Researchers proposed ConSPO, a new reinforcement learning method that addresses limitations in existing algorithms like GRPO. ConSPO uses length-normalized sequence log-probabilities as rollout scores and contrasts them against negative distractors, leading to improved performance on reasoning benchmarks.

A new offline reinforcement learning algorithm called CDQAC is introduced, which learns effective scheduling policies directly from static, suboptimal datasets. This approach is shown to be sample-efficient and outperforms recent online and offline RL baselines on Job Shop Scheduling and Flexible JSP benchmarks.

Researchers studied reinforcement learning under external influence, proposing an algorithm to tackle the problem and providing guarantees for policy improvement. The work has implications for AI agents operating in dynamic environments.

A new model-free reinforcement learning framework, Deep-BQRL, has been developed for risk-sensitive decision-making. It extends buffered-quantile learning to neural function approximation, allowing for more general applicability and efficient exploration. The method is compared to existing approaches in experiments, demonstrating its effectiveness in solving asset-selling and slippery FrozenLake problems.

Researchers propose a reinforcement learning objective, multi-axis max@K, to improve diversity in text-to-image generation models. This method aims to cover a predefined set of visually distinct modes in generated images. Initial results show improved fairness scores while maintaining image quality and text alignment.

Researchers propose a reinforcement learning framework to generate graphs with precise structural properties, such as assortativity, using degree-preserving rewirings. This enables the creation of graphs with controlled structure and function, which is useful for AI model evaluation and development.

Researchers conducted a systematic analysis of reinforcement learning design space for diffusion models, focusing on likelihood estimation. They found that using an evidence lower bound (ELBO) based model likelihood estimator enables effective, efficient, and stable RL optimization, improving performance on multiple tasks.

Researchers propose CURB, a reward-shaping framework to mitigate retaliatory algorithmic collusion in reinforcement learning agents. CURB detects and penalizes collusion by measuring the total variation distance between an agent's action distributions across cooperation and defection histories. This approach is guaranteed to convert any collusive fixed point into a trivial one, preventing sustained collusive equilibria.

EPIG-Tree is a new algorithm for gradient-efficient reinforcement learning, specifically designed for language models. It improves policy-gradient estimation by allocating branches in a tree-based rollout construction to reduce uncertainty about the policy gradient per unit of compute. This can lead to better gradient estimation and improved performance in tasks such as Wordle and cloned-state control.

Researchers investigate the transparency and scrutability of recommender systems enhanced by generated natural-language user profiles. They reproduce and extend the original study, demonstrating that user profile recommendation (UPR) achieves competitive performance and makes recommendations more transparent.

MAGMA-GEN is a data-generation pipeline that converts ambiguous failed rollouts into validated recovery supervision for hierarchical robotic systems executing long-horizon manipulation tasks. It uses a privileged coach to hypothesize early decision-level errors and propose localized corrections or recovery actions, improving task success and recovery capabilities.

Researchers proposed a new algorithm, Regularized Emphatic Temporal-Difference Learning (RETD), to stabilize the expected off-policy TD update in reinforcement learning. RETD is designed to improve stability and convergence in scenarios with constant step sizes.

EfficientTDMPC, a sample-efficient model-based reinforcement learning method, improves on the TD-MPC family of algorithms by introducing an ensemble of dynamics models and an uncertainty penalty to reduce estimation errors. This enables EfficientTDMPC to achieve state-of-the-art sample efficiency on several benchmarks, including HumanoidBench-Hard and DMC hard.

RLLBC-Lib is an open-source code library for reinforcement learning and learning-based control, aiming to simplify the learning process for students and developers. It provides a comprehensive library of tabular RL approaches and a deep RL library, as well as implementations illustrating core RL principles.

This research explores the application of reinforcement learning to adaptively couple full and reduced order models in hybrid simulations, particularly in transient problems where localized features propagate through the domain. The approach uses deep Q-networks to select between subdomain-local full order models and pre-trained operator inference reduced order models to balance accuracy, cost, and model-switching frequency.

Researchers developed a decentralized control method for multi-humanoid robots to pick up and transport objects. This approach uses a shared control abstraction, allowing robots to learn from single-robot pickup and transfer those skills to cooperative multi-robot transport. The method was tested in simulation and on hardware, demonstrating success in various scenarios.

Researchers propose CALOS, a runtime safety layer for safe deep reinforcement learning in quadrotor control. CALOS enforces attitude constraints without modifying the underlying learning algorithm, using a quadratic program to compute the minimum-norm correction to the nominal torque output. This improves performance by 55-60% and accelerates training convergence.

Researchers introduced Reinforcement Learning from Self-Feedback (RLSF), a post-training stage that refines large language models by leveraging their intrinsic confidence as a self-generated reward. This approach refines probability estimates and strengthens step-by-step reasoning, improving performance on arithmetic reasoning and multiple-choice question answering.

R3 is a novel reward modeling framework for language models that is rubric-agnostic, generalizable, and interpretable, enabling more transparent and flexible evaluation of language models.

LP-BTS is a learning-guided planning architecture for large dynamic action spaces, specifically designed for one-to-many mobile charging scenarios. It uses a graph proposal policy, a learned value critic, and edge-budgeted PUCT to evaluate and select actions. The architecture achieves state-of-the-art results in a controlled setting, outperforming various baselines and providing evidence for the effectiveness of learning-guided planning in complex action spaces.