A post-training tensor decomposition of multi-head attention weights with shared factor matrices improves reasoning accuracy on several LLM benchmarks while compressing attention parameters.
Language mod- els are few-shot learners,
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
background 1
citation-polarity summary
fields
cs.CL 1years
2025 1verdicts
CONDITIONAL 1roles
background 1polarities
unclear 1representative citing papers
citing papers explorer
-
TensorLLM: Tensorising Multi-Head Attention for Enhanced Reasoning and Compression in LLMs
A post-training tensor decomposition of multi-head attention weights with shared factor matrices improves reasoning accuracy on several LLM benchmarks while compressing attention parameters.