Reinforcement learning algorithms are fundamental to align large language models with human preferences and to enhance their reasoning capabilities. However, current reinforcement learning algorithms often suffer from training instability due to loose on-policy constraints and computational inefficiency due to auxiliary models. In this work, we propose On-Policy RL with Optimal reward baseline (OPO), a novel and simplified reinforcement learning algorithm designed to address these challenges. OPO emphasizes the importance of exact on-policy training, which empirically stabilizes the training process and enhances exploration. Moreover, OPO introduces the optimal reward baseline that theoretically minimizes gradient variance. We evaluate OPO on mathematical reasoning benchmarks. The results demonstrate its superior performance and training stability without additional models or regularization terms. Furthermore, OPO achieves lower policy shifts and higher output entropy, encouraging more diverse and less repetitive responses. These results highlight OPO as a promising direction for stable and effective reinforcement learning in large language model alignment and reasoning tasks. The implementation is provided at https://github.com/microsoft/LMOps/tree/main/opo.
Rewards serve as proxies for human preferences and play a crucial role in Reinforcement Learning from Human Feedback (RLHF). However, if these rewards are inherently imperfect, exhibiting various biases, they can adversely affect the alignment of large language models (LLMs). In this paper, we collectively define the various biases present in rewards as the problem of reward unfairness. We propose a bias-agnostic method to address the issue of reward fairness from a resource allocation perspective, without specifically designing for each type of bias, yet effectively mitigating them. Specifically, we model preference learning as a resource allocation problem, treating rewards as resources to be allocated while considering the trade-off between utility and fairness in their distribution. We propose two methods, Fairness Regularization and Fairness Coefficient, to achieve fairness in rewards. We apply our methods in both verification and reinforcement learning scenarios to obtain a fairness reward model and a policy model, respectively. Experiments conducted in these scenarios demonstrate that our approach aligns LLMs with human preferences in a more fair manner.
While reinforcement learning (RL) over chains of thought has significantly advanced language models in tasks such as mathematics and coding, visual reasoning introduces added complexity by requiring models to direct visual attention, interpret perceptual inputs, and ground abstract reasoning in spatial evidence. We introduce ViGoRL (Visually Grounded Reinforcement Learning), a vision-language model trained with RL to explicitly anchor each reasoning step to specific visual coordinates. Inspired by human visual decision-making, ViGoRL learns to produce spatially grounded reasoning traces, guiding visual attention to task-relevant regions at each step. When fine-grained exploration is required, our novel multi-turn RL framework enables the model to dynamically zoom into predicted coordinates as reasoning unfolds. Across a diverse set of visual reasoning benchmarks--including SAT-2 and BLINK for spatial reasoning, V*bench for visual search, and ScreenSpot and VisualWebArena for web-based grounding--ViGoRL consistently outperforms both supervised fine-tuning and conventional RL baselines that lack explicit grounding mechanisms. Incorporating multi-turn RL with zoomed-in visual feedback significantly improves ViGoRL's performance on localizing small GUI elements and visual search, achieving 86.4% on V*Bench. Additionally, we find that grounding amplifies other visual behaviors such as region exploration, grounded subgoal setting, and visual verification. Finally, human evaluations show that the model's visual references are not only spatially accurate but also helpful for understanding model reasoning steps. Our results show that visually grounded RL is a strong paradigm for imbuing models with general-purpose visual reasoning.
