Negation is Not Semantic: Diagnosing Dense Retrieval Failure Modes for Trade-offs in Contradiction-Aware Biomedical QA

A research team has published a critical diagnosis of why current AI systems fail at a fundamental task in medicine: correctly handling negation and contradiction. Their paper, "Negation is Not Semantic," reveals a catastrophic flaw they term 'Semantic Collapse' in dense retrieval models used by systems like GPT-4 and Claude. When these models convert text into vectors, signals for negation (like "not effective") become indistinguishable from positive statements ("effective"), leading to dangerously plausible but false answers. In tests for the TREC 2025 BioGen track, which requires AI to surface contradictory evidence, complex adversarial retrieval strategies scored a near-failure MRR of just 0.023.

To solve this, the researchers built a Decoupled Lexical Architecture that abandons an all-in-one vector approach. Instead, it uses a unified BM25 lexical search backbone to separately handle finding supporting evidence and pinpointing contradictions. This method achieved a balanced performance with a semantic support recall of 0.810 and a contradiction detection score of 0.750. For the final answer generation, they added Narrative Aware Reranking and One-Shot In-Context Learning, boosting citation coverage from 50% to 100%.

The results were validated in official competition. Their system ranked 2nd for contradiction detection (Task A) and 3rd out of 50 runs for citation coverage in Task B, achieving a remarkable 98.77% coverage with a zero citation contradiction rate. This proves the system can reliably scale to search the 30-million-document PubMed corpus. The work provides a blueprint for moving large language models from being unpredictable 'stochastic generators' to trustworthy 'evidence synthesizers,' where every claim is explicitly backed by and checked against source material.