We present CompassMax-V3-Thinking, a hundred-billion-scale MoE reasoning model trained with a new RL framework built on one principle: each prompt must matter. Scaling RL to this size exposes critical inefficiencies-zero-variance prompts that waste rollouts, unstable importance sampling over long horizons, advantage inversion from standard reward models, and systemic bottlenecks in rollout processing. To overcome these challenges, we introduce several unified innovations: (1) Multi-Stage Zero-Variance Elimination, which filters out non-informative prompts and stabilizes group-based policy optimization (e.g. GRPO) by removing wasted rollouts; (2) ESPO, an entropy-adaptive optimization method that balances token-level and sequence-level importance sampling to maintain stable learning dynamics; (3) a Router Replay strategy that aligns training-time MoE router decisions with inference-time behavior to mitigate train-infer discrepancies, coupled with a reward model adjustment to prevent advantage inversion; (4) a high-throughput RL system with FP8-precision rollouts, overlapped reward computation, and length-aware scheduling to eliminate performance bottlenecks. Together, these contributions form a cohesive pipeline that makes RL on hundred-billion-scale MoE models stable and efficient. The resulting model delivers strong performance across both internal and public evaluations.

When should an autonomous agent commit resources to a task? We introduce the Agent Capability Problem (ACP), a framework for predicting whether an agent can solve a problem under resource constraints. Rather than relying on empirical heuristics, ACP frames problem-solving as information acquisition: an agent requires $\Itotal$ bits to identify a solution and gains $\Istep$ bits per action at cost $\Cstep$, yielding an effective cost $\Ceff = (\Itotal/\Istep), \Cstep$ that predicts resource requirements before search. We prove that $\Ceff$ lower-bounds expected cost and provide tight probabilistic upper bounds. Experimental validation shows that ACP predictions closely track actual agent performance, consistently bounding search effort while improving efficiency over greedy and random strategies. The framework generalizes across LLM-based and agentic workflows, linking principles from active learning, Bayesian optimization, and reinforcement learning through a unified information-theoretic lens. \

This study presents a systematic comparison of three Reinforcement Learning (RL) algorithms (PPO, GRPO, and DAPO) for improving complex reasoning in large language models (LLMs). Our main contribution is a controlled transfer-learning evaluation: models are first fine-tuned on the specialized Countdown Game and then assessed on a suite of general-purpose reasoning benchmarks. Across all tasks, RL-trained models outperform their corresponding base models, although the degree of improvement differs by benchmark. Our parametric analysis offers practical guidance for RL-based LLM training. Increasing the group size in GRPO and DAPO leads to more stable training dynamics and higher accuracy, while the impact of the KL-penalty coefficient is non-monotonic. Additionally, we find that the Dynamic Sampling (DS) component in DAPO does not improve performance; in fact, the best overall results are achieved with DAPO when DS is disabled.

Abstract--We investigate model-based reinforcement learning in contextual Markov decision processes (C-MDPs) in which the context is unobserved and induces confounding in the offline dataset. In such settings, conventional model-learning methods are fundamentally inconsistent, as the transition and reward mechanisms generated under a behavioral policy do not correspond to the interventional quantities required for evaluating a state-based policy. T o address this issue, we adapt a proximal off-policy evaluation approach that identifies the confounded reward expectation using only observable state-action-reward trajectories under mild invertibility conditions on proxy variables. When combined with a behavior-averaged transition model, this construction yields a surrogate MDP whose Bellman operator is well defined and consistent for state-based policies, and which integrates seamlessly with the maximum causal entropy (MaxCausalEnt) model-learning framework. The proposed formulation enables principled model learning and planning in confounded environments where contextual information is unobserved, unavailable, or impractical to collect.

Abstract-- For full-size humanoid robots, even with recent advances in reinforcement learning-based control, achieving reliable locomotion on complex terrains, such as long staircases, remains challenging. In such settings, limited perception, ambiguous terrain cues, and insufficient adaptation of gait timing can cause even a single misplaced or mistimed step to result in rapid loss of balance. We introduce a perceptive locomotion framework that merges terrain sensing, gait regulation, and whole-body control into a single reinforcement learning policy. A downward-facing depth camera mounted under the base observes the support region around the feet, and a compact U-Net reconstructs a dense egocentric height map from each frame in real time, operating at the same frequency as the control loop. The perceptual height map, together with proprioceptive observations, is processed by a unified policy that produces joint commands and a global stepping-phase signal, allowing gait timing and whole-body posture to be adapted jointly to the commanded motion and local terrain geometry. We further adopt a single-stage successive teacher-student training scheme for efficient policy learning and knowledge transfer . Experiments conducted on a 31-DoF, 1.65 m humanoid robot demonstrate robust locomotion in both simulation and real-world settings, including forward and backward stair ascent and descent, as well as crossing a 46 cm gap.

Effective traffic control is essential for mitigating congestion in transportation networks. Conventional traffic management strategies, including route guidance, ramp metering, and traffic signal control, often rely on state feedback controllers, used for their simplicity and reactivity; however, they lack the adaptability required to cope with complex and time-varying traffic dynamics. This paper proposes a multi-agent reinforcement learning framework in which each agent adaptively tunes the parameters of a state feedback traffic controller, combining the reactivity of state feedback controllers with the adaptability of reinforcement learning. By tuning parameters at a lower frequency rather than directly determining control actions at a high frequency, the reinforcement learning agents achieve improved training efficiency while maintaining adaptability to varying traffic conditions. The multi-agent structure further enhances system robustness, as local controllers can operate independently in the event of partial failures. The proposed framework is evaluated on a simulated multi-class transportation network under varying traffic conditions. Results show that the proposed multi-agent framework outperforms the no control and fixed-parameter state feedback control cases, while performing on par with the single-agent RL-based adaptive state feedback control, with a much better resilience to partial failures.

