Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) are two prominent post-training paradigms for refining the capabilities and aligning the behavior of Large Language Models (LLMs). Existing approaches that integrate SFT and RL often face the risk of disrupting established response patterns and inducing overfitting to expert data. To address this, we present a novel investigation into the unified view of SFT and RL through an off-policy versus on-policy lens. We propose CHORD, a framework for Controllable Harmonization of On- and Off-Policy Reinforcement Learning via Dynamic Weighting, which reframes SFT not as a separate stage but as a dynamically weighted auxiliary objective within the on-policy RL process. Based on an analysis of off-policy expert data's influence at both holistic and granular levels, we incorporate a dual-control mechanism in CHORD. Specifically, the framework first employs a global coefficient to holistically guide the transition from off-policy imitation to on-policy exploration, and then applies a token-wise weighting function that enables granular learning from the expert, which promotes on-policy exploration and mitigates disruption from off-policy data. We conduct extensive experiments on mathematical reasoning problems and practical tool-use tasks, providing empirical evidence that CHORD achieves a stable and efficient learning process. By effectively harmonizing off-policy expert data with on-policy exploration, CHORD demonstrates significant improvements over baselines. We release the implementation at https://github.com/modelscope/Trinity-RFT/tree/main/examples/mix_chord to inspire further research.

Training effective Vision-Language Models (VLMs) for GUI agents typically depends on large-scale annotated datasets, whose collection is both labor-intensive and error-prone. We introduce K-step GUI Transition, a self-supervised inverse dynamics task in which VLMs learn GUI dynamics by predicting the initial action that causes a transition between two GUI states. This approach eliminates the need for natural language instructions and enables scalable dataset construction from existing GUI trajectories or automated exploration. Building on this task, we propose GUI-Shift, a reinforcement learning (RL) framework that combines rule-based optimization with data filtering to improve VLM performance. We conduct extensive experiments using multiple VLM backbones across four benchmarks, spanning GUI task automation (AndroidControl, GUI Odyssey) and GUI grounding (ScreenSpot-v2, ScreenSpot-Pro). Our results show that training on GUI-Shift generalizes well to both GUI automation and grounding tasks, yielding up to an 11.2% increase in GUI automation accuracy. This study underscores the potential of self-supervised RL to leverage unlabeled GUI trajectories and offers a scalable alternative to training with annotated samples.

Today's generative models thrive with large amounts of supervised data and informative reward functions characterizing the quality of the generation. They work under the assumptions that the supervised data provides knowledge to pre-train the model, and the reward function provides dense information about how to further improve the generation quality and correctness. However, in the hardest instances of important problems, two problems arise: (1) the base generative model attains a near-zero reward signal, and (2) calls to the reward oracle are expensive. This setting poses a fundamentally different learning challenge than standard reward-based post-training. To address this, we propose BaNEL (Bayesian Negative Evidence Learning), an algorithm that post-trains the model using failed attempts only, while minimizing the number of reward evaluations (NREs). Our method is based on the idea that the problem of learning regularities underlying failures can be cast as another, in-loop generative modeling problem. We then leverage this model to assess whether new data resembles previously seen failures and steer the generation away from them. We show that BaNEL can improve model performance without observing a single successful sample on several sparse-reward tasks, outperforming existing novelty-bonus approaches by up to several orders of magnitude in success rate, while using fewer reward evaluations.

Multi-agent pathfinding (MAPF) remains a critical problem in robotics and autonomous systems, where agents must navigate shared spaces efficiently while avoiding conflicts. Traditional centralized algorithms that have global information, such as Conflict-Based Search (CBS), provide high-quality solutions but become computationally expensive in large-scale scenarios due to the combinatorial explosion of conflicts that need resolution. Conversely, distributed approaches that have local information, particularly learning-based methods, offer better scalability by operating with relaxed information availability, yet often at the cost of solution quality. To address these limitations, we propose a hybrid framework that combines decentralized path planning with a lightweight centralized coordinator. Our framework leverages reinforcement learning (RL) for decentralized planning, enabling agents to adapt their planning based on minimal, targeted alerts--such as static conflict-cell flags or brief conflict tracks--that are dynamically shared information from the central coordinator for effective conflict resolution. We empirically study the effect of the information available to an agent on its planning performance. Our approach reduces the inter-agent information sharing compared to fully centralized and distributed methods, while still consistently finding feasible, collision-free solutions--even in large-scale scenarios having higher agent counts.

Group Relative Policy Optimization (GRPO) has significantly advanced the reasoning ability of large language models (LLMs), particularly by boosting their mathematical performance. However, GRPO and related entropy-regularization methods still face challenges rooted in the sparse token rewards inherent to chain-of-thought (CoT). Current approaches often rely on undifferentiated token-level entropy adjustments, which frequently lead to entropy collapse or model collapse. In this work, we propose TEPO, a novel token-level framework that incorporates Markov Likelihood (sequence likelihood) links group-level rewards with tokens via token-level aggregation. Experiments show that TEPO consistently outperforms existing baselines across key metrics (including @k and accuracy). It not only sets a new state of the art on mathematical reasoning tasks but also significantly enhances training stability.

