We present a novel approach for training small language models for reasoning-intensive document ranking that combines knowledge distillation with reinforcement learning optimization. While existing methods often rely on expensive human annotations or large black-box language models, our methodology leverages web data and a teacher LLM to automatically generate high-quality training examples with relevance explanations. By framing document ranking as a reinforcement learning problem and incentivizing explicit reasoning capabilities, we train a compact 3B parameter language model that achieves state-of-the-art performance on the BRIGHT benchmark. Our model ranks third on the leaderboard while using substantially fewer parameters than other approaches, outperforming models that are over 20 times larger. Through extensive experiments, we demonstrate that generating explanations during inference, rather than directly predicting relevance scores, enables more effective reasoning with smaller language models. The self-supervised nature of our method offers a scalable and interpretable solution for modern information retrieval systems.
Reasoning-Oriented Reinforcement Learning (RORL) enhances the reasoning ability of Large Language Models (LLMs). However, due to the sparsity of rewards in RORL, effective training is highly dependent on the selection of problems of appropriate difficulty. Although curriculum learning attempts to address this by adjusting difficulty, it often relies on static schedules, and even recent online filtering methods lack theoretical grounding and a systematic understanding of their effectiveness. In this work, we theoretically and empirically show that curating the batch with the problems that the training model achieves intermediate accuracy on the fly can maximize the effectiveness of RORL training, namely balanced online difficulty filtering. We first derive that the lower bound of the KL divergence between the initial and the optimal policy can be expressed with the variance of the sampled accuracy. Building on those insights, we show that balanced filtering can maximize the lower bound, leading to better performance. Experimental results across five challenging math reasoning benchmarks show that balanced online filtering yields an additional 10% in AIME and 4% improvements in average over plain GRPO. Moreover, further analysis shows the gains in sample efficiency and training time efficiency, exceeding the maximum reward of plain GRPO within 60% training time and the volume of the training set.
Deep neuroevolution is a highly scalable alternative to reinforcement learning due to its unique ability to encode network updates in a small number of bytes. Recent insights from traditional deep learning indicate high-dimensional models possess intrinsic, low-rank structure. In this work, we introduce low-rank, factorized neuroevolution: an indirect encoding through which we can search a small space of low-rank factors that enforce underlying structure across a network's weights. We compare our approach with non-factorized networks of similar and smaller size to understand how much performance can be attributed to the smaller search space. We evaluate our method on a language modeling task using transformers, as well as continuous and discrete vision-based reinforcement learning tasks. Our study shows that low-rank, factorized neuroevolution outperforms or is competitive with non-factorized neuroevolution, performing notably well on language modeling. Our results also suggest deleterious factorized mutations have a stronger negative impact on performance than deleterious non-factorized mutations, which significantly reduces the runtime on environments with early termination for bad performers. More broadly, these results show how we can use insights from backpropgation-based methods to enhance neuroevolution
Quantum computing is currently strongly limited by the impact of noise, in particular introduced by the application of two-qubit gates. For this reason, reducing the number of two-qubit gates is of paramount importance on noisy intermediate-scale quantum hardware. To advance towards more reliable quantum computing, we introduce a framework based on ZX calculus, graph-neural networks and reinforcement learning for quantum circuit optimization. By combining reinforcement learning and tree search, our method addresses the challenge of selecting optimal sequences of ZX calculus rewrite rules. Instead of relying on existing heuristic rules for minimizing circuits, our method trains a novel reinforcement learning policy that directly operates on ZX-graphs, therefore allowing us to search through the space of all possible circuit transformations to find a circuit significantly minimizing the number of CNOT gates. This way we can scale beyond hard-coded rules towards discovering arbitrary optimization rules. We demonstrate our method's competetiveness with state-of-the-art circuit optimizers and generalization capabilities on large sets of diverse random circuits.
