CoDeR: New retrieval method slashes constraint violations by 23 points

Traditional information retrieval systems rely on semantic similarity as a proxy for relevance, but this fails for constraint-sensitive queries—for example, when a document is topically close but supports the opposite constraint direction (e.g., affirming a negated relation). A new paper from Yin, Tang, and Du introduces CoDeR (Local Constraint-Compatible Retrieval), a dense retrieval method that explicitly separates topical relevance from constraint compatibility. CoDeR uses a standard topical encoder for candidate coverage and adds a compatibility scorer (a bi-encoder) trained with lexical-polarity supervision on contrastive satisfying and violating evidence. The compatibility signal can rescore topical candidates or retrieve an auxiliary compatibility-oriented set, producing a ranked list without external LLM calls at inference.

CoDeR was evaluated on three controlled diagnostic sets targeting antonymy, negation, and exclusion. Results show a reduction in V@2 (violation at rank 2) by 20.59, 23.53, and 5.77 points respectively compared to the strongest non-CoDeR baselines. The method also improves FVR (first violating rank) by pushing violating documents deeper in the ranking. This approach is particularly valuable for applications like legal document retrieval, medical information search, or any domain where negated or excluded conditions matter. By avoiding expensive LLM calls, CoDeR offers a practical, efficient solution for constraint-aware retrieval.