AIAny
Icon for item

LatentPress: Context Compression Beyond Text and Vision

Compresses conversational histories and long documents into short sequences of continuous soft memory tokens that a frozen decoder can read directly without text reconstruction. Uses a small reader-matched writer that trains only a tiny adapter, achieving 4–16× compression and much faster write/read latencies.

Introduction

Most long-context systems still store or transmit summaries as text or images that must be decoded back into tokens before an LLM can use them. LatentPress flips that pipeline: instead of reconstructing human-readable context, it writes compact continuous tokens that a frozen decoder ingests directly, trading interpretable format for speed and model-aligned compression.

Key Findings
  • Direct soft-token interface: A reader-matched writer maps multi-turn conversational histories and long documents to short continuous token sequences that are concatenated into a frozen decoder's input-embedding layer, eliminating runtime text reconstruction.
  • Strong compression with a tiny trainable footprint: The writer trains only a small adapter (≈4.2M–26.2M parameters, ~0.1% of the decoder) and achieves 4–16× compression while preserving or improving downstream accuracy in many settings.
  • Empirical gains and efficiency: On LongMemEval, LatentPress reached 0.504 accuracy at 7.70× compression (versus 0.490 with uncompressed evidence). Encoding is fast (~43 ms per conversation) and reading from the compressed prefix is 5–9× faster than reading raw context or cached OCR.
Who it's for and trade-offs

Great fit if you need machine-facing, high-throughput long-context storage or retrieval for frozen LLM decoders—e.g., scalable memory systems, long-document QA backends, or cross-session agents where latency and storage matter. Look elsewhere if human-readable archives, explainable reconstructed text, or full generalization across arbitrary downstream decoders are required—LatentPress favors direct machine consumption and depends on reader/decoder compatibility.

Where it fits

LatentPress sits between text-summary and encoder-decoder latent compressors: it keeps the downstream decoder frozen (reducing fine-tuning cost) while using soft tokens as a portable, fast context interface applicable beyond just text and vision.

Information

  • Websitearxiv.org
  • OrganizationsCornell University, Iowa State University
  • AuthorsZhengze Zhou, Hejian Sang
  • Published date2026/09/01

More Items

Studies on-policy distillation (OPD) at the data-minimal limit by training on a single query, measuring state coverage and alignment dynamics, and showing OPD is often data-overfed but algorithm-starved.

Compresses KV cache for long-chain reasoning by keeping prompt tokens and evicting remaining entries uniformly at random per attention head; across four models and six reasoning tasks it matches the strongest prior evictor while delivering 32–43% higher vLLM throughput. Relies on prompt protection and redundancy across heads/text to retain reasoning traces; suitable when static memory budgets and higher serving throughput are priorities.

Shows that fully quantizing all 496 linear layers—including the recurrent Gated DeltaNet—of a hybrid 27B LLM to 4-bit NVFP4 W4A4 preserves benchmark accuracy while reducing model size to 17.5 GiB and improving prefill speed; includes a calibrated NVFP4 checkpoint.