The functionality of Large Language Model (LLM) agents is primarily determined by two capabilities: action planning and answer summarization. The former, action planning, is the core capability that dictates an agent's performance. However, prevailing training paradigms employ end-to-end, multi-objective optimization that jointly trains both capabilities. This paradigm faces two critical challenges: imbalanced optimization objective allocation and scarcity of verifiable data, making it difficult to enhance the agent's planning capability. To address these challenges, we propose Reinforcement Learning with Tool-use Rewards (RLTR), a novel framework that decouples the training process to enable a focused, single-objective optimization of the planning module. Crucially, RLTR introduces a reward signal based on tool-use completeness to directly evaluate the quality of tool invocation sequences. This method offers a more direct and reliable training signal than assessing the final response content, thereby obviating the need for verifiable data. Our experiments demonstrate that RLTR achieves an 8%-12% improvement in planning performance compared to end-to-end baselines. Moreover, this enhanced planning capability, in turn, translates to a 5%-6% increase in the final response quality of the overall agent system.

In reinforcement learning with human feedback (RLHF), reward models can efficiently learn and amplify latent biases within multimodal datasets, which can lead to imperfect policy optimization through flawed reward signals and decreased fairness. Bias mitigation studies have often applied passive constraints, which can fail under causal confounding. Here, we present a counterfactual reward model that introduces causal inference with multimodal representation learning to provide an unsupervised, bias-resilient reward signal. The heart of our contribution is the Counterfactual Trust Score, an aggregated score consisting of four components: (1) counterfactual shifts that decompose political framing bias from topical bias; (2) reconstruction uncertainty during counterfactual perturbations; (3) demonstrable violations of fairness rules for each protected attribute; and (4) temporal reward shifts aligned with dynamic trust measures. We evaluated the framework on a multimodal fake versus true news dataset, which exhibits framing bias, class imbalance, and distributional drift. Following methodologies similar to unsupervised drift detection from representation-based distances [1] and temporal robustness benchmarking in language models [2], we also inject synthetic bias across sequential batches to test robustness. The resulting system achieved an accuracy of 89.12% in fake news detection, outperforming the baseline reward models. More importantly, it reduced spurious correlations and unfair reinforcement signals. This pipeline outlines a robust and interpretable approach to fairness-aware RLHF, offering tunable bias reduction thresholds and increasing reliability in dynamic real-time policy making.

We present a generative optimization approach for learning game-playing agents, where policies are represented as Python programs and refined using large language models (LLMs). Our method treats decision-making policies as self-evolving code, with current observation as input and an in-game action as output, enabling agents to self-improve through execution traces and natural language feedback with minimal human intervention. Applied to Atari games, our game-playing Python program achieves performance competitive with deep reinforcement learning (RL) baselines while using significantly less training time and much fewer environment interactions. This work highlights the promise of programmatic policy representations for building efficient, adaptable agents capable of complex, long-horizon reasoning.

Cyber defense requires automating defensive decision-making under stealthy, deceptive, and continuously evolving adversarial strategies. The FlipIt game provides a foundational framework for modeling interactions between a defender and an advanced adversary that compromises a system without being immediately detected. In FlipIt, the attacker and defender compete to control a shared resource by performing a Flip action and paying a cost. However, the existing FlipIt frameworks rely on a small number of heuristics or specialized learning techniques, which can lead to brittleness and the inability to adapt to new attacks. To address these limitations, we introduce PoolFlip, a multi-agent gym environment that extends the FlipIt game to allow efficient learning for attackers and defenders. Furthermore, we propose Flip-PSRO, a multi-agent reinforcement learning (MARL) approach that leverages population-based training to train defender agents equipped to generalize against a range of unknown, potentially adaptive opponents. Our empirical results suggest that Flip-PSRO defenders are $2\times$ more effective than baselines to generalize to a heuristic attack not exposed in training. In addition, our newly designed ownership-based utility functions ensure that Flip-PSRO defenders maintain a high level of control while optimizing performance.

Offline reinforcement learning (RL) often deals with suboptimal data when collecting large expert datasets is unavailable or impractical. This limitation makes it difficult for agents to generalize and achieve high performance, as they must learn primarily from imperfect or inconsistent trajectories. A central challenge is therefore how to best leverage scarce expert demonstrations alongside abundant but lower-quality data. We demonstrate that incorporating even a tiny amount of expert experience can substantially improve RL agent performance. We introduce Re:Frame (Retrieving Experience From Associative Memory), a plug-in module that augments a standard offline RL policy (e.g., Decision Transformer) with a small external Associative Memory Buffer (AMB) populated by expert trajectories drawn from a separate dataset. During training on low-quality data, the policy learns to retrieve expert data from the Associative Memory Buffer (AMB) via content-based associations and integrate them into decision-making; the same AMB is queried at evaluation. This requires no environment interaction and no modifications to the backbone architecture. On D4RL MuJoCo tasks, using as few as 60 expert trajectories (0.1% of a 6000-trajectory dataset), Re:Frame consistently improves over a strong Decision Transformer baseline in three of four settings, with gains up to +10.7 normalized points. These results show that Re:Frame offers a simple and data-efficient way to inject scarce expert knowledge and substantially improve offline RL from low-quality datasets.

