2025 · Speculative Decoding
Curvature-Guided Speculative Decoding for Large Language Models
Uses semantic curvature to predict which tokens will be accepted during speculative decoding, achieving 2.8× speedup over vanilla speculative decoding with no accuracy loss.
GIGI
Inference
DOI Pending
2025 · KV-Cache
Geometric KV-Cache Compression via Curvature-Aware Token Selection
Identifies which KV-cache entries to keep based on their geometric importance on the semantic manifold. Achieves 70% cache reduction with minimal perplexity increase.
MIRADOR
Memory
DOI Pending
2025 · Cognition
Spectral Geometry of Transformer Cognition
Analyzes the spectral properties of attention matrices through the lens of discrete Riemannian geometry. Shows that "reasoning" corresponds to geodesic paths on the semantic manifold.
Theory
DOI Pending
2025 · Holonomy
The Davis Conjecture on Semantic Coherence: Context Windows as Holonomy Groups
Proposes that context windows in LLMs function as holonomy groups on a semantic manifold. Context collapse occurs when the holonomy becomes trivial.
Conjecture
DOI Pending