Improving LLM Embeddings via Inter-Layer Geometry
June 2025 – PresentICLR 2026 GRAM Workshop (Tiny Paper Track) · First Author
A GNN-based method to enhance LLM embeddings for downstream NLP tasks by exploiting the geometry between transformer layers.
- · Developed a GNN-based method to enhance LLM embeddings for downstream NLP tasks.
- · Evaluated on 12 datasets across 9 LLMs, comparing multiple GNN architectures and baselines.