Format failure hides inside accuracy: the GGUF quantization level that breaks is not predictable from its label

Abstract Accuracy benchmarks score a non-answer and a wrong answer alike. We test whether that convention survives the 8-bit to 4-bit step of GGUF quantization. Across four instruction-tuned models from 0.5B to 3B, four tasks at 800 items each, and every published precision level, 70,394 paired responses separate the responses that answer from the responses that do not. Format failure is concentrated and not monotone in bit-width. On ARC-Challenge, Qwen2.5-0.5B produces no usable answer on 1.6% of responses at Q8_0, 5.4% at Q5_K_M, 19.4% at Q4_K_M, 9.1% at Q3_K_M and 10.4% at Q2_K: the level that fails is Q4_K_M, while the more compressed levels answer more often. Qwen2.5-3B fails at Q3_K_M instead, at 30.6% against 3.1% at Q4_K_M, and Gemma-3-1B and Llama-3.2-3B never exceed 0.1% at any level. Counting non-answers as errors raises the measured cost of the 4-bit step at 0.5B from 9.0% to 10.9%; on responses that do answer, the cost falls with scale, 9.0%, 6.8% and 3.1% against a 0.4% gain at Llama-3.2-3B. A label does not tell a practitioner whether the artifact will answer, and the level that fails cannot be inferred from its neighbours.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-12
DOI
https://doi.org/10.21203/rs.3.rs-11704703/v1
Primary Topic
Advanced Neural Network Applications
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Format failure hides inside accuracy: the GGUF quantization level that breaks is not predictable from its label

Shaoxuan Zhang
Research Square
Advanced Neural Network Applications
preprint

Format failure hides inside accuracy: the GGUF quantization level that breaks is not predictable from its label

Shaoxuan Zhang
preprint en

Abstract

Abstract Accuracy benchmarks score a non-answer and a wrong answer alike. We test whether that convention survives the 8-bit to 4-bit step of GGUF quantization. Across four instruction-tuned models from 0.5B to 3B, four tasks at 800 items each, and every published precision level, 70,394 paired responses separate the responses that answer from the responses that do not. Format failure is concentrated and not monotone in bit-width. On ARC-Challenge, Qwen2.5-0.5B produces no usable answer on 1.6% of responses at Q8_0, 5.4% at Q5_K_M, 19.4% at Q4_K_M, 9.1% at Q3_K_M and 10.4% at Q2_K: the level that fails is Q4_K_M, while the more compressed levels answer more often. Qwen2.5-3B fails at Q3_K_M instead, at 30.6% against 3.1% at Q4_K_M, and Gemma-3-1B and Llama-3.2-3B never exceed 0.1% at any level. Counting non-answers as errors raises the measured cost of the 4-bit step at 0.5B from 9.0% to 10.9%; on responses that do answer, the cost falls with scale, 9.0%, 6.8% and 3.1% against a 0.4% gain at Llama-3.2-3B. A label does not tell a practitioner whether the artifact will answer, and the level that fails cannot be inferred from its neighbours.

Research Square
Harbin Institute of Technology (CN)
Advanced Neural Network Applications
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon