Contextual bandit algorithms have transformed modern experimentation by enabling real-time adaptation for personalized treatment and efficient use of data. Yet these advantages create challenges for statistical inference due to adaptivity. A fundamental property that supports valid inference is policy convergence, meaning that action-selection probabilities converge in probability given the context. Convergence ensures replicability of adaptive experiments and stability of online algorithms. In this paper, we highlight a previously overlooked issue: widely used algorithms such as LinUCB may fail to converge when the reward model is misspecified, and such non-convergence creates fundamental obstacles for statistical inference. This issue is practically important, as misspecified models -- such as linear approximations of complex dynamic system -- are often employed in real-world adaptive experiments to balance bias and variance. Motivated by this insight, we propose and analyze a broad class of algorithms that are guaranteed to converge even under model misspecification. Building on this guarantee, we develop a general inference framework based on an inverse-probability-weighted Z-estimator (IPW-Z) and establish its asymptotic normality with a consistent variance estimator. Simulation studies confirm that the proposed method provides robust and data-efficient confidence intervals, and can outperform existing approaches that exist only in the special case of offline policy evaluation. Taken together, our results underscore the importance of designing adaptive algorithms with built-in convergence guarantees to enable stable experimentation and valid statistical inference in practice.

As AI systems become increasingly autonomous, reliably aligning their decision-making with human preferences is essential. Inverse reinforcement learning (IRL) offers a promising approach to infer preferences from demonstrations. These preferences can then be used to produce an apprentice policy that performs well on the demonstrated task. However, in domains like autonomous driving or robotics, where errors can have serious consequences, we need not just good average performance but reliable policies with formal guarantees -- yet obtaining sufficient human demonstrations for reliability guarantees can be costly. Active IRL addresses this challenge by strategically selecting the most informative scenarios for human demonstration. We introduce PAC-EIG, an information-theoretic acquisition function that directly targets probably-approximately-correct (PAC) guarantees for the learned policy -- providing the first such theoretical guarantee for active IRL with noisy expert demonstrations. Our method maximises information gain about the regret of the apprentice policy, efficiently identifying states requiring further demonstration. We also present Reward-EIG as an alternative when learning the reward itself is the primary objective. Focusing on finite state-action spaces, we prove convergence bounds, illustrate failure modes of prior heuristic methods, and demonstrate our method's advantages experimentally.

In reinforcement learning with sparse rewards, demonstrations can accelerate learning, but determining when to imitate them remains challenging. We propose Smooth Policy Regularisation from Demonstrations (SPReD), a framework that addresses the fundamental question: when should an agent imitate a demonstration versus follow its own policy? SPReD uses ensemble methods to explicitly model Q-value distributions for both demonstration and policy actions, quantifying uncertainty for comparisons. We develop two complementary uncertainty-aware methods: a probabilistic approach estimating the likelihood of demonstration superiority, and an advantage-based approach scaling imitation by statistical significance. Unlike prevailing methods (e.g. Q-filter) that make binary imitation decisions, SPReD applies continuous, uncertainty-proportional regularisation weights, reducing gradient variance during training. Despite its computational simplicity, SPReD achieves remarkable gains in experiments across eight robotics tasks, outperforming existing approaches by up to a factor of 14 in complex tasks while maintaining robustness to demonstration quality and quantity. Our code is available at https://github.com/YujieZhu7/SPReD.

Online reinforcement learning (RL) has been central to post-training language models, but its extension to diffusion models remains challenging due to intractable likelihoods. Recent works discretize the reverse sampling process to enable GRPO-style training, yet they inherit fundamental drawbacks, including solver restrictions, forward-reverse inconsistency, and complicated integration with classifier-free guidance (CFG). We introduce Diffusion Negative-aware FineTuning (DiffusionNFT), a new online RL paradigm that optimizes diffusion models directly on the forward process via flow matching. DiffusionNFT contrasts positive and negative generations to define an implicit policy improvement direction, naturally incorporating reinforcement signals into the supervised learning objective. This formulation enables training with arbitrary black-box solvers, eliminates the need for likelihood estimation, and requires only clean images rather than sampling trajectories for policy optimization. DiffusionNFT is up to $25\times$ more efficient than FlowGRPO in head-to-head comparisons, while being CFG-free. For instance, DiffusionNFT improves the GenEval score from 0.24 to 0.98 within 1k steps, while FlowGRPO achieves 0.95 with over 5k steps and additional CFG employment. By leveraging multiple reward models, DiffusionNFT significantly boosts the performance of SD3.5-Medium in every benchmark tested.

This study introduces a novel method for revealing human internal attention patterns from gameplay data alone, leveraging offline attention techniques from reinforcement learning (RL). We propose contextualized, task-relevant (CTR) attention networks, which generate attention maps from both human and RL agent gameplay in Atari environments. To evaluate whether the human CTR maps reveal internal attention, we validate our model by quantitative and qualitative comparison to the agent maps as well as to a temporally integrated overt attention (TIOA) model based on human eye-tracking data. Our results show that human CTR maps are more sparse than the agent ones and align better with the TIOA maps. Following a qualitative visual comparison we conclude that they likely capture patterns of internal attention. As a further application, we use these maps to guide RL agents, finding that human internal attention-guided agents achieve slightly improved and more stable learning compared to baselines. This work advances the understanding of human-agent attention differences and provides a new approach for extracting and validating internal attention from behavioral data.