We close open theoretical gaps in Multi-Agent Imitation Learning (MAIL) by characterizing the limits of non-interactive MAIL and presenting the first interactive algorithm with near-optimal sample complexity. In the non-interactive setting, we prove a statistical lower bound that identifies the all-policy deviation concentrability coefficient as the fundamental complexity measure, and we show that Behavior Cloning (BC) is rate-optimal. For the interactive setting, we introduce a framework that combines reward-free reinforcement learning with interactive MAIL and instantiate it with an algorithm, MAIL-WARM. It improves the best previously known sample complexity from $\mathcal{O}(\varepsilon^{-8})$ to $\mathcal{O}(\varepsilon^{-2}),$ matching the dependence on $\varepsilon$ implied by our lower bound. Finally, we provide numerical results that support our theory and illustrate, in environments such as grid worlds, where Behavior Cloning fails to learn.

Abstract--Learning-based motion planning can quickly generate near-optimal trajectories. However, it often requires either large training datasets or costly collection of human demonstrations. This work proposes an alternative approach that quickly generates smooth, near-optimal collision-free 3D Cartesian trajectories from a single artificial demonstration. The demonstration is encoded as a Dynamic Movement Primitive (DMP) and iteratively reshaped using policy-based reinforcement learning to create a diverse trajectory dataset for varying obstacle configurations. This dataset is used to train a neural network that takes as inputs the task parameters describing the obstacle dimensions and location, derived automatically from a point cloud, and outputs the DMP parameters that generate the trajectory. The approach is validated in simulation and real-robot experiments, outperforming a RRT -Connect baseline in terms of computation and execution time, as well as trajectory length, while supporting multi-modal trajectory generation for different obstacle geometries and end-effector dimensions. Videos and the implementation code are available at https://github.com/ A motion planner for autonomous robotic manipulation should be able to quickly generate smooth optimal trajectories in different scenarios [1]. Sampling-based motion planners often struggle to quickly find near-optimal trajectories due to frequent online resampling [2], [3].

This paper explores the application of deep Q-learning to hedging at-the-money options on the S\&P~500 index. We develop an agent based on the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm, trained to simulate hedging decisions without making explicit model assumptions on price dynamics. The agent was trained on historical intraday prices of S\&P~500 call options across years 2004--2024, using a single time series of six predictor variables: option price, underlying asset price, moneyness, time to maturity, realized volatility, and current hedge position. A walk-forward procedure was applied for training, which led to nearly 17~years of out-of-sample evaluation. The performance of the deep reinforcement learning (DRL) agent is benchmarked against the Black--Scholes delta-hedging strategy over the same period. We assess both approaches using metrics such as annualized return, volatility, information ratio, and Sharpe ratio. To test the models' adaptability, we performed simulations across varying market conditions and added constraints such as transaction costs and risk-awareness penalties. Our results show that the DRL agent can outperform traditional hedging methods, particularly in volatile or high-cost environments, highlighting its robustness and flexibility in practical trading contexts. While the agent consistently outperforms delta-hedging, its performance deteriorates when the risk-awareness parameter is higher. We also observed that the longer the time interval used for volatility estimation, the more stable the results.

Current knowledge-enhanced large language models (LLMs) rely on static, pre-constructed knowledge bases that suffer from coverage gaps and temporal obsolescence, limiting their effectiveness in dynamic information environments. We present Agentic-KGR, a novel framework enabling co-evolution between LLMs and knowledge graphs (KGs) through multi-round reinforcement learning (RL). Our approach introduces three key innovations: (1) a dynamic schema expansion mechanism that systematically extends graph ontologies beyond pre-defined boundaries during training; (2) a retrieval-augmented memory system enabling synergistic co-evolution between model parameters and knowledge structures through continuous optimization; (3) a learnable multi-scale prompt compression approach that preserves critical information while reducing computational complexity through adaptive sequence optimization. Experimental results demonstrate substantial improvements over supervised baselines and single-round RL approaches in knowledge extraction tasks. When integrated with GraphRAG, our method achieves superior performance in downstream QA tasks, with significant gains in both accuracy and knowledge coverage compared to existing methods.

Learning from demonstrations enables experts to teach robots complex tasks using interfaces such as kinesthetic teaching, joystick control, and sim-to-real transfer. However, these interfaces often constrain the expert's ability to demonstrate optimal behavior due to indirect control, setup restrictions, and hardware safety. For example, a joystick can move a robotic arm only in a 2D plane, even though the robot operates in a higher-dimensional space. As a result, the demonstrations collected by constrained experts lead to suboptimal performance of the learned policies. This raises a key question: Can a robot learn a better policy than the one demonstrated by a constrained expert? We address this by allowing the agent to go beyond direct imitation of expert actions and explore shorter and more efficient trajectories. We use the demonstrations to infer a state-only reward signal that measures task progress, and self-label reward for unknown states using temporal interpolation. Our approach outperforms common imitation learning in both sample efficiency and task completion time. On a real WidowX robotic arm, it completes the task in 12 seconds, 10x faster than behavioral cloning, as shown in real-robot videos on https://sites.google.com/view/constrainedexpert .