PhD student, Institute of Artificial Intelligence, Beihang University
1 paper at NeurIPS 2025
We propose manifold steering that projects the steering direction of model overthinking on the low-dimensional activation manifold, effectively reducing output tokens while maintaining accuracy.