The ability to continuously and efficiently transfer skills across tasks is a hallmark of biological intelligence and a long-standing goal in artificial systems. Reinforcement learning (RL), a dominant paradigm for learning in high-dimensional control tasks, is known to suffer from brittleness to task variations and catastrophic forgetting. Neuroevolution (NE) has recently gained attention for its robustness, scalability, and capacity to escape local optima. In this paper, we investigate an understudied dimension of NE: its transfer learning capabilities. To this end, we introduce two benchmarks: a) in stepping gates, neural networks are tasked with emulating logic circuits, with designs that emphasize modular repetition and variation b) ecorobot extends the Brax physics engine with objects such as walls and obstacles and the ability to easily switch between different robotic morphologies. Crucial in both benchmarks is the presence of a curriculum that enables evaluating skill transfer across tasks of increasing complexity. Our empirical analysis shows that NE methods vary in their transfer abilities and frequently outperform RL baselines. Our findings support the potential of NE as a foundation for building more adaptable agents and highlight future challenges for scaling NE to complex, real-world problems.
Large language models excel at complex tasks by breaking down problems into structured reasoning steps. However, reasoning traces often extend beyond reaching a correct answer, causing wasted computation, reduced readability, and hallucinations. To address this, we introduce a novel hyperparameter-free conciseness score used as a reward signal within a reinforcement learning framework to guide models toward generating correct and concise reasoning traces. This score is evaluated by a large language model acting as a judge, enabling dynamic, context-aware feedback beyond simple token length. Our method achieves state-of-the-art efficiency-accuracy trade-offs on the MATH dataset, reducing token usage by up to 31x on simple problems while improving accuracy by 7%, and on the hardest problems, it outperforms full reasoning by +7.5% accuracy with up to 3.6x fewer tokens. On TheoremQA, our method improves accuracy by +2.2% using 12.5x fewer tokens. We also conduct ablation studies on the judge model, reward composition, and problem difficulty, showing that our method dynamically adapts reasoning length based on problem difficulty and benefits significantly from stronger judges. The code, model weights, and datasets are open-sourced at https://github.com/RazvanDu/ConciseRL.
Recent advances have demonstrated the effectiveness of Reinforcement Learning (RL) in improving the reasoning capabilities of Large Language Models (LLMs). However, existing works inevitably rely on high-quality instructions and verifiable rewards for effective training, both of which are often difficult to obtain in specialized domains. In this paper, we propose Self-play Reinforcement Learning(SeRL) to bootstrap LLM training with limited initial data. Specifically, SeRL comprises two complementary modules: self-instruction and self-rewarding. The former module generates additional instructions based on the available data at each training step, employing robust online filtering strategies to ensure instruction quality, diversity, and difficulty. The latter module introduces a simple yet effective majority-voting mechanism to estimate response rewards for additional instructions, eliminating the need for external annotations. Finally, SeRL performs conventional RL based on the generated data, facilitating iterative self-play learning. Extensive experiments on various reasoning benchmarks and across different LLM backbones demonstrate that the proposed SeRL yields results superior to its counterparts and achieves performance on par with those obtained by high-quality data with verifiable rewards. Our code is available at https://github.com/wantbook-book/SeRL.
Reinforcement learning (RL) is increasingly applied to real-world problems involving complex and structured decisions, such as routing, scheduling, and assortment planning. These settings challenge standard RL algorithms, which struggle to scale, generalize, and exploit structure in the presence of combinatorial action spaces. We propose Structured Reinforcement Learning (SRL), a novel actor-critic framework that embeds combinatorial optimization layers into the actor neural network. We enable end-to-end learning of the actor via Fenchel-Young losses and provide a geometric interpretation of SRL as a primal-dual algorithm in the dual of the moment polytope. Across six environments with exogenous and endogenous uncertainty, SRL matches or surpasses the performance of unstructured RL and imitation learning on static tasks and improves over these baselines by up to 92% on dynamic problems, with improved stability and convergence speed.
