Overview
- RouteRAG applies end-to-end reinforcement learning to jointly control graph‑text retrieval and generation, reporting significant gains over RAG baselines across five QA benchmarks with a two‑stage training scheme for efficiency.
- SEAL-RAG introduces a fixed-budget “replace, don’t expand” controller that counters context dilution, improving HotpotQA accuracy by 3–13 points and evidence precision by 12–18 points, and beating Adaptive‑k on 2WikiMultiHopQA by 8 points with p<0.001; code is released.
- CoopRAG coordinates a retriever and LLM via sub‑question unrolling and layer‑contrasting reranking, and the authors report consistent improvements over state‑of‑the‑art methods on three multi‑hop datasets and a simple QA set; code is available.
- FlippedRAG demonstrates transfer-based black-box attacks that reverse-engineer retrievers and poison targets, boosting attack success by 16.7%, shifting response polarity by about 50%, and defeating tested mitigations.
- Developer guides show RAG pipelines working with late‑interaction retrieval (ColBERT), web search, and SQL via txtai 9.3+, underscoring that effective retrieval can extend beyond standard vector search.