FRESCO benchmark reveals re-rankers consistently favor older Wikipedia documents even when factually obsolete, and instruction optimization yields up to 27% gains on evolving-knowledge tasks while preserving non-evolving performance.
Each new prompt must be a cohesive instruction set that aims to solve all provided examples by intelligently combining the best of A and B
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.IR 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
FRESCO: Benchmarking and Optimizing Re-rankers for Evolving Semantic Conflict in Retrieval-Augmented Generation
FRESCO benchmark reveals re-rankers consistently favor older Wikipedia documents even when factually obsolete, and instruction optimization yields up to 27% gains on evolving-knowledge tasks while preserving non-evolving performance.