Pessimistic Auxiliary Policy for Offline Reinforcement Learning

Offline reinforcement learning aims to learn an agent from pre-collected datasets, avoiding unsafe and inefficient real-time interaction. However, inevitable access to out-ofdistribution actions during the learning process introduces approximation errors, causing the error accumulation and considerable overestimation. In this paper, we construct a new pessimistic auxiliary policy for sampling reliable actions. Specifically, we develop a pessimistic auxiliary strategy by maximizing the lower confidence bound of the Q-function. The pessimistic auxiliary strategy exhibits a relatively high value and low uncertainty in the vicinity of the learned policy, avoiding the learned policy sampling high-value actions with potentially high errors during the learning process. Less approximation error introduced by sampled action from pessimistic auxiliary strategy leads to the alleviation of error accumulation. Extensive experiments on offline reinforcement learning benchmarks reveal that utilizing the pessimistic auxiliary strategy can effectively improve the efficacy of other offline RL approaches.

Competitive Multi-Operator Reinforcement Learning for Joint Pricing and Fleet Rebalancing in AMoD Systems

Autonomous Mobility-on-Demand (AMoD) systems promise to revolutionize urban transportation by providing affordable on-demand services to meet growing travel demand. However, realistic AMoD markets will be competitive, with multiple operators competing for passengers through strategic pricing and fleet deployment. While reinforcement learning has shown promise in optimizing single-operator AMoD control, existing work fails to capture competitive market dynamics. We investigate the impact of competition on policy learning by introducing a multi-operator reinforcement learning framework where two operators simultaneously learn pricing and fleet rebalancing policies. By integrating discrete choice theory, we enable passenger allocation and demand competition to emerge endogenously from utility-maximizing decisions. Experiments using real-world data from multiple cities demonstrate that competition fundamentally alters learned behaviors, leading to lower prices and distinct fleet positioning patterns compared to monopolistic settings. Notably, we demonstrate that learning-based approaches are robust to the additional stochasticity of competition, with competitive agents successfully converging to effective policies while accounting for partially unobserved competitor strategies.

Reward-Conditioned Reinforcement Learning

RL agents are typically trained under a single, fixed reward function, which makes them brittle to reward misspecification and limits their ability to adapt to changing task preferences. We introduce Reward-Conditioned Reinforcement Learning (RCRL), a framework that trains a single agent to optimize a family of reward specifications while collecting experience under only one nominal objective. RCRL conditions the agent on reward parameterizations and learns multiple reward objectives from a shared replay data entirely off-policy, enabling a single policy to represent reward-specific behaviors. Across single-task, multi-task, and vision-based benchmarks, we show that RCRL not only improves performance under the nominal reward parameterization, but also enables efficient adaptation to new parameterizations. Our results demonstrate that RCRL provides a scalable mechanism for learning robust, steerable policies without sacrificing the simplicity of single-task training.

Decoupling Task and Behavior: A Two-Stage Reward Curriculum in Reinforcement Learning for Robotics

Deep Reinforcement Learning is a promising tool for robotic control, yet practical application is often hindered by the difficulty of designing effective reward functions. Real-world tasks typically require optimizing multiple objectives simultaneously, necessitating precise tuning of their weights to learn a policy with the desired characteristics. To address this, we propose a two-stage reward curriculum where we decouple task-specific objectives from behavioral terms. In our method, we first train the agent on a simplified task-only reward function to ensure effective exploration before introducing the full reward that includes auxiliary behavior-related terms such as energy efficiency. Further, we analyze various transition strategies and demonstrate that reusing samples between phases is critical for training stability. We validate our approach on the DeepMind Control Suite, ManiSkill3, and a mobile robot environment, modified to include auxiliary behavioral objectives. Our method proves to be simple yet effective, substantially outperforming baselines trained directly on the full reward while exhibiting higher robustness to specific reward weightings.

