We investigate a failure mode that arises during the training of reasoning models, where the diversity of generations begins to collapse, leading to suboptimal test-time scaling. Notably, the Pass@1 rate reliably improves during supervised finetuning (SFT), but Pass@k rapidly deteriorates. Surprisingly, a simple intervention of interpolating the weights of the latest SFT checkpoint with an early checkpoint, otherwise known as WiSE-FT, almost completely recovers Pass@k while also improving Pass@1. The WiSE-FT variant achieves better test-time scaling (Best@k, majority vote) and achieves superior results with less data when tuned further by reinforcement learning. Finally, we find that WiSE-FT provides complementary performance gains that cannot be achieved only through diversity-inducing decoding strategies, like temperature scaling. We formalize a bias-variance tradeoff of Pass@k with respect to the expectation and variance of Pass@1 over the test distribution. We find that WiSE-FT can reduce bias and variance simultaneously, while temperature scaling inherently trades off between bias and variance.

Offline goal-conditioned reinforcement learning (GCRL) trains policies that reach user-specified goals at test time, providing a simple, unsupervised, domain-agnostic way to extract diverse behaviors from unlabeled, reward-free datasets. Nonetheless, long-horizon decision making remains difficult for GCRL agents due to temporal credit assignment and error accumulation, and the offline setting amplifies these effects. To alleviate this issue, we introduce Test-Time Graph Search (TTGS), a lightweight planning approach to solve the GCRL task. TTGS accepts any state-space distance or cost signal, builds a weighted graph over dataset states, and performs fast search to assemble a sequence of subgoals that a frozen policy executes. When the base learner is value-based, the distance is derived directly from the learned goal-conditioned value function, so no handcrafted metric is needed. TTGS requires no changes to training, no additional supervision, no online interaction, and no privileged information, and it runs entirely at inference. On the OGBench benchmark, TTGS improves success rates of multiple base learners on challenging locomotion tasks, demonstrating the benefit of simple metric-guided test-time planning for offline GCRL.

Real-world robotic agents must act under partial observability and long horizons, where key cues may appear long before they affect decision making. However, most modern approaches rely solely on instantaneous information, without incorporating insights from the past. Standard recurrent or transformer models struggle with retaining and leveraging long-term dependencies: context windows truncate history, while naive memory extensions fail under scale and sparsity. We propose ELMUR (External Layer Memory with Update/Rewrite), a transformer architecture with structured external memory. Each layer maintains memory embeddings, interacts with them via bidirectional cross-attention, and updates them through an Least Recently Used (LRU) memory module using replacement or convex blending. ELMUR extends effective horizons up to 100,000 times beyond the attention window and achieves a 100% success rate on a synthetic T-Maze task with corridors up to one million steps. In POPGym, it outperforms baselines on more than half of the tasks. On MIKASA-Robo sparse-reward manipulation tasks with visual observations, it nearly doubles the performance of strong baselines. These results demonstrate that structured, layer-local external memory offers a simple and scalable approach to decision making under partial observability.

Compliance with maritime traffic rules is essential for the safe operation of autonomous vessels, yet training reinforcement learning (RL) agents to adhere to them is challenging. The behavior of RL agents is shaped by the training scenarios they encounter, but creating scenarios that capture the complexity of maritime navigation is non-trivial, and real-world data alone is insufficient. To address this, we propose a falsification-driven RL approach that generates adversarial training scenarios in which the vessel under test violates maritime traffic rules, which are expressed as signal temporal logic specifications. Our experiments on open-sea navigation with two vessels demonstrate that the proposed approach provides more relevant training scenarios and achieves more consistent rule compliance.

Edge devices have limited resources, which inevitably leads to situations where stream processing services cannot satisfy their needs. While existing autoscaling mechanisms focus entirely on resource scaling, Edge devices require alternative ways to sustain the Service Level Objectives (SLOs) of competing services. To address these issues, we introduce a Multi-dimensional Autoscaling Platform (MUDAP) that supports fine-grained vertical scaling across both service- and resource-level dimensions. MUDAP supports service-specific scaling tailored to available parameters, e.g., scale data quality or model size for a particular service. To optimize the execution across services, we present a scaling agent based on Regression Analysis of Structural Knowledge (RASK). The RASK agent efficiently explores the solution space and learns a continuous regression model of the processing environment for inferring optimal scaling actions. We compared our approach with two autoscalers, the Kubernetes VPA and a reinforcement learning agent, for scaling up to 9 services on a single Edge device. Our results showed that RASK can infer an accurate regression model in merely 20 iterations (i.e., observe 200s of processing). By increasingly adding elasticity dimensions, RASK sustained the highest request load with 28% less SLO violations, compared to baselines.