Abstract--In bus arrival time prediction, the process of organizing road infrastructure network data into homogeneous entities is known as segmentation. Segmenting a road network is widely recognized as the first and most critical step in developing an arrival time prediction system, particularly for auto-regressive-based approaches. Traditional methods typically employ a uniform segmentation strategy, which fails to account for varying physical constraints along roads, such as road conditions, intersections, and points of interest, thereby limiting prediction efficiency. In this paper, we propose a Reinforcement Learning (RL)-based approach to efficiently and adaptively learn non-uniform road segments for arrival time prediction. Our method decouples the prediction process into two stages: 1) Nonuniform road segments are extracted based on their impact scores using the proposed RL framework; and 2) A linear prediction model is applied to the selected segments to make predictions. This method ensures optimal segment selection while maintaining computational efficiency, offering a significant improvement over traditional uniform approaches. Furthermore, our experimental results suggest that the linear approach can even achieve better performance than more complex methods. Extensive experiments demonstrate the superiority of the proposed method, which not only enhances efficiency but also improves learning performance on large-scale benchmarks.

Adaptive morphogenetic robots adapt their morphology and control policies to meet changing tasks and environmental conditions. Many such systems leverage soft components, which enable shape morphing but also introduce simulation and control challenges. Soft-body simulators remain limited in accuracy and computational tractability, while rigid-body simulators cannot capture soft-material dynamics. Here, we present a surrogate compliance modeling approach: rather than explicitly modeling soft-body physics, we introduce indirect variables representing soft-material deformation within a rigid-body simulator. We validate this approach using our amphibious robotic turtle, a quadruped with soft morphing limbs designed for multi-environment locomotion. By capturing deformation effects as changes in effective limb length and limb center of mass, and by applying reinforcement learning with extensive randomization of these indirect variables, we achieve reliable policy learning entirely in a rigid-body simulation. The resulting gaits transfer directly to hardware, demonstrating high-fidelity sim-to-real performance on hard, flat substrates and robust, though lower-fidelity, transfer on rheologically complex terrains. The learned closed-loop gaits exhibit unprecedented terrestrial maneuverability and achieve an order-of-magnitude reduction in cost of transport compared to open-loop baselines. Field experiments with the robot further demonstrate stable, multi-gait locomotion across diverse natural terrains, including gravel, grass, and mud.

Currently, there is a noticeable lack of AI in the medical field to support doctors in treating heterogenous brain tumors such as Glioblastoma Multiforme (GBM), the deadliest human cancer in the world with a five-year survival rate of just 5.1%. This project develops an AI system offering the only end-to-end solution by aiding doctors with both diagnosis and treatment planning. In the diagnosis phase, a sequential decision-making framework consisting of 4 classification models (Convolutional Neural Networks and Support Vector Machine) are used. Each model progressively classifies the patient's brain into increasingly specific categories, with the final step being named diagnosis. For treatment planning, an RL system consisting of 3 generative models is used. First, the resection model (diffusion model) analyzes the diagnosed GBM MRI and predicts a possible resection outcome. Second, the radiotherapy model (Spatio-Temporal Vision Transformer) generates an MRI of the brain's progression after a user-defined number of weeks. Third, the chemotherapy model (Diffusion Model) produces the post-treatment MRI. A survival rate calculator (Convolutional Neural Network) then checks if the generated post treatment MRI has a survival rate within 15% of the user defined target. If not, a feedback loop using proximal policy optimization iterates over this system until an optimal resection location is identified. When compared to existing solutions, this project found 3 key findings: (1) Using a sequential decision-making framework consisting of 4 small diagnostic models reduced computing costs by 22.28x, (2) Transformers regression capabilities decreased tumor progression inference time by 113 hours, and (3) Applying Augmentations resembling Real-life situations improved overall DICE scores by 2.9%. These results project to increase survival rates by 0.9%, potentially saving approximately 2,250 lives.

Phishing is a cybercrime in which individuals are deceived into revealing personal information, often resulting in financial loss. These attacks commonly occur through fraudulent messages, misleading advertisements, and compromised legitimate websites. This study proposes a Quantile Regression Deep Q-Network (QR-DQN) approach that integrates RoBERTa semantic embeddings with handcrafted lexical features to enhance phishing detection while accounting for uncertainties. Unlike traditional DQN methods that estimate single scalar Q-values, QR-DQN leverages quantile regression to model the distribution of returns, improving stability and generalization on unseen phishing data. A diverse dataset of 105,000 URLs was curated from PhishTank, OpenPhish, Cloudflare, and other sources, and the model was evaluated using an 80/20 train-test split. The QR-DQN framework achieved a test accuracy of 99.86%, precision of 99.75%, recall of 99.96%, and F1-score of 99.85%, demonstrating high effectiveness. Compared to standard DQN with lexical features, the hybrid QR-DQN with lexical and semantic features reduced the generalization gap from 1.66% to 0.04%, indicating significant improvement in robustness. Five-fold cross-validation confirmed model reliability, yielding a mean accuracy of 99.90% with a standard deviation of 0.04%. These results suggest that the proposed hybrid approach effectively identifies phishing threats, adapts to evolving attack strategies, and generalizes well to unseen data.