Tsinghua Latent Compaction Hits 96.7% RULER at 32x KV Compression via Test-Time Training

Tsinghua's Latent Compaction treats the document-specific KV cache as a trainable neural network, fitting attention outputs and log-normalizers to the full-cache model with base weights frozen and using only generic prompts about the document. On RULER at 32x compression it hits 96.7% accuracy versus 40.0% for KVZip+, with transfer across Qwen3 and MiniCPM5 model families and sizes.