Reinforcement fine-tuning (RFT) has emerged as a powerful post-training technique to incentivize the reasoning ability of large language models (LLMs). However, LLMs can respond very inconsistently to RFT: some show substantial performance gains, while others plateau or even degrade. To understand this divergence, we analyze the per-step influence of the RL objective and identify two key conditions for effective post-training: (1) RL-informative rollout accuracy, and (2) strong data co-influence, which quantifies how much the training data affects performance on other samples. Guided by these insights, we propose behavior injection, a task-agnostic data-augmentation scheme applied prior to RL. Behavior injection enriches the supervised finetuning (SFT) data by seeding exploratory and exploitative behaviors, effectively making the model more RL-ready. We evaluate our method across two reasoning benchmarks with multiple base models. The results demonstrate that our theoretically motivated augmentation can significantly increases the performance gain from RFT over the pre-RL model.
In this paper, we propose a robust and reinforcement-learning-enhanced network intrusion detection system (NIDS) designed for class-imbalanced and few-shot attack scenarios in Industrial Internet of Things (IIoT) environments. Our model integrates a TabTransformer for effective tabular feature representation with Proximal Policy Optimization (PPO) to optimize classification decisions via policy learning. Evaluated on the TON\textunderscore IoT benchmark, our method achieves a macro F1-score of 97.73\% and accuracy of 98.85\%. Remarkably, even on extremely rare classes like man-in-the-middle (MITM), our model achieves an F1-score of 88.79\%, showcasing strong robustness and few-shot detection capabilities. Extensive ablation experiments confirm the complementary roles of TabTransformer and PPO in mitigating class imbalance and improving generalization. These results highlight the potential of combining transformer-based tabular learning with reinforcement learning for real-world NIDS applications.
We trained 13,440 large language models and found that entropy minimization requires only a single unlabeled data and 10 steps optimization to achieve performance improvements comparable to or even greater than those obtained using thousands of data and carefully designed rewards in rule-based reinforcement learning. This striking result may prompt a rethinking of post-training paradigms for large language models. Our code is avaliable at https://github.com/zitian-gao/one-shot-em.
Large language models (LLMs) have demonstrated impressive performance on reasoning-intensive tasks, but enhancing their reasoning abilities typically relies on either reinforcement learning (RL) with verifiable signals or supervised fine-tuning (SFT) with high-quality long chain-of-thought (CoT) demonstrations, both of which are expensive. In this paper, we study a novel problem of incentivizing the reasoning capacity of LLMs without expensive high-quality demonstrations and reinforcement learning. We investigate whether the reasoning capabilities of LLMs can be effectively incentivized via supervision from significantly weaker models. We further analyze when and why such weak supervision succeeds in eliciting reasoning abilities in stronger models. Our findings show that supervision from significantly weaker reasoners can substantially improve student reasoning performance, recovering close to 94% of the gains of expensive RL at a fraction of the cost. Experiments across diverse benchmarks and model architectures demonstrate that weak reasoners can effectively incentivize reasoning in stronger student models, consistently improving performance across a wide range of reasoning tasks. Our results suggest that this simple weak-to-strong paradigm is a promising and generalizable alternative to costly methods for incentivizing strong reasoning capabilities at inference-time in LLMs. The code is publicly available at https://github.com/yuanyige/W2SR.
We present REARANK, a large language model (LLM)-based listwise reasoning reranking agent. REARANK explicitly reasons before reranking, significantly improving both performance and interpretability. Leveraging reinforcement learning and data augmentation, REARANK achieves substantial improvements over baseline models across popular information retrieval benchmarks, notably requiring only 179 annotated samples. Built on top of Qwen2.5-7B, our REARANK-7B demonstrates performance comparable to GPT-4 on both in-domain and out-of-domain benchmarks and even surpasses GPT-4 on reasoning-intensive BRIGHT benchmarks. These results underscore the effectiveness of our approach and highlight how reinforcement learning can enhance LLM reasoning capabilities in reranking.
