A lightweight iterative cross-attention latent head on frozen LLMs beats major text-regression baselines with only ~3.4–3.7M trainable parameters.
Yelin, and Roel Van Beeumen
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
RELISH: LLM REgression with a Latent Iterative State Head
A lightweight iterative cross-attention latent head on frozen LLMs beats major text-regression baselines with only ~3.4–3.7M trainable parameters.