KV-Cache Grafting – Boosting frozen 12B LLMs to 93.3% AIME accuracy

Name: KV-Cache Grafting – Boosting frozen 12B LLMs to 93.3% AIME accuracy
Availability: InStock
Author: Corbenic

by Corbenic·Jul 17, 2026·2 points·0 comments

Visit Project View on HN

AI Analysis

●●●●GemWizardryBig BrainZero to One

Frozen 12B model hits 93.3% AIME by grafting verified KV states, not retraining.

Strengths

•Byte-exact restore with SHA-256 verification and zero KL divergence across fifty samples.
•Answers 8 unsolved AIME problems in 61 tokens versus 401k token budget, 8,700x energy savings.
•Extends usable context from 32k to 2.8M tokens at zero extra accelerator memory.

Weaknesses

•Engine is proprietary with no code release, limiting reproducibility and adoption.
•Only works on same-architecture machines for byte-identical state transfer.

Similar Projects

AI/ML●●●Banger

Tier – Adaptive tool routing that makes small LLMs 10pt more accurate

Makes 1.5B models 10% more accurate by hiding 90% of tool descriptions.

Big BrainNiche Gem

pranabsarkar

433mo ago

AI/ML●●Solid

Lumisift – improves data retention in RAG from ~40% to 87%

Information density scoring beats semantic similarity for scientific RAG retrieval.

Big BrainNiche Gem

benmora

103mo ago

AI/ML●Mid

Is the "frozen weights" paradigm the main bottleneck for AGI?

The repo openly rejects the 'frozen weights' assumption and tries to prototype an assistant that rewires online — you can see the scaffolding in files like autonomous_ai.py, view_graph.py, a configs folder, a streamlit_apps dir and chroma_data. That's an interesting, contrarian direction, but the project is clearly early-stage: the UI and repo layout are tidy, yet there’s little in-repo evidence of benchmarks, experiments, or reproducible results to back the big claim.

Bold BetRabbit Hole

8lamaster8

105mo ago

AI/ML○Pass