Atomic data determined by analysis of observed atomic spectra are essential for plasma diagnostics. For each low-ionisation open d- and f-subshell atomic species, around $10^3$ fine structure level energies can be determined through years of analysis of $10^4$ observable spectral lines. We propose the automation of this task by casting the analysis procedure as a Markov decision process and solving it by graph reinforcement learning using reward functions learned on historical human decisions. In our evaluations on existing spectral line lists and theoretical calculations for Co II and Nd II-III, hundreds of level energies were computed within hours, agreeing with published values in 95% of cases for Co II and 54-87% for Nd II-III. As the current efficiency in atomic fine structure determination struggles to meet growing atomic data demands from astronomy and fusion science, our new artificial intelligence approach sets the stage for closing this gap.

Deep reinforcement learning (RL) is emerging as a viable strategy for automated cyber defense (ACD). The traditional RL approach represents networks as a list of computers in various states of safety or threat. Unfortunately, these models are forced to overfit to specific network topologies, rendering them ineffective when faced with even small environmental perturbations. In this work, we frame ACD as a two-player context-based partially observable Markov decision problem with observations represented as attributed graphs. This approach allows our agents to reason through the lens of relational inductive bias. Agents learn how to reason about hosts interacting with other system entities in a more general manner, and their actions are understood as edits to the graph representing the environment. By introducing this bias, we will show that our agents can better reason about the states of networks and zero-shot adapt to new ones. We show that this approach outperforms the state-of-the-art by a wide margin, and makes our agents capable of defending never-before-seen networks against a wide range of adversaries in a variety of complex, and multi-agent environments.

Designing effective reward functions remains a major challenge in reinforcement learning (RL), often requiring considerable human expertise and iterative refinement. Recent advances leverage Large Language Models (LLMs) for automated reward design, but these approaches are limited by hallucinations, reliance on human feedback, and challenges with handling complex, multi-step tasks. In this work, we introduce Reward Evolution with Graph-of-Thoughts (RE-GoT), a novel bi-level framework that enhances LLMs with structured graph-based reasoning and integrates Visual Language Models (VLMs) for automated rollout evaluation. RE-GoT first decomposes tasks into text-attributed graphs, enabling comprehensive analysis and reward function generation, and then iteratively refines rewards using visual feedback from VLMs without human intervention. Extensive experiments on 10 RoboGen and 4 ManiSkill2 tasks demonstrate that RE-GoT consistently outperforms existing LLM-based baselines. On RoboGen, our method improves average task success rates by 32.25%, with notable gains on complex multi-step tasks. On ManiSkill2, RE-GoT achieves an average success rate of 93.73% across four diverse manipulation tasks, significantly surpassing prior LLM-based approaches and even exceeding expert-designed rewards. Our results indicate that combining LLMs and VLMs with graph-of-thoughts reasoning provides a scalable and effective solution for autonomous reward evolution in RL.

Figure 1: Our Solo12 loco-manipulation policy is able to execute and smoothly transition between locomotion and manipulation to reach a specified target. Abstract-- Although humanoid and quadruped robots provide a wide range of capabilities, current control methods, such as Deep Reinforcement Learning, focus mainly on single skills. This approach is inefficient for solving more complicated tasks where high-level goals, physical robot limitations and desired motion style might all need to be taken into account. A more effective approach is to first train a multipurpose motion policy that acquires low-level skills through imitation, while providing latent space control over skill execution. Then, this policy can be used to efficiently solve downstream tasks. This method has already been successful for controlling characters in computer graphics. In this work, we apply the approach to humanoid and quadrupedal loco-manipulation by imitating either simple synthetic motions or kinematically retargeted dog motions. We extend the original formulation to handle constraints, ensuring deployment safety, and use a diffusion discriminator for better imitation quality.

This paper investigates artificial intelligence (AI) methodologies for the synthesis and transpilation of permutation circuits across generic topologies. Our approach uses Reinforcement Learning (RL) techniques to achieve near-optimal synthesis of permutation circuits up to 25 qubits. Rather than developing specialized models for individual topologies, we train a foundational model on a generic rectangular lattice, and employ masking mechanisms to dynamically select subsets of topologies during the synthesis. This enables the synthesis of permutation circuits on any topology that can be embedded within the rectangular lattice, without the need to re-train the model. In this paper we show results for 5x5 lattice and compare them to previous AI topology-oriented models and classical methods, showing that they outperform classical heuristics, and match previous specialized AI models, and performs synthesis even for topologies that were not seen during training. We further show that the model can be fine tuned to strengthen the performance for selected topologies of interest. This methodology allows a single trained model to efficiently synthesize circuits across diverse topologies, allowing its practical integration into transpilation workflows.