SpectralQuant
A KV-cache compression method for LLM inference that improves on an existing method by exploiting a structural property shared across models, described in a NeurIPS 2026 submission.
SpectralQuant is a KV-cache compression method for large language model inference, using calibrated eigenbasis rotation and water-filled bit allocation to compress the cache more aggressively than a prior method while relying on a structural property observed across multiple models and architecture families. The repository is the canonical, consolidated version of the project, combining the original public release with an expanded evidence layer developed afterward, and is written up in a NeurIPS 2026 submission with its LaTeX sources, figures, and a submission audit of round-by-round revisions included.
Earlier manuscripts and technical reports are kept in the repository for traceability but are explicitly marked as no longer the current paper, and the repository documents its own renaming history so that older identifiers used in archived experiment results, storage volume names, and file paths can still be traced back to the current project.