Training large language models (LLMs) for complex reasoning via Reinforcement Learning with Verifiable Rewards (RLVR) is effective but limited by reliance on costly, domain-specific supervision. We explore Reinforcement Learning from Internal Feedback (RLIF), a framework that enables LLMs to learn from intrinsic signals without external rewards or labeled data. We propose Intuitor, an RLIF method that uses a model's own confidence, termed self-certainty, as its sole reward signal. Intuitor replaces external rewards in Group Relative Policy Optimization (GRPO) with self-certainty scores, enabling fully unsupervised learning. Experiments demonstrate that Intuitor matches GRPO's performance on mathematical benchmarks while achieving superior generalization to out-of-domain tasks like code generation, without requiring gold solutions or test cases. Our findings show that intrinsic model signals can drive effective learning across domains, offering a scalable alternative to RLVR for autonomous AI systems where verifiable rewards are unavailable. Code is available at https://github.com/sunblaze-ucb/Intuitor
Reasoning large language models (LLMs) excel in complex tasks, which has drawn significant attention to reinforcement learning (RL) for LLMs. However, existing approaches allocate an equal number of rollouts to all questions during the RL process, which is inefficient. This inefficiency stems from the fact that training on simple questions yields limited gains, whereas more rollouts are needed for challenging questions to sample correct answers. Furthermore, while RL improves response precision, it limits the model's exploration ability, potentially resulting in a performance cap below that of the base model prior to RL. To address these issues, we propose a mechanism for dynamically allocating rollout budgets based on the difficulty of the problems, enabling more efficient RL training. Additionally, we introduce an adaptive dynamic temperature adjustment strategy to maintain the entropy at a stable level, thereby encouraging sufficient exploration. This enables LLMs to improve response precision while preserving their exploratory ability to uncover potential correct pathways. The code and data is available on: https://github.com/LiaoMengqi/E3-RL4LLMs
Recent advances in large language models (LLMs) have introduced latent reasoning as a promising alternative to autoregressive reasoning. By performing internal computation with hidden states from previous steps, latent reasoning benefit from more informative features rather than sampling a discrete chain-of-thought (CoT) path. Yet latent reasoning approaches are often incompatible with LLMs, as their continuous paradigm conflicts with the discrete nature of autoregressive generation. Moreover, these methods rely on CoT traces for training and thus fail to exploit the inherent reasoning patterns of LLMs. In this work, we explore latent reasoning by leveraging the intrinsic capabilities of LLMs via reinforcement learning (RL). To this end, we introduce hybrid reasoning policy optimization (HRPO), an RL-based hybrid latent reasoning approach that (1) integrates prior hidden states into sampled tokens with a learnable gating mechanism, and (2) initializes training with predominantly token embeddings while progressively incorporating more hidden features. This design maintains LLMs' generative capabilities and incentivizes hybrid reasoning using both discrete and continuous representations. In addition, the hybrid HRPO introduces stochasticity into latent reasoning via token sampling, thereby enabling RL-based optimization without requiring CoT trajectories. Extensive evaluations across diverse benchmarks show that HRPO outperforms prior methods in both knowledge- and reasoning-intensive tasks. Furthermore, HRPO-trained LLMs remain interpretable and exhibit intriguing behaviors like cross-lingual patterns and shorter completion lengths, highlighting the potential of our RL-based approach and offer insights for future work in latent reasoning.
Sparse-reward reinforcement learning (RL) can model a wide range of highly complex tasks. Solving sparse-reward tasks is RL's core premise - requiring efficient exploration coupled with long-horizon credit assignment - and overcoming these challenges is key for building self-improving agents with superhuman ability. We argue that solving complex and high-dimensional tasks requires solving simpler tasks that are relevant to the target task. In contrast, most prior work designs strategies for selecting exploratory tasks with the objective of solving any task, making exploration of challenging high-dimensional, long-horizon tasks intractable. We find that the sense of direction, necessary for effective exploration, can be extracted from existing RL algorithms, without needing any prior information. Based on this finding, we propose a method for directed sparse-reward goal-conditioned very long-horizon RL (DISCOVER), which selects exploratory goals in the direction of the target task. We connect DISCOVER to principled exploration in bandits, formally bounding the time until the target task becomes achievable in terms of the agent's initial distance to the target, but independent of the volume of the space of all tasks. Empirically, we perform a thorough evaluation in high-dimensional environments. We find that the directed goal selection of DISCOVER solves exploration problems that are beyond the reach of prior state-of-the-art exploration methods in RL.
