Pre-Evidence Procedural Failures in Agentic RAG Systems

A recent paper on arXiv (2608.02011) explores a particular failure mode in retrieval-augmented generation (RAG) systems, where agents obtain candidate snippets but do not review them before providing answers. The research categorizes incorrect responses into two types: pre-evidence discipline failures and post-gold-read failures, utilizing recorded tool-call traces, retrieved evidence, read passages, and final outputs. Analyzing 12,000 paired trajectories from HotpotQA, 2WikiMultiHopQA, and MuSiQue reveals that these failure types are mostly distinct, with a both-trigger rate ranging from 11.2% to 13.1% for regex and spaCy entity extractors. The authors introduce Read-Gate, a simple runtime requirement mandating agents to read after searching and prior to finalizing answers. This enforced reading enhances LLM-Acc by 14.9-19.9 points for trajectories that would normally bypass reading and by 3.2-9.4 points for complete minimal-reasoning cells. Additional diagnostics are included, although the abstract is truncated. This study is significant for advancing dependable AI systems in information retrieval and question answering.