In RL, the spectrum concentrates at all learning rates, while SFT expands it at moderate rates; RL also rotates singular subspaces faster than SFT at low rates.
Explore technology by industry and role
Newest additions first. Explore past and current notes by industry and role.
All notes
Newest additions first. Explore past and current notes by industry and role.
Structured reasoning chains enhance decision transparency in driving scenarios
A dynamic weighting mechanism fuses heterogeneous datasets to improve prediction accuracy
LightEMMA introduces a unified, lightweight evaluation protocol that assesses intrinsic driving capability without model-specific tuning or prompt engineering
A natural reflection pattern is used as a trigger to induce response delays in the model
Proposes a method to distinguish object instances using only camera input for 3D panoptic scene completion
A virtualization-based shared battery optimization system simultaneously reduces electricity costs and transformer aging
A model-free deep reinforcement learning method is proposed for joint real-time scheduling of heterogeneous battery systems and electric vehicles
Perceptual and localization uncertainty-aware association and registration are introduced
Integrates multi-sensor data with LLM-based reasoning for contextual adaptation
A coding agent reconstructs tasks from a single calibration episode and generates realistic simulation data
Sparse and localized feature maps generate interpretable decision pathways from raw inputs
Introduces a lightweight latent probe that enables early failure detection with calibrated accuracy and minimal overhead.
Point-plane projections enable effective feature learning from 2D representations using only LiDAR data
Automatically synthesizes realistic safety-critical scenarios from real-world geographic data.