Live intelligence

Latest AI updates

1 updates
Research

TANSAR: Editable Correlation Matrix Memories for Multi-Hop Knowledge Editing

The paper analyzes Correlation Matrix Memories under correlated keys using frame-theoretic coherence, deriving capacity, deletion-collateral, and attribution bounds. It introduces TANSAR, which traverses a directed acyclic association graph, prunes invalid paths with sound Horn constraints, and injects verified-path representations into a frozen language model's residual stream through a selective-risk gate. In multi-hop knowledge editing, Horn pruning reduced answer contradictions from 25.1% to 0.5%, and an end-to-end edit transaction took 13.6 ms on CPU and 2.37 ms on GPU. The symbolic component matched gold labels across 600 ProofWriter theories and reached 94.0% agreement on FOLIO.

Correlation Matrix Memoriesknowledge editingneuro-symbolic reasoning