LectureBank-Full Decomposition: Diffusion+Quotas Drive ~18 R@10 Points; Contrast Gating Adds At Most ~1 Point (Statistically Tied)
On LectureBank-Full, the strict-parity audit decomposes the retrieval gain by component: deterministic bidirectional diffusion plus role-aware quotas lifts R@10 by roughly 18 points relative to static parent expansion, while the contrast-gating (adaptive-depth) component adds at most a nominal +1.05 pp in paired-bootstrap delta with 95% CI and . The contrast gate is therefore statistically tied with the fixed-depth hierarchical baseline, and the main retrieval gain is attributable to diffusion plus role-aware quotas rather than to adaptive depth.
0
1
Tags
Science
Auditable Strict-Parity Evaluation of Prerequisite-Graph Retrieval for RAG under Leakage Controls
Related
LectureBank-Full Decomposition: Diffusion+Quotas Drive ~18 R@10 Points; Contrast Gating Adds At Most ~1 Point (Statistically Tied)
Bounded Benchmark Validity: Two Question Families and 21/18 Unique Held-Out Targets Cap Statistical Power
Restrained Claim Scope: No Automatic-Judge Validation, No Semantic-Evasion Claim, No Robust End-to-End QA Gains
Language-Matched Seeding as a Prerequisite for Graph-Expansion Gains
HotpotQA External-Validity Probe: Adaptive Depth Does Not Transfer to a Denser Non-Prerequisite Graph (FullWiki-1k: Flat 93.4 / Hier 92.9 / Adaptive 94.0 R@10)
Seed Quality and Graph Noise (Not Depth Control) Are the Main Unresolved Bottlenecks
LectureBank-Full Decomposition: Diffusion+Quotas Drive ~18 R@10 Points; Contrast Gating Adds At Most ~1 Point (Statistically Tied)
LectureBank-Full R@10 Gain from Diffusion and Role-Aware Quotas
LectureBank-Full Configuration Used in Hierarchical Prerequisite RAG (208 Concepts, 899 Edges, 1,421 QA)
LectureBank-Full Target-Disjoint R@10 Result (n=164): Diffusion Gain Survives, Adaptive Tied
LectureBank-Full Generation Diagnostic: Token-F1 1.9 → 18.3, EM Stays 0.0
LectureBank-Full Error Taxonomy: Residual Misses Are Near-Misses Along the Local Prerequisite Graph
LectureBank-Full Paired Delta: Adaptive vs Hierarchical Baseline = +0.7 [-2.1, +3.6]
Token-Cap Comparison on LectureBank-Full: Adaptive Loses More as Cap Tightens
LectureBank-Full Tight-Budget Advantage of Adaptive Depth Gating (Mean ΔR@k = +2.13 over k∈{1,2,3,4})
LectureBank-Full ΔR@k Peaks at k=4 (+6.4 Points, CI [1.0, 11.7])
LectureBank-Full Diffusion Gain over Static Parent Expansion (~18 R@10 Points)
Bounded Held-Out Targets After Strictest Leakage Control (21 LectureBank-Full, 18 MOOC-CS)
LectureBank-Full Decomposition: Diffusion+Quotas Drive ~18 R@10 Points; Contrast Gating Adds At Most ~1 Point (Statistically Tied)