A collaborative technique in which a fine-tuned bidirectional encoder (such as ELECTRA) supplies its predicted label, class probabilities, and retrieved similar examples as extra context inside the prompt given to a large language model for three-way sentiment classification. Students learn how encoder-derived signals can raise a general LLM's classification accuracy with no further training, why the same augmentation can instead degrade a model that has itself been fine-tuned, and how to reason about the resulting cost-versus-performance trade-offs when building affordable text classifiers.
Bidirectional transformers excel at sentiment analysis, and Large Language Models (LLM) are effective zero-shot learners. Might they perform better as a team? This paper explores collaborative approa…