Expanding linear RNN state-transition matrix eigenvalues to include negatives improves state-tracking tasks and language modeling without added training or inference costs.

DeltaProduct is a new linear recurrent neural network architecture that uses products of generalized Householder matrices as state-transition matrices enabling better expressivity, state-tracking, and language modeling performance than DeltaNet.

Seamlessly track the visibility state of your document across different browsers. Know when your users switch to another window or tab.

Official implementation of "DZ-TDPO: Non-Destructive Temporal Alignment for Mutable State Tracking". SOTA on Multi-Session Chat with negligible alignment tax.

GoldEvidenceBench is a benchmark and harness for engineers/researchers building RAG/LLM systems. It generates long, noisy logs with oracle labels to separate retrieval, selection, attribution, and authority failures, and provides a self‑teaching loop to train selectors and measure evidence‑grounded state tracking.