Retrieved but not ranked: surface-form bias in structural retrieval, from mathematics to agent trajectories

In 30 Sekunden

Was passiert ist

Researchers found that embedding models fail to retrieve items with matching structure but different wording, instead matching on surface-level text similarity across mathematics and agent tasks.

Warum es zählt

Engineers building retrieval systems should care when deploying embeddings for semantic search where paraphrasing or structural variation matters more than lexical overlap.

Achtung

LLM rerankers recover some performance but show domain-dependent gains; mathematics recovery partly reflects memorization rather than true structural understanding of problems.