Study Finds Injecting Factual But Distracting Evidence Can Drop Search Agent Accuracy by Average 5.9 Points

Researchers have identified a failure mode called "lazy grounding," demonstrating that search agents can be misled by completely factual but distracting information, not just misinformation. The study, authored by Yulin Zhang et al., was submitted to arXiv on August 31, 2026, revised to v3 on October 13, 2026, and accepted to the EMNLP 2026 Main Conference.

While retrieval-based approaches mitigate hallucination by grounding answers in web results, they introduce an attack surface. Previous concerns focused on poisoned corpora containing false or malicious documents. This work shows that even truthful documents can degrade performance if they support a neighboring rewritten question rather than the original query, leading agents to misapply valid evidence.

The team exposed lazy grounding by injecting nearby evidence from answer-changing rewrites of benchmark questions into search corpora. Each injected document contained factual evidence supporting a neighboring rewritten question but was retrieved for the original question. Across 12 model-benchmark pairs, the attack caused the accuracy of search agents' responses to drop by 5.9 points on average and by up to 17.3 points, inducing nearby-answer adoption in every setting. The effect was stronger when the distracting evidence appeared later in the retrieval list or was more answer-shaped.

These results indicate that robust search agents must defend against the misapplication of factual evidence, not just misinformation. Code is publicly available at github.com/frankyzha/lazy-grounding. These are first-party benchmark results; independent reproduction has not yet been reported.