This research focuses on enhancing reinforcement learning (RL) algorithms by integrating penalty functions to guide agents in avoiding unwanted actions while optimizing rewards. The goal is to improve the learning process by ensuring that agents learn not only suitable actions but also which actions to avoid. Additionally, we reintroduce a bidirectional learning approach that enables agents to learn from both initial and terminal states, thereby improving speed and robustness in complex environments. Our proposed Penalty-Based Bidirectional methodology is tested against Mani skill benchmark environments, demonstrating an optimality improvement of success rate of approximately 4% compared to existing RL implementations. The findings indicate that this integrated strategy enhances policy learning, adaptability, and overall performance in challenging scenarios
We propose MORAL (a multimodal reinforcement learning framework for decision making in autonomous laboratories) that enhances sequential decision-making in autonomous robotic laboratories through the integration of visual and textual inputs. Using the BridgeData V2 dataset, we generate fine-tuned image captions with a pretrained BLIP-2 vision-language model and combine them with visual features through an early fusion strategy. The fused representations are processed using Deep Q-Network (DQN) and Proximal Policy Optimization (PPO) agents. Experimental results demonstrate that multimodal agents achieve a 20% improvement in task completion rates and significantly outperform visual-only and textual-only baselines after sufficient training. Compared to transformer-based and recurrent multimodal RL models, our approach achieves superior performance in cumulative reward and caption quality metrics (BLEU, METEOR, ROUGE-L). These results highlight the impact of semantically aligned language cues in enhancing agent learning efficiency and generalization. The proposed framework contributes to the advancement of multimodal reinforcement learning and embodied AI systems in dynamic, real-world environments.
Grasping large flat objects, such as books or keyboards lying horizontally, presents significant challenges for single-arm robotic systems, often requiring extra actions like pushing objects against walls or moving them to the edge of a surface to facilitate grasping. In contrast, dual-arm manipulation, inspired by human dexterity, offers a more refined solution by directly coordinating both arms to lift and grasp the object without the need for complex repositioning. In this paper, we propose a model-free deep reinforcement learning (DRL) framework to enable dual-arm coordination for grasping large flat objects. We utilize a large-scale grasp pose detection model as a backbone to extract high-dimensional features from input images, which are then used as the state representation in a reinforcement learning (RL) model. A CNN-based Proximal Policy Optimization (PPO) algorithm with shared Actor-Critic layers is employed to learn coordinated dual-arm grasp actions. The system is trained and tested in Isaac Gym and deployed to real robots. Experimental results demonstrate that our policy can effectively grasp large flat objects without requiring additional maneuvers. Furthermore, the policy exhibits strong generalization capabilities, successfully handling unseen objects. Importantly, it can be directly transferred to real robots without fine-tuning, consistently outperforming baseline methods.
Deep reinforcement learning agents, trained on raw pixel inputs, often fail to generalize beyond their training environments, relying on spurious correlations and irrelevant background details. To address this issue, object-centric agents have recently emerged. However, they require different representations tailored to the task specifications. Contrary to deep agents, no single object-centric architecture can be applied to any environment. Inspired by principles of cognitive science and Occam's Razor, we introduce Object-Centric Attention via Masking (OCCAM), which selectively preserves task-relevant entities while filtering out irrelevant visual information. Specifically, OCCAM takes advantage of the object-centric inductive bias. Empirical evaluations on Atari benchmarks demonstrate that OCCAM significantly improves robustness to novel perturbations and reduces sample complexity while showing similar or improved performance compared to conventional pixel-based RL. These results suggest that structured abstraction can enhance generalization without requiring explicit symbolic representations or domain-specific object extraction pipelines.
