Deterministic Tie-Breaking Rule in Role-Aware Ranking
When two diffusion candidates compete for the same role slot, the role-aware ranker breaks ties deterministically in a fixed order: (1) lower hierarchy level wins, (2) then higher dense similarity, (3) then stable document order. Because every step is deterministic, the returned context is reproducible across runs and removes the stochastic ancestor sampling used by several graph retrievers, which is what makes the policy auditable under strict parity.
0
1
Tags
Science
Auditable Strict-Parity Evaluation of Prerequisite-Graph Retrieval for RAG under Leakage Controls
Related
Bidirectional Diffusion Hyperparameter Settings (LectureBank-Full Main Runs)
Child/Bridge/Global Role Partition of Diffusion Candidates
Retrieval-Stage Role Quota (1, 4, 1) for Child/Bridge/Global
Deterministic Tie-Breaking Rule in Role-Aware Ranking
LectureBank-Full Diffusion Gain over Static Parent Expansion (~18 R@10 Points)
Retrieval-Stage Role Quota (1, 4, 1) for Child/Bridge/Global
Child/Bridge/Global Role Partition of Diffusion Candidates
Deterministic Tie-Breaking Rule in Role-Aware Ranking