We propose that reinforcement learning (RL) from partial expert demonstrations is not merely a training heuristic, but a promising framework for solving complex sequence generation tasks. Supervised fine-tuning (SFT) relies on dense ground-truth labels, which become increasingly costly as sequence length grows. RL, on the other hand, struggles with sparse rewards and a combinatorially large output space. We address this by introducing adaptive backtracking (AdaBack), a per-sample curriculum learning algorithm that reveals only a partial prefix of the target output during training. The supervision length is adjusted dynamically for each sample based on the model's past reward signal, allowing it to incrementally learn to complete reasoning chains by conditioning on correct partial solutions. We investigate this intermediate regime between SFT and RL and argue that per-sample curriculum learning is more than a trade-off between efficiency and generality, it can succeed in tasks with long sequences of latent dependencies where SFT and RL both fail to generalize. Using a synthetic task with latent parity constraints, we show that our adaptive curriculum over partial answers reliably solves problems that are otherwise intractable. On mathematical reasoning benchmarks (MATH, GSM8k), we find that curriculum learning enables models to solve problems that RL alone cannot, acquiring new reasoning capabilities through incremental exposure to partial solutions.

Supporting decision-making has long been a central vision in the field of spatio-temporal intelligence. While prior work has improved the timeliness and accuracy of spatio-temporal forecasting, converting these forecasts into actionable strategies remains a key challenge. A main limitation is the decoupling of the prediction and the downstream decision phases, which can significantly degrade the downstream efficiency. For example, in emergency response, the priority is successful resource allocation and intervention, not just incident prediction. To this end, it is essential to propose an Adaptive Spatio-Temporal Early Decision model (ASTER) that reforms the forecasting paradigm from event anticipation to actionable decision support. This framework ensures that information is directly used for decision-making, thereby maximizing overall effectiveness. Specifically, ASTER introduces a new Resource-aware Spatio-Temporal interaction module (RaST) that adaptively captures long- and short-term dependencies under dynamic resource conditions, producing context-aware spatiotemporal representations. To directly generate actionable decisions, we further design a Preference-oriented decision agent (Poda) based on multi-objective reinforcement learning, which transforms predictive signals into resource-efficient intervention strategies by deriving optimal actions under specific preferences and dynamic constraints. Experimental results on four benchmark datasets demonstrate the state-of-the-art performance of ASTER in improving both early prediction accuracy and resource allocation outcomes across six downstream metrics.

Reinforcement learning (RL) for auto-bidding has shifted from using simplistic offline simulators (Simulation-based RL Bidding, SRLB) to offline RL on fixed real datasets (Offline RL Bidding, ORLB). However, ORLB policies are limited by the dataset's state space coverage, offering modest gains. While SRLB expands state coverage, its simulator-reality gap risks misleading policies. This paper introduces Model-based RL Bidding (MRLB), which learns an environment model from real data to bridge this gap. MRLB trains policies using both real and model-generated data, expanding state coverage beyond ORLB. To ensure model reliability, we propose: 1) A permutation equivariant model architecture for better generalization, and 2) A robust offline Q-learning method that pessimistically penalizes model errors. These form the Permutation Equivariant Model-based Offline RL (PE-MORL) algorithm. Real-world experiments show that PE-MORL outperforms state-of-the-art auto-bidding methods.

In real-world recommendation systems, users would engage in variety scenarios, such as homepages, search pages, and related recommendation pages. Each of these scenarios would reflect different aspects users focus on. However, the user interests may be inconsistent in different scenarios, due to differences in decision-making processes and preference expression. This variability complicates unified modeling, making multi-scenario learning a significant challenge. To address this, we propose a novel reinforcement learning approach that models user preferences across scenarios by modeling user interest evolution across multiple scenarios. Our method employs Double Q-learning to enhance next-item prediction accuracy and optimizes contrastive learning loss using Q-value to make model performance better. Experimental results demonstrate that our approach surpasses state-of-the-art methods in multi-scenario recommendation tasks. Our work offers a fresh perspective on multi-scenario modeling and highlights promising directions for future research.

--Vision-Language-Action models (VLA) have demonstrated remarkable capabilities and promising potential in solving complex robotic manipulation tasks. However, their substantial parameter sizes and high inference latency pose significant challenges for real-world deployment, particularly on resource-constrained robotic platforms. T o address this issue, we begin by conducting an extensive empirical study to explore the effectiveness of model compression techniques when applied to VLAs. Building on the insights gained from these preliminary experiments, we propose RLRC, a three-stage recovery method for compressed VLAs, including structured pruning, performance recovery based on SFT and RL, and further quantization. RLRC achieves up to an 8 reduction in memory usage and a 2.3 improvement in inference throughput, while maintaining or even surpassing the original VLA's task success rate. Extensive experiments show that RLRC consistently outperforms existing compression baselines, demonstrating strong potential for on-device deployment of VLAs. I. INTRODUCTION Recent advances in the field of robot learning have demonstrated new breakthroughs in both the accuracy and generalization of robotic policies for task execution. Since the introduction of RT -2 [1], Vision-Language-Action (VLA) models have attracted increasing attention. These models, built upon large foundation models, exhibit strong generalization capabilities, suggesting a promising path toward the development of general-purpose robots capable of performing a wide range of manipulation tasks. VLA models leverage the general knowledge embedded in pretrained Vision-Language Models (VLMs), while possessing the capability to comprehend language instructions, perceive the visual environment, and generate appropriate actions [2][3][4].