Ballbot (i.e. Ball balancing robot) navigation usually relies on methods rooted in control theory (CT), and works that apply Reinforcement learning (RL) to the problem remain rare while generally being limited to specific subtasks (e.g. balance recovery). Unlike CT based methods, RL does not require (simplifying) assumptions about environment dynamics (e.g. the absence of slippage between the ball and the floor). In addition to this increased accuracy in modeling, RL agents can easily be conditioned on additional observations such as depth-maps without the need for explicit formulations from first principles, leading to increased adaptivity. Despite those advantages, there has been little to no investigation into the capabilities, data-efficiency and limitations of RL based methods for ballbot control and navigation. Furthermore, there is a notable absence of an open-source, RL-friendly simulator for this task. In this paper, we present an open-source ballbot simulation based on MuJoCo, and show that with appropriate conditioning on exteroceptive observations as well as reward shaping, policies learned by classical model-free RL methods are capable of effectively navigating through randomly generated uneven terrain, using a reasonable amount of data (four to five hours on a system operating at 500hz).
This paper presents a pioneering exploration of reinforcement learning (RL) via group relative policy optimization for unified multimodal large language models (ULMs), aimed at simultaneously reinforcing generation and understanding capabilities. Through systematic pilot studies, we uncover the significant potential of ULMs to enable the synergistic co-evolution of dual capabilities within a shared policy optimization framework. Building on this insight, we introduce CoRL, a co-reinforcement learning framework comprising a unified RL stage for joint optimization and a refined RL stage for task-specific enhancement. With the proposed CoRL, our resulting model, ULM-R1, achieves average improvements of 7% on three text-to-image generation datasets and 23% on nine multimodal understanding benchmarks. These results demonstrate the effectiveness of CoRL and highlight the substantial benefit of reinforcement learning in facilitating cross-task synergy and optimization for ULMs. Code is available at https://github.com/mm-vl/ULM-R1.
Meta-reinforcement learning trains a single reinforcement learning agent on a distribution of tasks to quickly generalize to new tasks outside of the training set at test time. From a Bayesian perspective, one can interpret this as performing amortized variational inference on the posterior distribution over training tasks. Among the various meta-reinforcement learning approaches, a common method is to represent this distribution with a point-estimate using a recurrent neural network. We show how one can augment this point estimate to give full distributions through the Laplace approximation, either at the start of, during, or after learning, without modifying the base model architecture. With our approximation, we are able to estimate distribution statistics (e.g., the entropy) of non-Bayesian agents and observe that point-estimate based methods produce overconfident estimators while not satisfying consistency. Furthermore, when comparing our approach to full-distribution based learning of the task posterior, our method performs on par with variational baselines while having much fewer parameters.
Reinforcement learning (RL) with delays is challenging as sensory perceptions lag behind the actual events: the RL agent needs to estimate the real state of its environment based on past observations. State-of-the-art (SOTA) methods typically employ recursive, step-by-step forecasting of states. This can cause the accumulation of compounding errors. To tackle this problem, our novel belief estimation method, named Directly Forecasting Belief Transformer (DFBT), directly forecasts states from observations without incrementally estimating intermediate states step-by-step. We theoretically demonstrate that DFBT greatly reduces compounding errors of existing recursively forecasting methods, yielding stronger performance guarantees. In experiments with D4RL offline datasets, DFBT reduces compounding errors with remarkable prediction accuracy. DFBT's capability to forecast state sequences also facilitates multi-step bootstrapping, thus greatly improving learning efficiency. On the MuJoCo benchmark, our DFBT-based method substantially outperforms SOTA baselines.