The Approximation Rank of Softmax Attention: Sharp Geometric Laws and Robust Interaction Dimension
arXiv · 2026
Yuhe Sui
I study the mathematical structure of learning and decision-making, and the design of reliable AI systems for research.
Research interests
- Attention & Representation
- Mathematical structure, approximation and representation complexity, intrinsic geometry, and invariances of attention mechanisms and learned representations.
- Learning & Decision-Making
- Reinforcement learning, policy optimisation, in-context learning of algorithms, stochastic control, and structured decision procedures.
- Scientific AI Systems
- Reliable long-horizon AI systems for research, with emphasis on hierarchical execution, verification, provenance, recoverability, and human control.
- Mathematical & Quantitative Methods
- Probability, optimisation, discrete mathematics, statistical learning, and quantitative applications.
Selected publications
All publications →NTU-QRS AlphaAllocator: Bounded Transformer News Overlays for Chinese Macro-Asset Allocation
NLPCC 2026 · 2026
Shared Task 4: Track 1 winner; Track 2 6th.
No matching results. Try another keyword or clear the filters.
Selected research
All research →AI-Native Workflow
A hierarchical long-horizon research workflow centred on durable state, explicit handoffs, independent verification, recovery, and bounded agent authority.
The Approximation Rank of Softmax Attention
Sharp Geometric Laws and Robust Interaction Dimension
Studies the approximation complexity of softmax attention through support geometry and softmax-visible interaction dimension.
NTU-QRS AlphaAllocator
A bounded transformer-news overlay for Chinese macro-asset allocation with explicit portfolio controls and frozen evaluation.
No matching results. Try another keyword or clear the filters.