Asaf Yehudai

Teaching agents *what* unstated information to hunt down—rather than just *when* to act—enables small models to beat 15× larger baselines at multi-hop task completion while asking fewer questions.

Systematic gaps in AI evaluation reporting are exposed, revealing inconsistencies that hinder reliable comparisons across thousands of models and benchmarks.

Agents that excel on traditional benchmarks may crumble under the pressure of newly synthesized tasks, revealing the limitations of current evaluation methods.