· via dev.to (home feed)
Retrieval agents stay overconfident when cited evidence is corrupted, DeepStress finds
A dev.to post covering the DeepStress benchmark reports that search agents keep giving confident answers even when cited sources carry explicit low-credibility signals, with Search-R1 flat at 78-80% accuracy.

A benchmark that breaks the evidence
Most evaluations of retrieval-augmented agents rest on a quiet assumption: the documents the retriever hands back are either correct or at least plausibly useful. According to a post on dev.to, the DeepStress benchmark, introduced in a paper titled "DeepStress: Stress-Testing Deep Search Agents", drops that assumption. It deliberately corrupts source documents at scale and then measures what agents do as their evidence base degrades. The post argues that this failure mode is invisible to standard benchmarks, which score right-or-wrong answers against clean retrieval and never probe how an agent behaves when the material it cites is riddled with errors.
Confident answers from zero-trust sources
The central finding is blunt. Even when the retrieval layer returns documents explicitly labeled as completely untrustworthy, every tested system still produced a definitive answer rather than abstaining in many runs. The paper's authors, as quoted in the post, describe the low-credibility signals in those documents as unmistakable to the point of caricature. The warnings were about as obvious as credibility flags get, and the agents answered anyway.
Flat accuracy, flat abstention
The accuracy numbers are just as telling. Search-R1 stayed between 78 and 80 percent correct across the entire trustworthiness spectrum, and its rate of abstaining showed no clear change at any point. In other words, the agent behaved much the same whether its sources looked solid or were marked as junk. That points to a calibration problem rather than plain incompetence: under noisy retrieval the model does not become more wrong, it simply never signals doubt.
Where the results stop
The post is careful about scope. DeepStress evaluates synthetic degradations on two datasets, HotpotQA and BrowseComp-Plus, and looks only at final-answer confidence, not at intermediate uncertainty signals. It does not test naturalistic noise, larger model families, alternative prompting strategies, or memory-augmented routing. As the author concedes, there is no basis yet for claiming that the same brittleness will appear in production assistants built on different techniques.
Why it matters
Citations have become the de facto trust mechanism for agent answers: if the assistant shows its sources, users treat the answer as verified. This research suggests the citation is doing less work than assumed, because an agent will cite documents it has been told are worthless and still speak with full confidence. Anyone consuming agent-sourced answers, or deploying agents in settings where knowing when to stay silent matters, is relying on behavior that ordinary leaderboards never measure.
The post closes with two concrete recommendations. Safety pipelines should enforce a credibility filter that forces agents to abstain whenever retrieved evidence falls below a minimal trust threshold. And benchmarks should adopt DeepStress-style stress testing as a standard sanity check. The gap between answering correctly and knowing when not to answer is precisely the gap production systems live or die on, and right now it is mostly untested.
- #ai-agents
- #llm
- #benchmarks
- #retrieval-augmented-generation
- #evaluation