Task-Adaptive Embedding Refinement via Test-time LLM Guidance

Researchers found that using an LLM at test time to refine queries can boost embedding-based search and classification.

The method adapts query embeddings using feedback from a generative LLM on a small set of documents, rather than the full corpus.

Results were consistently positive across benchmarks, with improvements reaching up to 25% on harder tasks.

The approach could offer a cheaper alternative to end-to-end LLM pipelines for task-specific retrieval and classification.