BandPO: Bridging Trust Regions and Ratio Clipping via Probability-Aware Bounds for LLM Reinforcement Learning

Proximal constraints are fundamental to the stability of the Large Language Model reinforcement learning. While the canonical clipping mechanism in PPO serves as an efficient surrogate for trust regions, we identify a critical bottleneck: fixed bounds strictly constrain the upward update margin of low-probability actions, disproportionately suppressing high-advantage tail strategies and inducing rapid entropy collapse. To address this, we introduce Band-constrained Policy Optimization (BandPO). BandPO replaces canonical clipping with Band, a unified theoretical operator that projects trust regions defined by f-divergences into dynamic, probability-aware clipping intervals. Theoretical analysis confirms that Band effectively resolves this exploration bottleneck. We formulate this mapping as a convex optimization problem, guaranteeing a globally optimal numerical solution while deriving closed-form solutions for specific divergences. Extensive experiments across diverse models and datasets demonstrate that BandPO consistently outperforms canonical clipping and Clip-Higher, while robustly mitigating entropy collapse.

KARL: Knowledge Agents via Reinforcement Learning

We present a system for training enterprise search agents via reinforcement learning that achieves state-of-the-art performance across a diverse suite of hard-to-verify agentic search tasks. Our work makes four core contributions. First, we introduce KARLBench, a multi-capability evaluation suite spanning six distinct search regimes, including constraint-driven entity search, cross-document report synthesis, tabular numerical reasoning, exhaustive entity retrieval, procedural reasoning over technical documentation, and fact aggregation over internal enterprise notes. Second, we show that models trained across heterogeneous search behaviors generalize substantially better than those optimized for any single benchmark. Third, we develop an agentic synthesis pipeline that employs long-horizon reasoning and tool use to generate diverse, grounded, and high-quality training data, with iterative bootstrapping from increasingly capable models. Fourth, we propose a new post-training paradigm based on iterative large-batch off-policy RL that is sample efficient, robust to train-inference engine discrepancies, and naturally extends to multi-task training with out-of-distribution generalization. Compared to Claude 4.6 and GPT 5.2, KARL is Pareto-optimal on KARLBench across cost-quality and latency-quality trade-offs, including tasks that were out-of-distribution during training. With sufficient test-time compute, it surpasses the strongest closed models. These results show that tailored synthetic data in combination with multi-task reinforcement learning enables cost-efficient and high-performing knowledge agents for grounded reasoning.

SCoUT: Scalable Communication via Utility-Guided Temporal Grouping in Multi-Agent Reinforcement Learning

Communication can improve coordination in partially observed multi-agent reinforcement learning (MARL), but learning \emph{when} and \emph{who} to communicate with requires choosing among many possible sender-recipient pairs, and the effect of any single message on future reward is hard to isolate. We introduce \textbf{SCoUT} (\textbf{S}calable \textbf{Co}mmunication via \textbf{U}tility-guided \textbf{T}emporal grouping), which addresses both these challenges via temporal and agent abstraction within traditional MARL. During training, SCoUT resamples \textit{soft} agent groups every \(K\) environment steps (macro-steps) via Gumbel-Softmax; these groups are latent clusters that induce an affinity used as a differentiable prior over recipients. Using the same assignments, a group-aware critic predicts values for each agent group and maps them to per-agent baselines through the same soft assignments, reducing critic complexity and variance. Each agent is trained with a three-headed policy: environment action, send decision, and recipient selection. To obtain precise communication learning signals, we derive counterfactual communication advantages by analytically removing each sender's contribution from the recipient's aggregated messages. This counterfactual computation enables precise credit assignment for both send and recipient-selection decisions. At execution time, all centralized training components are discarded and only the per-agent policy is run, preserving decentralized execution. Project website, videos and code: \hyperlink{https://scout-comm.github.io/}{https://scout-comm.github.io/}

Boosting ASR Robustness via Test-Time Reinforcement Learning with Audio-Text Semantic Rewards

