The problem, the contribution, and an interactive explanation—grounded in the original research.
2026-10-18
The authors introduce Windowed Interaction RL (WIRL) for autonomous long-horizon machine learning engineering optimization, demonstrating improved performance on MLE-Bench Lite from 37.9% to 72.7% medal rate.
Read briefing and explore the method →2026-10-18
The authors develop a robust reinforcement learning (RL) framework for learning adaptive treatment policies that generalize across patients and remain reliable under errors in the measured biomarker and shifts in the disease dynamics.
Read briefing and explore the method →2026-10-17
Explore how control variates and baselines improve the stability and efficiency of policy gradient algorithms, with practical implementations in modern RL frameworks.
Read briefing and explore the method →2026-10-17
The authors introduce T2‑Z, a test‑time optimization method that augments a robotic world model with a learnable environment latent to improve generalization to new environments.
Read briefing and explore the method →2026-10-16
A robotic manipulation benchmark measures task completion separately from satisfaction of language-expressible preferences.
Read briefing and explore the method →