We introduce a tool for rigorous and automated verification of large language model (LLM)- based policies in memoryless sequential decision-making tasks. Given a Markov decision process (MDP) representing the sequential decision-making task, an LLM policy, and a safety requirement expressed as a PCTL formula, our approach incrementally constructs only the reachable portion of the MDP guided by the LLM's chosen actions. Each state is encoded as a natural language prompt, the LLM's response is parsed into an action, and reachable successor states by the policy are expanded. The resulting formal model is checked with Storm to determine whether the policy satisfies the specified safety property. In experiments on standard grid world benchmarks, we show that open source LLMs accessed via Ollama can be verified when deterministically seeded, but generally underperform deep reinforcement learning baselines. Our tool natively integrates with Ollama and supports PRISM-specified tasks, enabling continuous benchmarking in user-specified sequential decision-making tasks and laying a practical foundation for formally verifying increasingly capable LLMs.

In this work, we introduce dual goal representations for goal-conditioned reinforcement learning (GCRL). A dual goal representation characterizes a state by "the set of temporal distances from all other states"; in other words, it encodes a state through its relations to every other state, measured by temporal distance. This representation provides several appealing theoretical properties. First, it depends only on the intrinsic dynamics of the environment and is invariant to the original state representation. Second, it contains provably sufficient information to recover an optimal goal-reaching policy, while being able to filter out exogenous noise. Based on this concept, we develop a practical goal representation learning method that can be combined with any existing GCRL algorithm. Through diverse experiments on the OGBench task suite, we empirically show that dual goal representations consistently improve offline goal-reaching performance across 20 state- and pixel-based tasks.

Recent progress in vision and language foundation models has significantly advanced multimodal understanding, reasoning, and generation, inspiring a surge of interest in extending such capabilities to embodied settings through vision-language-action (VLA) models. Yet, most VLA models are still trained with supervised fine-tuning (SFT), which struggles to generalize under distribution shifts due to error accumulation. Reinforcement learning (RL) offers a promising alternative by directly optimizing task performance through interaction, but existing attempts remain fragmented and lack a unified platform for fair and systematic comparison across model architectures and algorithmic designs. To address this gap, we introduce RLinf-VLA, a unified and efficient framework for scalable RL training of VLA models. The system adopts a highly flexible resource allocation design that addresses the challenge of integrating rendering, training, and inference in RL+VLA training. In particular, for GPU-parallelized simulators, RLinf-VLA implements a novel hybrid fine-grained pipeline allocation mode, achieving a 1.61x-1.88x speedup in training. Through a unified interface, RLinf-VLA seamlessly supports diverse VLA architectures (e.g., OpenVLA, OpenVLA-OFT), multiple RL algorithms (e.g., PPO, GRPO), and various simulators (e.g., ManiSkill, LIBERO). In simulation, a unified model achieves 98.11\% across 130 LIBERO tasks and 97.66\% across 25 ManiSkill tasks. Beyond empirical performance, our study distills a set of best practices for applying RL to VLA training and sheds light on emerging patterns in this integration. Furthermore, we present preliminary deployment on a real-world Franka robot, where RL-trained policies exhibit stronger generalization than those trained with SFT. We envision RLinf-VLA as a foundation to accelerate and standardize research on embodied intelligence.

The Forward-Forward (FF) Algorithm is a recently proposed learning procedure for neural networks that employs two forward passes instead of the traditional forward and backward passes used in backpropagation. However, FF remains largely confined to supervised settings, leaving a gap at domains where learning signals can be yielded more naturally such as RL. In this work, inspired by FF's goodness function using layer activity statistics, we introduce Action-conditioned Root mean squared Q-Functions (ARQ), a novel value estimation method that applies a goodness function and action conditioning for local RL using temporal difference learning. Despite its simplicity and biological grounding, our approach achieves superior performance compared to state-of-the-art local backprop-free RL methods in the MinAtar and the DeepMind Control Suite benchmarks, while also outperforming algorithms trained with backpropagation on most tasks. Code can be found at https://github.com/agentic-learning-ai-lab/arq.

The objective of this study is to develop a model-free workspace trajectory planner for space manipulators using a Twin Delayed Deep Deterministic Policy Gradient (TD3) agent to enable safe and reliable debris capture. A local control strategy with singularity avoidance and manipulability enhancement is employed to ensure stable execution. The manipulator must simultaneously track a capture point on a non-cooperative target, avoid self-collisions, and prevent unintended contact with the target. To address these challenges, we propose a curriculum-based multi-critic network where one critic emphasizes accurate tracking and the other enforces collision avoidance. A prioritized experience replay buffer is also used to accelerate convergence and improve policy robustness. The framework is evaluated on a simulated seven-degree-of-freedom KUKA LBR iiwa mounted on a free-floating base in Matlab/Simulink, demonstrating safe and adaptive trajectory generation for debris removal missions.