Core Concepts
Backward dependencies enhance sentence embeddings in large language models.
Abstract
Sentence embeddings are crucial for semantic similarity measurement.
BeLLM introduces backward dependency modeling to improve LLMs.
Experimental results show BeLLM outperforms previous SOTA models.
Ablation study highlights the importance of balancing uni- and bi-directional layers.
Case study demonstrates BeLLM's superior performance in semantic retrieval tasks.
Discussion on enhanced dependency shows the effectiveness of incorporating backward dependencies in LLMs.
Stats
Existing LLMs mainly adopt autoregressive architecture without explicit backward dependency modeling.
BeLLM achieves state-of-the-art performance in various semantic textual similarity tasks.
BeLLM significantly outperforms previous SOTA models, such as SimCSE and RoBERTa, across different benchmarks.
Quotes
"Most advanced NLP models adopted autoregressive architectures with forward dependency modeling only."
"BeLLM achieves a notable 2.5% improvement compared to the previous SOTA model."
"BeLLM performs the best in all S-STS datasets."