A block-scaled logarithmic format with exact integer-index products matches FP8-class perplexity on 7B-9B language models with seven bits and shows parity with FP8 in 3B-token pretraining.
(pt) ↑ BF16 FP8 CurveFP8CurveFP7 Runtime format 0 -0.25 -0.24 -0.45 0 -0.08 -0.07 -0.02 0 -0.03 -0.13 +0.01 (c) Five-shot Δacc
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
CurveFP: Co-Designing Numerical Representation and Product Arithmetic for Language Models
A block-scaled logarithmic format with exact integer-index products matches FP8-class perplexity on 7B-9B language models with seven bits and shows parity with FP8 in 3B-token pretraining.