institute
active
institute:mats-ml-alignment-theory-scholarsMATS (ML Alignment & Theory Scholars)
Program under which work was conducted.
Neighborhood — ranked by edge-count
Thinkers (1)
thinker
- Matthew Weardenaffiliated_with
Papers (1)
paper
Institutes (1)
institute
- Program that supported Tim Hua and Andrew Qin during this research.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- The primary contribution of the paper: a bidirectional causal method that learns rotation matrices for each model to uncover and compare causally relevant latent subspaces across neural networks.
- Architecture of Mixtral-8x7B; uses sparse expert routing affecting how hidden states are computed across layers.
- MAS reduces number of required alignment matrices for n-model comparison from n(n-1) or n^2 (stitching) to nfinding0.733Key computational efficiency advantage of MAS over traditional model stitching for multi-model comparisons.
- Skyline for RAVEL achieving 0.87.
- Machine Learning Research Group, Department of Engineering Science, University of Oxfordinstitute0.725Research group where Philip Ball is based
- Independent component alignment for multi-task learning.
- Source paper for the MT-Bench evaluation benchmark used to assess capabilities post-SOO fine-tuning
- Alignment map ϕ(h)=W_orth*h using orthogonal matrix; assumes linear representation hypothesis