Infoblock alpha
cs-ml/interpretability IB-arxiv-2607.28292-degradation-reduction measured

On a 4-bit quantized OpenLLaMA-3B with a UK financial corpus of 88,021 documents, CACHE-UK reduces knowledge degradation by 11-17% relative to adapted baselines under identical 4-bit constraints.

Active Evidence state is recorded explicitly and may be revised by a later event.

Observed quantities

corpus_documents 88021 · model_params_billion 3 · quant_bits 4 · degradation_reduction_pct_min 11 · degradation_reduction_pct_max 17

01

Primary source

  1. arxiv · 2607.28292 · v1 · abstract
Title
CACHE-UK: A Stability-Aware Memory Editor for Sequentially Updated Quantized LLMs in Finance
Authors
Anubhav Lakra, Yue Feng
Version
v1
Check
Support quote found word for word in the arXiv abstract v1, 2026-10-08
Supporting quote from the source
Evaluated on a 4-bit quantized OpenLLaMA-3B model with a curated UK financial corpus of 88,021 documents, CACHE-UK reduces knowledge degradation by 11-17% relative to adapted baselines under identical 4-bit constraints

02

Boundary of this block

No additional public limitation was recorded.

03

Connections

Outgoing

No outgoing public edges.

Incoming

No incoming public edges.