Recently, Automatic Speech Recognition (ASR) systems (e.g., Whisper) have achieved remarkable accuracy improvements but remain highly sensitive to real-world unseen data (data with large distribution shifts), including noisy environments and diverse accents. To address this issue, test-time adaptation (TTA) has shown great potential in improving the model adaptability at inference time without ground-truth labels, and existing TTA methods often rely on pseudo-labeling or entropy minimization. However, by treating model confidence as a learning signal, these methods may reinforce high-confidence errors, leading to confirmation bias that undermines adaptation. To overcome these limitations, we present ASR-TRA, a novel Test-time Reinforcement Adaptation framework inspired by causal intervention. More precisely, our method introduces a learnable decoder prompt and utilizes temperature-controlled stochastic decoding to generate diverse transcription candidates. These are scored by a reward model that measures audio-text semantic alignment, and the resulting feedback is used to update both model and prompt parameters via reinforcement learning. Comprehensive experiments on LibriSpeech with synthetic noise and L2 Arctic accented English datasets demonstrate that our method achieves higher accuracy while maintaining lower latency than existing TTA baselines. Ablation studies further confirm the effectiveness of combining audio and language-based rewards, highlighting our method's enhanced stability and interpretability. Overall, our approach provides a practical and robust solution for deploying ASR systems in challenging real-world conditions.

Diffusion Policy through Conditional Proximal Policy Optimization

Reinforcement learning (RL) has been extensively employed in a wide range of decision-making problems, such as games and robotics. Recently, diffusion policies have shown strong potential in modeling multi-modal behaviors, enabling more diverse and flexible action generation compared to the conventional Gaussian policy. Despite various attempts to combine RL with diffusion, a key challenge is the difficulty of computing action log-likelihood under the diffusion model. This greatly hinders the direct application of diffusion policies in on-policy reinforcement learning. Most existing methods calculate or approximate the log-likelihood through the entire denoising process in the diffusion model, which can be memory- and computationally inefficient. To overcome this challenge, we propose a novel and efficient method to train a diffusion policy in an on-policy setting that requires only evaluating a simple Gaussian probability. This is achieved by aligning the policy iteration with the diffusion process, which is a distinct paradigm compared to previous work. Moreover, our formulation can naturally handle entropy regularization, which is often difficult to incorporate into diffusion policies. Experiments demonstrate that the proposed method produces multimodal policy behaviors and achieves superior performance on a variety of benchmark tasks in both IsaacLab and MuJoCo Playground.

Breaking Contextual Inertia: Reinforcement Learning with Single-Turn Anchors for Stable Multi-Turn Interaction

While LLMs demonstrate strong reasoning capabilities when provided with full information in a single turn, they exhibit substantial vulnerability in multi-turn interactions. Specifically, when information is revealed incrementally or requires updates, models frequently fail to integrate new constraints, leading to a collapse in performance compared to their single-turn baselines. We term the root cause as \emph{Contextual Inertia}: a phenomenon where models rigidly adhere to previous reasoning traces. Even when users explicitly provide corrections or new data in later turns, the model ignores them, preferring to maintain consistency with its previous (incorrect) reasoning path. To address this, we introduce \textbf{R}einforcement \textbf{L}earning with \textbf{S}ingle-\textbf{T}urn \textbf{A}nchors (\textbf{RLSTA}), a generalizable training approach designed to stabilize multi-turn interaction across diverse scenarios and domains. RLSTA leverages the model's superior single-turn capabilities as stable internal anchors to provide reward signals. By aligning multi-turn responses with these anchors, RLSTA empowers models to break contextual inertia and self-calibrate their reasoning based on the latest information. Experiments show that RLSTA significantly outperforms standard fine-tuning and abstention-based methods. Notably, our method exhibits strong cross-domain generalization (e.g., math to code) and proves effective even without external verifiers, highlighting its potential for general-domain applications.