Reproduce “Transformer-FSQ: Discrete Latents for Extreme KV-Cache Compression” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Andras Szecsenyi, Thomas L. Lee, Vaisakh Shaj, Elliot J. Crowley, Amos Storkey
Transformer-FSQ compresses a language model’s key-value cache by storing one learned, quantized code per token in place of separate keys and values. Its finite scalar quantization is trained into inserted projections while the backbone stays frozen. On Qwen3-8B-Base, it scores 0.49 on RULER at 32× compression, versus 0.24 for token eviction; evidence covers retrieval and document tasks, not general reasoning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.