Natural Language to SQL (NL2SQL) has emerged as a critical task for enabling seamless interaction with databases. Recent advancements in Large Language Models (LLMs) have demonstrated remarkable performance in this domain. However, existing NL2SQL methods predominantly rely on closed-source LLMs leveraging prompt engineering, while open-source models typically require fine-tuning to acquire domain-specific knowledge. Despite these efforts, open-source LLMs struggle with complex NL2SQL tasks due to the indirect expression of user query objectives and the semantic gap between user queries and database schemas. Inspired by the application of reinforcement learning in mathematical problem-solving to encourage step-by-step reasoning in LLMs, we propose LearNAT (Learning NL2SQL with AST-guided Task Decomposition), a novel framework that improves the performance of open-source LLMs on complex NL2SQL tasks through task decomposition and reinforcement learning. LearNAT introduces three key components: (1) a Decomposition Synthesis Procedure that leverages Abstract Syntax Trees (ASTs) to guide efficient search and pruning strategies for task decomposition, (2) Margin-aware Reinforcement Learning, which employs fine-grained step-level optimization via DPO with AST margins, and (3) Adaptive Demonstration Reasoning, a mechanism for dynamically selecting relevant examples to enhance decomposition capabilities. Extensive experiments on two benchmark datasets, Spider and BIRD, demonstrate that LearNAT enables a 7B-parameter open-source LLM to achieve performance comparable to GPT-4, while offering improved efficiency and accessibility.
Without exact knowledge of the true system dynamics, optimal control of non-linear continuous-time systems requires careful treatment of epistemic uncertainty. In this work, we propose a probabilistic extension to Pontryagin's maximum principle by minimizing the mean Hamiltonian with respect to epistemic uncertainty. We show minimization of the mean Hamiltonian is a necessary optimality condition when optimizing the mean cost, and propose a multiple shooting numerical method scalable to large-scale probabilistic dynamical models, including ensemble neural ordinary differential equations. Comparisons against state-of-the-art methods in online and offline model-based reinforcement learning tasks show that our probabilistic Hamiltonian formulation leads to reduced trial costs in offline settings and achieves competitive performance in online scenarios. By bridging optimal control and reinforcement learning, our approach offers a principled and practical framework for controlling uncertain systems with learned dynamics.
We propose a learning architecture that allows symbolic control and guidance in reinforcement learning with deep neural networks. We introduce SymDQN, a novel modular approach that augments the existing Dueling Deep Q-Networks (DuelDQN) architecture with modules based on the neuro-symbolic framework of Logic Tensor Networks (LTNs). The modules guide action policy learning and allow reinforcement learning agents to display behaviour consistent with reasoning about the environment. Our experiment is an ablation study performed on the modules. It is conducted in a reinforcement learning environment of a 5x5 grid navigated by an agent that encounters various shapes, each associated with a given reward. The underlying DuelDQN attempts to learn the optimal behaviour of the agent in this environment, while the modules facilitate shape recognition and reward prediction. We show that our architecture significantly improves learning, both in terms of performance and the precision of the agent. The modularity of SymDQN allows reflecting on the intricacies and complexities of combining neural and symbolic approaches in reinforcement learning.
We propose a decentralized reinforcement learning solution for multi-agent shepherding of non-cohesive targets using policy-gradient methods. Our architecture integrates target-selection with target-driving through Proximal Policy Optimization, overcoming discrete-action constraints of previous Deep Q-Network approaches and enabling smoother agent trajectories. This model-free framework effectively solves the shepherding problem without prior dynamics knowledge. Experiments demonstrate our method's effectiveness and scalability with increased target numbers and limited sensing capabilities.
