CIPO rewards agents for actually using retrieved evidence, not just retrieving it.

arXiv 2608.06128 attacks confirmation bias in search agents, where the model decides from parametric memory and uses retrieval to confirm itself. CIPO assigns dense turn-level credit specifically to reasoning actions influenced by retrieved information, plus a global outcome reward, with no human process annotations or separate reward model. It reduces prior-driven reasoning across seven in-domain and out-of-domain benchmarks. "It cited a source" and "the source changed its answer" are different measurements, and almost nobody instruments the second one.