Chronological feed of everything captured from Brandon Amos.
Contrary to the dominant narrative that compounding model errors limit model-based RL, this ICML 2026 paper demonstrates that search itself is the problem — it can degrade performance even with highly accurate models. The root cause is overestimation bias in value functions, not prediction horizon errors. The authors show that taking the minimum over an ensemble of value functions effectively mitigates this bias, enabling search to achieve state-of-the-art results across standard benchmarks.