LatentPress: 4.2M adapter matches raw context at 7.7× compression

A 4.2M–26.2M parameter adapter can compress conversational history and long documents into vectors a frozen language model reads directly, reaching 0.504 accuracy on the LongMemEval memory-QA benchmark at 7.70× compression, slightly above the 0.490 the same model scores on the uncompressed evidence.

That is the claim in LatentPress: Context Compression Beyond Text and Vision, a new preprint from Zhengze Zhou and Hejian Sang. "Compressed context is usually carried as human-readable text or as rendered images that must be decoded, even when its consumer is a language model," the paper argues, proposing "continuous memory tokens that a frozen decoder reads directly through its input-embedding interface, with no text reconstruction at inference."

The margin over weaker baselines is wide. At comparable settings, text summaries score 0.184 on LongMemEval and OCR-based compression lands between 0.312 and 0.426. LatentPress writes a conversation in 43 milliseconds and reads 5-9× faster than raw context or cached OCR, with the trained adapter sitting at roughly 0.1% of the decoder's parameters.

The picture on LongBench-QA is more mixed. In-domain writers match or exceed raw-context reading at 4-8× compression, and 16× brings a performance decline. The 0.504-vs-0.490 gap on LongMemEval is narrow, the abstract does not name the frozen decoder, and no per-domain LongBench numbers appear.