Tag
The paper introduces Counterfactual Gate–Support Decomposition (CGSD), a four-arm budget-matched experimental design that separates within-support selection from between-support redistribution effects in training-data quality gates, showing on Korean–English/Korean–Japanese retrieval that only within-support selection reliably exceeds the calibrated null.
Introduces XBCP (Cross-lingual BrowseComp-Plus), a benchmark for evaluating deep research agents and retrievers in cross-lingual and multilingual settings. Results show significant performance degradation when evidence is in a different language from the query, highlighting both retrieval failures and agent-side difficulty in integrating language-mismatched evidence.