LatentPress: Context Compression Beyond Text and Vision
LatentPress stores context as soft memory tokens a frozen language model can read directly, skipping text or image reconstruction.
The paper reports 4x-16x compression using a small writer adapter matched to the reader model. On LongMemEval, it reaches 0.504 accuracy at 7.70x compression, above the 0.490 result for uncompressed evidence in the abstract. It also beats text summaries and OCR-based compression in the reported comparisons. The authors say writing takes 43ms per conversation and reading is 5x-9x faster than raw context or cached OCR. ArXiv · AI/CL/LG's note
The paper reports 4x-16x compression using a small writer adapter matched to the reader model. On LongMemEval, it reaches 0.504 accuracy at 7.70x compression, above the 0.490 result for uncompressed evidence in the abstract. It also beats text summaries and OCR-based compression in the reported comparisons. The authors say writing takes 43ms per conversation and reading is 5x-9x faster than raw context or cached OCR. ArXiv · AI/CL/LG's note
score 6