Matteo Merler

Physics engines do not need to be replaced by black-box neural networks—patching them with Bayesian-inferred code enables robots to actively discover missing physical mechanisms with drastically fewer real-world interactions.

Selective querying of VLMs can lead to autonomous policies that outperform their teachers while drastically reducing reliance on expensive model calls.

Simple prompting methods consistently outperform advanced dense supervision techniques, challenging the current assumptions in LLM training strategies.

Unlabeled data can be transformed into powerful training examples, boosting classification performance by nearly 20 points without the need for costly annotations.