We present the first mechanistic evidence that model-free reinforcement learning agents can learn to plan. This is achieved by applying a methodology based on concept-based interpretability to a model-free agent in Sokoban -- a commonly used benchmark for studying planning. Specifically, we demonstrate that DRC, a generic model-free agent introduced by Guez et al. (2019), uses learned concept representations to internally formulate plans that both predict the long-term effects of actions on the environment and influence action selection. Our methodology involves: (1) probing for planning-relevant concepts, (2) investigating plan formation within the agent's representations, and (3) verifying that discovered plans (in the agent's representations) have a causal effect on the agent's behavior through interventions. We also show that the emergence of these plans coincides with the emergence of a planning-like property: the ability to benefit from additional test-time compute. Finally, we perform a qualitative analysis of the planning algorithm learned by the agent and discover a strong resemblance to parallelized bidirectional search. Our findings advance understanding of the internal mechanisms underlying planning behavior in agents, which is important given the recent trend of emergent planning and reasoning capabilities in LLMs through RL
Reinforcement learning (RL) -- algorithms that teach artificial agents to interact with environments by maximising reward signals -- has achieved significant success in recent years. These successes have been facilitated by advances in algorithms (e.g., deep Q-learning, deep deterministic policy gradients, proximal policy optimisation, trust region policy optimisation, and soft actor-critic) and specialised computational resources such as GPUs and TPUs. One promising research direction involves introducing goals to allow multimodal policies, commonly through hierarchical or curriculum reinforcement learning. These methods systematically decompose complex behaviours into simpler sub-tasks, analogous to how humans progressively learn skills (e.g. we learn to run before we walk, or we learn arithmetic before calculus). However, fully automating goal creation remains an open challenge. We present a novel probabilistic curriculum learning algorithm to suggest goals for reinforcement learning agents in continuous control and navigation tasks.
Dynamic metabolic control can enhance bioprocess flexibility and expand the available optimization degrees of freedom via real-time modulation of metabolic enzyme expression. This allows target metabolic fluxes to be dynamically tuned throughout the process. However, identifying optimal dynamic control policies is challenging due to the presence of potential metabolic burden, cytotoxic effects, and the generally high-dimensional solution space, making exhaustive experimentation impractical. Here, we propose an approach based on reinforcement learning to derive optimal dynamic metabolic control policies by allowing an agent or controller to interact with a surrogate dynamic model $\textit{in silico}$. To incorporate and test robustness, we apply domain randomization, enabling the controller to generalize across system uncertainties. Our approach provides an alternative to conventional model-based control such as model predictive control, which requires differentiating the models with respect to decision variables; an often impractical task when dealing with complex stochastic, nonlinear, stiff, or piecewise-defined dynamics. In contrast, our approach only requires forward integration, making the task computationally much simpler with off-the-shelf solvers. We demonstrate our approach with a case study on the dynamic control of acetyl-CoA carboxylase in $\textit{Escherichia coli}$ for fatty acid biosynthesis. The derived dynamic metabolic control policies outperform static control, achieving up to 40 % higher titers while remaining robust under uncertainty.
This paper proposes tackling safety-critical stochastic Reinforcement Learning (RL) tasks with a samplebased, model-based approach. At the core of the method lies a Model Predictive Control (MPC) scheme that acts as function approximation, providing a model-based predictive control policy. To ensure safety, a probabilistic Control Barrier Function (CBF) is integrated into the MPC controller. A sample-based approach with guarantees is employed to approximate the effects of stochasticies in the optimal control formulation and to guarantee the probabilistic CBF condition. A learnable terminal cost formulation is included in the MPC objective to counterbalance the additional computational burden due to sampling. An RL algorithm is deployed to learn both the terminal cost and the CBF constraint. Results from our numerical experiment on a constrained LTI problem corroborate the effectiveness of the proposed methodology in reducing computation time while preserving control performance and safety.
The economic feasibility of nuclear microreactors will depend on minimizing operating costs through advancements in autonomous control, especially when these microreactors are operating alongside other types of energy systems (e.g., renewable energy). This study explores the application of deep reinforcement learning (RL) for real-time drum control in microreactors, exploring performance in regard to load-following scenarios. By leveraging a point kinetics model with thermal and xenon feedback, we first establish a baseline using a single-output RL agent, then compare it against a traditional proportional-integral-derivative (PID) controller. This study demonstrates that RL controllers, including both single- and multi-agent RL (MARL) frameworks, can achieve similar or even superior load-following performance as traditional PID control across a range of load-following scenarios. In short transients, the RL agent was able to reduce the tracking error rate in comparison to PID. Over extended 300-minute load-following scenarios in which xenon feedback becomes a dominant factor, PID maintained better accuracy, but RL still remained within a 1% error margin despite being trained only on short-duration scenarios. This highlights RL's strong ability to generalize and extrapolate to longer, more complex transients, affording substantial reductions in training costs and reduced overfitting. Furthermore, when control was extended to multiple drums, MARL enabled independent drum control as well as maintained reactor symmetry constraints without sacrificing performance -- an objective that standard single-agent RL could not learn. We also found that, as increasing levels of Gaussian noise were added to the power measurements, the RL controllers were able to maintain lower error rates than PID, and to do so with less control effort.
