A new type of attack, Micro-Collaborative Poisoning, is introduced that targets Retrieval-Augmented Generation (RAG) systems. This attack involves dividing a false target claim across multiple documents, making it difficult to detect. The attack is evaluated across various RAG configurations, and the results show that it can be effective even with diverse and strong retrievers. This highlights the need for improved defenses against such attacks.

Researchers explore the impact of chunking and representation on retrieval-augmented generation in chemistry-aware question answering. They evaluate 41 embedding models on a new benchmark and find that retrieval-tuned E5, BGE, and Nomic models perform well. They also identify optimal chunking strategies and sizes for practical application.

Researchers introduce RAFT, a stateful retrieval-augmented generation framework for troubleshooting agents, addressing the limitations of existing RAG systems. RAFT abstracts historical cases as directed chains and retrieves cases with matching intermediate states, improving case hit rates compared to vanilla and GraphRAG baselines.

Researchers propose a novel framework for improving interpretability of spreadsheets in LLM-driven RAG systems by splitting them into interpretable chunks using cell role annotation. This approach aims to address the 'spreadsheet-to-LLM bottleneck' by developing dimensionality-reduction techniques to flatten 2D unstructured spreadsheets into 1D text.

Researchers introduce DRAG, a query-adaptive framework for selecting retriever-generator configurations in Retrieval-Augmented Generation (RAG) systems, leading to improved effectiveness-efficiency trade-off. The framework includes two approaches: DRAG_QPP, a training-free routing method using query performance prediction, and DRAG_SFT, a supervised routing method that fine-tunes a large language model to predict retriever-generator configurations.

A new graph-based retrieval-augmented generation system, EffiRAG, reduces the cost of graph construction and query processing in RAG systems while maintaining or improving answer quality. This affects AI agent developers by providing an alternative approach to traditional RAG systems, allowing them to consider cost-effectiveness in addition to answer quality when designing their agents.

RAG-CT is a novel defense mechanism that identifies malicious queries in Retrieval-Augmented Generation (RAG) systems, preventing personally identifiable information (PII) leakage. It analyzes entropy and margin distributions to detect and mitigate PII extraction attacks on LLMs, outperforming existing defenses without modifying the underlying LLM or retriever.

A new paper proposes a solution to the 'Semantic Shadowing' failure mode in Retrieval-Augmented Generation (RAG) architectures, which are used in long-horizon autonomous agents. The proposed solution, GC-Mem, is a strict inference-time consistency protocol that uses a temporal dominance operator and contradiction detection to excise shadowed context. The authors demonstrate its effectiveness through rigorous benchmarking and establish deployment thresholds to ensure state convergence.

Researchers introduce ORDER, a query-conditioned Retrieval-Augmented Generation (RAG) framework that adapts indexing and retrieval to incoming queries, improving performance in complex expert-domain environments.

A new method for retrieval-augmented generation (RAG) pipelines is proposed to improve source-risk triage by treating source relationships to the query as query-dependent. The method routes canonical query families for enhanced review and assigns retrieved pages to pass, contextualize, exclude, or review. This has implications for the accuracy and reliability of AI agents using RAG pipelines, particularly in terms of page-level frequency vs family-level exposure and calibration.