Dense retrieval architectures are fundamentally flawed at detecting negation and contradictions due to 'Semantic Collapse' in vector space.
March 19, 2026
Original Paper
Negation is Not Semantic: Diagnosing Dense Retrieval Failure Modes for Trade-offs in Contradiction-Aware Biomedical QA
arXiv · 2603.17580
The Takeaway
The paper demonstrates that even advanced adversarial dense models fail at contradiction detection (MRR 0.023). It advocates for a return to 'Decoupled Lexical Architectures' for clinical and high-stakes RAG, where accurately surfacing contradictory evidence is vital for safety.
From the abstract
Large Language Models (LLMs) have demonstrated strong capabilities in biomedical question answering, yet their tendency to generate plausible but unverified claims poses serious risks in clinical settings. To mitigate these risks, the TREC 2025 BioGen track mandates grounded answers that explicitly surface contradictory evidence (Task A) and the generation of narrative driven, fully attributed responses (Task B). Addressing the absence of target ground truth, we present a proxy-based development