While unmanned aerial vehicles (UAVs) with flexible mobility are envisioned to enhance physical layer security in wireless communications, the efficient security design that adapts to such high network dynamics is rather challenging. The conventional approaches extended from optimization perspectives are usually quite involved, especially when jointly considering factors in different scales such as deployment and transmission in UAV-related scenarios. In this paper, we address the UAV-enabled multi-user secure communications by proposing a deep graph reinforcement learning framework. Specifically, we reinterpret the security beamforming as a graph neural network (GNN) learning task, where mutual interference among users is managed through the message-passing mechanism. Then, the UAV deployment is obtained through soft actor-critic reinforcement learning, where the GNN-based security beamforming is exploited to guide the deployment strategy update. Simulation results demonstrate that the proposed approach achieves near-optimal security performance and significantly enhances the efficiency of strategy determination. Moreover, the deep graph reinforcement learning framework offers a scalable solution, adaptable to various network scenarios and configurations, establishing a robust basis for information security in UAV-enabled communications.
Efficient and robust bioprocess control is essential for maximizing performance and adaptability in advanced biotechnological systems. In this work, we present a reinforcement-learning framework for multi-setpoint and multi-trajectory tracking. Tracking multiple setpoints and time-varying trajectories in reinforcement learning is challenging due to the complexity of balancing multiple objectives, a difficulty further exacerbated by system uncertainties such as uncertain initial conditions and stochastic dynamics. This challenge is relevant, e.g., in bioprocesses involving microbial consortia, where precise control over population compositions is required. We introduce a novel return function based on multiplicative reciprocal saturation functions, which explicitly couples reward gains to the simultaneous satisfaction of multiple references. Through a case study involving light-mediated cybergenetic growth control in microbial consortia, we demonstrate via computational experiments that our approach achieves faster convergence, improved stability, and superior control compliance compared to conventional quadratic-cost-based return functions. Moreover, our method enables tuning of the saturation function's parameters, shaping the learning process and policy updates. By incorporating system uncertainties, our framework also demonstrates robustness, a key requirement in industrial bioprocessing. Overall, this work advances reinforcement-learning-based control strategies in bioprocess engineering, with implications in the broader field of process and systems engineering.
As large language models (LLMs) continue to evolve, ensuring their alignment with human goals and values remains a pressing challenge. A key concern is \textit{instrumental convergence}, where an AI system, in optimizing for a given objective, develops unintended intermediate goals that override the ultimate objective and deviate from human-intended goals. This issue is particularly relevant in reinforcement learning (RL)-trained models, which can generate creative but unintended strategies to maximize rewards. In this paper, we explore instrumental convergence in LLMs by comparing models trained with direct RL optimization (e.g., the o1 model) to those trained with reinforcement learning from human feedback (RLHF). We hypothesize that RL-driven models exhibit a stronger tendency for instrumental convergence due to their optimization of goal-directed behavior in ways that may misalign with human intentions. To assess this, we introduce InstrumentalEval, a benchmark for evaluating instrumental convergence in RL-trained LLMs. Initial experiments reveal cases where a model tasked with making money unexpectedly pursues instrumental objectives, such as self-replication, implying signs of instrumental convergence. Our findings contribute to a deeper understanding of alignment challenges in AI systems and the risks posed by unintended model behaviors.