--Simulated environments have proven invaluable in Autonomous Cyber Operations (ACO) where Reinforcement Learning (RL) agents can be trained without the computational overhead of emulation. These environments must accurately represent cybersecurity scenarios while producing the necessary signals to support RL training. In this study, we present a framework where we first extend CybORG's Cage Challenge 2 environment by implementing three new actions: Patch, Isolate, and Unisolate, to better represent the capabilities available to human operators in real-world settings. We then propose a design for agent development where we modify the reward signals and the agent's feature space to enhance training performance. T o validate these modifications, we train DQN and PPO agents in the updated environment. Our study demonstrates that CybORG can be extended with additional realistic functionality, while maintaining its ability to generate informative training signals for RL agents.

Learning robust driving policies from large-scale, real-world datasets is a central challenge in autonomous driving, as online data collection is often unsafe and impractical. While Behavioral Cloning (BC) offers a straightforward approach to imitation learning, policies trained with BC are notoriously brittle and suffer from compounding errors in closed-loop execution. This work presents a comprehensive pipeline and a comparative study to address this limitation. We first develop a series of increasingly sophisticated BC baselines, culminating in a Transformer-based model that operates on a structured, entity-centric state representation. While this model achieves low imitation loss, we show that it still fails in long-horizon simulations. We then demonstrate that by applying a state-of-the-art Offline Reinforcement Learning algorithm, Conservative Q-Learning (CQL), to the same data and architecture, we can learn a significantly more robust policy. Using a carefully engineered reward function, the CQL agent learns a conservative value function that enables it to recover from minor errors and avoid out-of-distribution states. In a large-scale evaluation on 1,000 unseen scenarios from the Waymo Open Motion Dataset, our final CQL agent achieves a 3.2x higher success rate and a 7.4x lower collision rate than the strongest BC baseline, proving that an offline RL approach is critical for learning robust, long-horizon driving policies from static expert data.

With the recent rapid advancement of Agentic Intelligence, agentic tool use in LLMs has become increasingly important. During multi-turn interactions between agents and users, the dynamic, uncertain, and stochastic nature of user demands poses significant challenges to the agent's tool invocation capabilities. Agents are no longer expected to simply call tools to deliver a result; rather, they must iteratively refine their understanding of user needs through communication while simultaneously invoking tools to resolve user queries. Existing reinforcement learning (RL) approaches for tool use lack the integration of genuinely dynamic users during the RL training process. To bridge this gap, we introduce MUA-RL (Multi-turn User-interacting Agent Reinforcement Learning for agentic tool use), a novel reinforcement learning framework that, for the first time in the field of agentic tool use, integrates LLM-simulated users into the reinforcement learning loop. MUA-RL aims to enable autonomous learning of models to communicate with users efficiently and use various tools to solve practical problems in dynamic multi-turn interactions. Evaluations are done on several multi-turn tool-using benchmarks (see Figure 1). Specifically, MUA-RL-32B achieves 67.3 on TAU2 Retail, 45.4 on TAU2 Airline, 28.3 on TAU2 Telecom, 28.4 on BFCL-V3 Multi Turn, and 82.5 on ACEBench Agent -- outperforming or matching the performance of larger open-source models such as DeepSeek-V3-0324 and Qwen3-235B-A22B in non-thinking settings.

As the embodiment gap between a robot and a human narrows, new opportunities arise to leverage datasets of humans interacting with their surroundings for robot learning. We propose a novel technique for training sensorimotor policies with reinforcement learning by imitating predictive models of human motions. Our key insight is that the motion of keypoints on human-inspired robot end-effectors closely mirrors the motion of corresponding human body keypoints. This enables us to use a model trained to predict future motion on human data \emph{zero-shot} on robot data. We train sensorimotor policies to track the predictions of such a model, conditioned on a history of past robot states, while optimizing a relatively sparse task reward. This approach entirely bypasses gradient-based kinematic retargeting and adversarial losses, which limit existing methods from fully leveraging the scale and diversity of modern human-scene interaction datasets. Empirically, we find that our approach can work across robots and tasks, outperforming existing baselines by a large margin. In addition, we find that tracking a human motion model can substitute for carefully designed dense rewards and curricula in manipulation tasks. Code, data and qualitative results available at https://jirl-upenn.github.io/track_reward/.

Dynamic rotational maneuvers, such as front flips, inherently involve large angular momentum generation and intense impact forces, presenting major challenges for reinforcement learning and sim-to-real transfer. In this work, we propose a general framework for learning and deploying impact-rich, rotation-intensive behaviors through centroidal velocity-based rewards and actuator-aware sim-to-real techniques. We identify that conventional link-level reward formulations fail to induce true whole-body rotation and introduce a centroidal angular velocity reward that accurately captures system-wide rotational dynamics. To bridge the sim-to-real gap under extreme conditions, we model motor operating regions (MOR) and apply transmission load regularization to ensure realistic torque commands and mechanical robustness. Using the one-leg hopper front flip as a representative case study, we demonstrate the first successful hardware realization of a full front flip. Our results highlight that incorporating centroidal dynamics and actuator constraints is critical for reliably executing highly dynamic motions. A supplementary video is available at: https://youtu.be/atMAVI4s1RY