New TRACES benchmark finds LLMs fail 71% of scientific integrity probes

Large language models are increasingly proposed as autonomous agents in scientific workflows, yet their ability to distinguish reliable from fraudulent literature has gone unmeasured. The new TRACES benchmark, introduced by Valentin Rodionov and Shamil Assylbekov, directly targets this gap with a probe corpus of 42 retracted, fraudulent, and pseudoscientific papers. Each probe pairs a near-verbatim preamble from a target paper with a plausible study-design request, spanning five claim types: fabricated observation, pseudophysical mechanism, magical premise, legitimization bridge, and cargo-cult experiment.

Across 30 models and 10 repeated runs, aggregate rejection scores appear strong at first glance: IFR-a (outright rejection) is 0.93 ± 0.004, and IFR-i (recognition of unreliability while engaging) is 0.809 ± 0.009. But the deeper results are alarming. Models engaged with untenable premises in 95% of all non-empty responses, and every evaluated model failed more than 71% of agentic probes. Perhaps most telling, 22 of 30 models failed over 90% of the time. The rejections that do occur are concentrated on a small number of high-notoriety topics and vanish under matched-structure controls. The authors argue this is consistent with topic-keyed safety behavior rather than robust epistemic competence, and they issue an urgent call for guardrail infrastructure before LLMs are deployed in scientific settings.