CROSS introduces a pre-commitment localization layer using continuous SE(3) pose branches and Gaussian-mixture filtering to reject false matches, improving long-term robot relocalization and semantic navigation under severe scene changes.
TACT detects overthinking and overacting as linear drift axes in hidden states and applies activation steering to pull agents back toward calibrated behavior, boosting resolution rates up to 5.8 points and cutting steps by 26%.
MiroEval benchmarks multimodal deep research agents via process and outcome evaluation across 100 real-world tasks, finding process quality predicts outcomes and multimodal tasks reduce scores by 3, 10 points.