-- Residual Reinforcement Learning (RL) is a popular approach for adapting pretrained policies by learning a lightweight residual policy that provides corrective actions. While Residual RL is more sample-efficient than finetuning the entire base policy, existing methods struggle with sparse rewards and are designed for deterministic base policies. We propose two improvements to Residual RL that further enhance its sample efficiency and make it suitable for stochastic base policies. First, we leverage uncertainty estimates of the base policy to focus exploration on regions in which the base policy is not confident. Second, we propose a simple modification to off-policy residual learning that allows it to observe base actions and better handle stochastic base policies. We evaluate our method with both Gaussian-based and Diffusion-based stochastic base policies on tasks from Robosuite and D4RL, and compare against state-of-the-art finetuning methods, demo-augmented RL methods, and other residual RL methods. Our algorithm significantly outperforms existing baselines in a variety of simulation benchmark environments. We also deploy our learned polices in the real world to demonstrate their robustness with zero-shot sim-to-real transfer . Residual Reinforcement Learning is popular in robotics to improve the performance of pretrained policies by training a separate policy that outputs residual actions [1], [2].

We introduce a novel reinforcement learning (RL) framework that treats distribution parameters as actions, redefining the boundary between agent and environment. This reparameterization makes the new action space continuous, regardless of the original action type (discrete, continuous, mixed, etc.). Under this new parameterization, we develop a generalized deterministic policy gradient estimator, Distribution Parameter Policy Gradient (DPPG), which has lower variance than the gradient in the original action space. Although learning the critic over distribution parameters poses new challenges, we introduce interpolated critic learning (ICL), a simple yet effective strategy to enhance learning, supported by insights from bandit settings. Building on TD3, a strong baseline for continuous control, we propose a practical DPPG-based actor-critic algorithm, Distribution Parameter Actor-Critic (DPAC). Empirically, DPAC outperforms TD3 in MuJoCo continuous control tasks from OpenAI Gym and DeepMind Control Suite, and demonstrates competitive performance on the same environments with discretized action spaces.

Reinforcement learning is emerging as a primary driver for improving language model reasoning capabilities. A fundamental question is whether current reinforcement learning algorithms -- such as Group Relative Policy Optimization (GRPO), the de facto standard algorithm used to improve language model reasoning -- merely sharpen the base model's distribution around problems it can already solve. We investigate this question in the context of formal theorem proving, which has access to a perfect verifier. We identify a degenerate rank bias in GRPO in which highly probable trajectories are reinforced and rare ones are neglected. This results in distribution sharpening: the model can solve some problems with fewer samples, but underperforms simply sampling more solutions from the original model. To overcome GRPO's rank bias we introduce unlikeliness reward, a simple method for explicitly up-weighting rare but correct solutions. We show that unlikeliness reward mitigates rank bias and improves pass@$N$ across a large range of $N$ in both synthetic and real theorem proving settings. We also uncover an unexpected link between rank bias and a seemingly mundane hyperparameter -- the number of updates per batch -- that leads to a second, complementary mitigation. We combine our insights into a revised GRPO training recipe for formal theorem proving, yielding an open pipeline that achieves competitive performance to DeepSeek-Prover-V1.5-RL on the miniF2F-test benchmark. We release our implementation at https://github.com/AndreHe02/rewarding-unlikely-release

We contribute the first provable guarantees of global convergence to Nash equilibria (NE) in two-player zero-sum convex Markov games (cMGs) by using independent policy gradient methods. Convex Markov games, recently defined by Gemp et al. (2024), extend Markov decision processes to multi-agent settings with preferences that are convex over occupancy measures, offering a broad framework for modeling generic strategic interactions. However, even the fundamental min-max case of cMGs presents significant challenges, including inherent nonconvexity, the absence of Bellman consistency, and the complexity of the infinite horizon. We follow a two-step approach. First, leveraging properties of hidden-convex--hidden-concave functions, we show that a simple nonconvex regularization transforms the min-max optimization problem into a nonconvex-proximal Polyak-Lojasiewicz (NC-pPL) objective. Crucially, this regularization can stabilize the iterates of independent policy gradient methods and ultimately lead them to converge to equilibria. Second, building on this reduction, we address the general constrained min-max problems under NC-pPL and two-sided pPL conditions, providing the first global convergence guarantees for stochastic nested and alternating gradient descent-ascent methods, which we believe may be of independent interest.

To use robots in more unstructured environments, we have to accommodate for more complexities. Robotic systems need more awareness of the environment to adapt to uncertainty and variability. Although cameras have been predominantly used in robotic tasks, the limitations that come with them, such as occlusion, visibility and breadth of information, have diverted some focus to tactile sensing. In this thesis, we explore the use of tactile sensing to determine the pose of the object using the temporal features. We then use reinforcement learning with tactile collisions to reduce the number of attempts required to grasp an object resulting from positional uncertainty from camera estimates. Finally, we use information provided by these tactile sensors to a reinforcement learning agent to determine the trajectory to take to remove an object from a restricted passage while reducing training time by pertaining from human examples.