this situation, optimization can only be conducted via utilizing offline historical data, which yields offline black-box optimization. Different from the traditional goal that is to pursue the optimal solution, this paper emphasizes that the goal of offline optimization is to stably surpass the offline dataset during optimization procedure. Although benchmarks called Design-Bench already exist in this emerging field, it can hardly evaluate the stability of offline optimization and mainly provides real-world offline tasks and the corresponding offline datasets. To this end, this paper proposes benchmarks named SOO-Bench (i.e., Stable Offline Optimization Benchmarks) for offline black-box optimization algorithms, so as to systematically evaluate the stability of surpassing the offline dataset under different data distributions. Along with SOO-Bench, we also propose a stability indicator to measure the degree of stability. Specifically, SOO-Bench includes various real-world offline optimization tasks and offline datasets under different data distributions, involving the fields of satellites, materials science, structural mechanics, and automobile manufacturing. Empirically, baseline and state-of-the-art algorithms are tested and analyzed on SOO-Bench. Hopefully, SOO-Bench is expected to serve as a catalyst for the rapid developments of more novel and stable offline optimization methods. The code is available at \url{https://github.com/zhuyiyi-123/SOO-Bench}.

The automation of scientific discovery has been a long-standing goal within the research community, driven by the potential to accelerate knowledge creation. While significant progress has been made using commercial large language models (LLMs) as research assistants or idea generators, the possibility of automating the entire research process with open-source LLMs remains largely unexplored. This paper explores the feasibility of using open-source post-trained LLMs as autonomous agents capable of performing the full cycle of automated research and review, from literature review and manuscript preparation to peer review and paper refinement. Our iterative preference training framework consists of CycleResearcher, which conducts research tasks, and CycleReviewer, which simulates the peer review process, providing iterative feedback via reinforcement learning. To train these models, we develop two new datasets, Review-5k and Research-14k, reflecting real-world machine learning research and peer review dynamics. Our results demonstrate that CycleReviewer achieves promising performance with a 26.89\% reduction in mean absolute error (MAE) compared to individual human reviewers in predicting paper scores, indicating the potential of LLMs to effectively assist expert-level research evaluation. In research, the papers generated by the CycleResearcher model achieved a score of 5.36 in simulated peer reviews, showing some competitiveness in terms of simulated review scores compared to the preprint level of 5.24 from human experts, while still having room for improvement compared to the accepted paper level of 5.69. This work represents a significant step toward fully automated scientific inquiry, providing ethical safeguards and exploring AI-driven research capabilities. The code, dataset and model weight are released at https://wengsyx.github.io/Researcher.

General intelligence requires quick adaptation across tasks. While existing reinforcement learning (RL) methods have made progress in generalization, they typically assume only distribution changes between source and target domains. In this paper, we explore a wider range of scenarios where not only the distribution but also the environment spaces may change. For example, in the CoinRun environment, we train agents from easy levels and generalize them to difficulty levels where there could be new enemies that have never occurred before. To address this challenging setting, we introduce a causality-guided self-adaptive representation-based approach, called CSR, that equips the agent to generalize effectively across tasks with evolving dynamics. Specifically, we employ causal representation learning to characterize the latent causal variables within the RL system. Such compact causal representations uncover the structural relationships among variables, enabling the agent to autonomously determine whether changes in the environment stem from distribution shifts or variations in space, and to precisely locate these changes. We then devise a three-step strategy to fine-tune the causal model under different scenarios accordingly. Empirical experiments show that CSR efficiently adapts to the target domains with only a few samples and outperforms state-of-the-art baselines on a wide range of scenarios, including our simulated environments, CartPole, CoinRun and Atari games.

In safe offline reinforcement learning, the objective is to develop a policy that maximizes cumulative rewards while strictly adhering to safety constraints, utilizing only offline data. Traditional methods often face difficulties in balancing these constraints, leading to either diminished performance or increased safety risks. We address these issues with a novel approach that begins by learning a conservatively safe policy through the use of Conditional Variational Autoencoders, which model the latent safety constraints. Subsequently, we frame this as a Constrained Reward-Return Maximization problem, wherein the policy aims to optimize rewards while complying with the inferred latent safety constraints. This is achieved by training an encoder with a reward-Advantage Weighted Regression objective within the latent constraint space. Our methodology is supported by theoretical analysis, including bounds on policy performance and sample complexity. Extensive empirical evaluation on benchmark datasets, including challenging autonomous driving scenarios, demonstrates that our approach not only maintains safety compliance but also excels in cumulative reward optimization, surpassing existing methods. Additional visualizations provide further insights into the effectiveness and underlying mechanisms of our approach.