Evaluation and verification — from LLM judges to realistic agent behavior

Evaluation increasingly tests complete workflows, evidence integrity, ambiguity handling, and grounded execution. KEEP-Bench targets whether compression preserves reasoning-relevant evidence, while StreamEval and SpatialScout point to gaps in interactive video generation and active 3D mapping; the short reports provide little methodological or results detail. Into the Unknown, OR-License, and Nemotron 3.5 Super VL coverage continue to underline the gap between benchmark scores and reliable real-world outcomes.