Synthetic Monitoring Environments for Reinforcement Learning

As large language model agents advance beyond software engineering (SWE) tasks toward machine learning engineering (MLE), verifying agent behavior becomes orders of magnitude more expensive: while SWE tasks can be verified via fast-executing unit tests, MLE verification requires running full ML pipelines -- data preprocessing, model training, and metric evaluation -- on large datasets at each rollout step, rendering trajectory-wise on-policy reinforcement learning (RL) prohibitively slow. Existing approaches retreat to supervised fine-tuning (SFT) or offline proxy rewards, sacrificing the exploration and generalization benefits of on-policy RL. We observe that sandbox data size is the primary source of this bottleneck. Based on this insight, we introduce SandMLE, a multi-agent framework that generates diverse, verifiable synthetic MLE environments from a small number of seed tasks, preserving the structural and technical complexity of real-world problems while constraining datasets to micro-scale (each task is paired with only 50-200 training samples). Through extensive experiments, we show that SandMLE reduces execution time by over 13 times, enabling large-scale, on-policy trajectory-wise RL for the first time in the MLE domain. On MLE-bench-lite, SandMLE yields significant gains over SFT baselines across Qwen3-8B, 14B, and 30B-A3B, with relative medal rate improvements ranging from 20.3% to 66.9%. Furthermore, the trained policy generalizes across unseen agentic scaffolds, achieving up to 32.4% better HumanRank score on MLE-Dojo.

Reliable reinforcement learning (RL) agents must maintain operational integrity amidst sensor malfunctions, dynamic disturbances, and slow environmental shifts. The detection of out-of-distribution conditions is pivotal to determining when an agent's observations, transitions, or trajectory dynamics deviate from the assumptions underpinning its policy training. Current out-of-distribution (OOD) detection benchmarks typically evaluate image classifiers or static low-dimensional datasets, failing to account for the complex, action-dependent temporal structure inherent in RL trajectories. To address this gap, we present OOD-RL-Bench, a comprehensive and extensible framework designed to evaluate OOD detectors against categories of anomalies injected into RL trajectories. Detectors and anomaly injectors are integrated through shared interfaces and configuration, which allows new scoring methods and perturbation families to be evaluated without modification of the core benchmark loop. We evaluate the utility of the framework using a Deep Q-Network policy within the LunarLander-v3 environment. We assess the performance of each detector across a suite of anomaly types using matched-time AUROC, matched-time AUPRC, matched-time false-positive rate, detection delay, and segmented-onset metrics. Our analysis reveals significant performance variance across anomaly types: observation perturbations and regime switches are identified with high accuracy by several methods, while observation delay and action-conditioned dynamics remain difficult even when post-onset anomaly scores are compared against clean scores from the same timesteps. We make the framework, trained policy checkpoint, and complete results publicly available as a reproducible artefact.

Benchmarking in reinforcement learning is usually relative: algorithms are ranked by returns on environments with unknown optima. Distance to optimality remains unobservable, while conclusions may depend on reward design, stochasticity, hyperparameters, environment selection, and simulator details. We develop a ground-truth-first framework based on converse optimality, constructing each task with a certified optimal value function and policy withheld from learners and used only for evaluation. The framework unifies two modes: sensitivity benchmarking samples certified environments from continuous-time generators to assess robustness to task variation, while oracle-referenced comparison uses native discrete-time generators to evaluate policies via optimality gap, regret, and policy error. Continuous-time models support physically interpretable sensitivity studies but introduce dependence on the simulator step ($Δt$) and numerical scheme. Native discrete-time models align the environment, oracle, and learner on the same transition law, eliminating external integration choices as a confounder. We reproduce quantitative results covering optimizer sensitivity and oracle-referenced comparisons on serial planar arms, dynamically extended nonholonomic vehicles, and quadrotor-inspired tasks. The methodology extends beyond smooth affine Gaussian systems: virtual-input pullback handles non-affine raw-action interfaces and finite-covariance non-Gaussian disturbances. A non-smooth converse optimality formulation enables certified benchmarks with kinked value functions and switching surfaces. For partially observed evaluation, an information-state formulation reduces benchmarking to converse optimality on a sufficient statistic and provides an information-penalty reference. Code, fixtures, and diagnostics are available on the public project page.