From 7c716e01565eba67a3e63fc4845899db166071d6 Mon Sep 17 00:00:00 2001 From: TheTom Date: Mon, 6 Apr 2026 22:23:31 -0500 Subject: [PATCH] Remove nihilistau reference from asymmetric K/V paper Co-Authored-By: tturney@psyguard.ai Co-Authored-By: Claude Opus 4.6 (1M context) --- docs/papers/asymmetric-kv-compression.md | 6 ------ 1 file changed, 6 deletions(-) diff --git a/docs/papers/asymmetric-kv-compression.md b/docs/papers/asymmetric-kv-compression.md index e538b22..186d048 100644 --- a/docs/papers/asymmetric-kv-compression.md +++ b/docs/papers/asymmetric-kv-compression.md @@ -344,12 +344,6 @@ These findings have been independently confirmed by multiple researchers: - Originally hit corrupt multilingual output on Madreag's fork with symmetric turbo3. Madreag's fork needs to pick up the same fix separately - First end-user confirmation of the head_dim=256 fix on Blackwell (SM 120) consumer hardware -**@nihilistau** — [Vilenkin basis spectral analysis, Phi-2 + turbo3](https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16442465) (2026-04-03): -- Independent harmonic analysis of K and V cache vectors in the Vilenkin (p-adic/mixed-radix) basis suggests a spectral explanation for why asymmetric compression works -- K vectors appear to concentrate energy in a small number of universal spectral indices (6 indices at 100% universality across all positions). V vectors appear to spread energy diffusely across the entire spectrum (max 20% universality, no dominant modes) -- K and V appear to occupy disjoint spectral bands with zero overlap, suggesting they compress independently without interference -- If confirmed across architectures, this would provide a mathematical basis for the empirical observation that V compression is free while K precision is critical -- **Caveat:** tested on Phi-2 only. Needs validation on Qwen, Llama, and Gemma architectures before drawing general conclusions. The spectral structure may be model-dependent ---