Generic Triple-Latent Compression with Gated Associative Retrieval

In three linesStudy of triple-latent sequence models maintaining running token state and compressed pair-memory pathway to capture higher-order token interactions. Improvements on byte-level WikiText-2 and MiniMind benchmark, with gated associative retrieval extension improving recall but remaining seed-sensitive and slow.

Summary generated by Claude — human-verified