ADEPT: An Entropy-Driven Dual-Strategy Agent for Interactive Video Retrieval

This RSS collection summarizes a set of 2026 research preprints focused on retrieval, embeddings, and retrieval-augmented generation (RAG). Key themes include: (1) RAG robustness and failure modes — empirical sensitivity/stability analyses showing non-monotonic downstream behavior, preprocessing-induced answer loss, and retrieval corruption effects; (2) high-dimensional embedding geometry — distance/cosine concentration, contrast collapse, hubness, and retrieval instability that can degrade grounding; (3) emergent population-level effects in dense embedding spaces — spatial crowding that can “