Study Shows LLM Attention Values Better Capture Sentence Meaning Than Hidden States
Researchers propose that attention value vectors in large language models capture sentence semantics more effectively than the hidden states typically used for embeddings. The study introduces Value Aggregation (VA) and Aligned Weighted Value Aggregation (AlignedWVA) methods that outperform existing training-free LLM-based embedding approaches. This finding could improve how sentence representations are extracted from LLMs for downstream NLP applications.
A new arXiv paper challenges the conventional approach of using final-layer hidden states for LLM-based sentence embeddings, arguing that attention value vectors better encode sentence-level semantics. The researchers propose Value Aggregation (VA), a training-free method that pools token values across multiple layers, and demonstrate it outperforms existing LLM embedding methods including the ensemble-based MetaEOL. They further refine this approach with Aligned Weighted Value Aggregation (AlignedWVA), which interprets layer attention outputs as weighted value vectors aligned through the LLM's output projection matrix, achieving state-of-the-art performance among training-free methods. The paper also suggests that fine-tuning Value Aggregation could yield even stronger embedding models. This work provides a new perspective on how to extract meaningful sentence representations from LLMs, which are foundational to many NLP applications.
What's missing
The paper does not discuss computational efficiency or inference time comparisons between VA/AlignedWVA and baseline methods, which would be relevant for practical deployment. Additionally, the generalization of these findings to non-English languages or domain-specific applications is not addressed.
What different sources said
- arXiv cs.CLCenter
LLM-based Embeddings: Attention Values Encode Sentence Semantics Better Than Hidden States
Related
Topology-Aware Thermodynamics Improves DNA Probe Specificity Design
Researchers developed a new framework for designing DNA probes that accounts for the spatial organization of matched sequences, not just overall thermodynamic stability. Traditional methods rely on scalar measures like melting temperature and free energy, which miss how mismatches are distributed along the probe. The approach could improve diagnostic accuracy in applications like HPV detection and gene expression profiling.
Study Identifies Optimal Thermal Dose for Combining Focused Ultrasound with Immunotherapy in Tumors
Researchers used multimodal PET imaging to identify an optimal thermal dose range for focused ultrasound ablation that destroys tumor tissue while preserving conditions for immunotherapy delivery. The study found that excessive heating collapses blood vessels needed for antibody access, while insufficient heating fails to adequately reduce tumor burden. The findings could guide clinical design of combination treatments pairing thermal ablation with immunotherapies.
Plant MSH1 Protein Functions as Mismatch-Directed Nuclease for Organelle Genome Maintenance
Researchers have identified the precise mechanism by which the AtMSH1 protein in Arabidopsis plants recognizes and cleaves DNA mismatches and lesions, preventing mutations in organellar genomes. The protein combines a DNA mismatch recognition module with a nuclease domain that makes staggered cuts at specific positions relative to DNA damage. This discovery explains how plants maintain unusually low mutation rates in their mitochondrial and chloroplast DNA compared to other eukaryotes.