Unified SDT Model Reveals Human Memory Beats RAG on Retrieval Interference (0.41 vs. 0.67)

Humans suffer 63% less retrieval interference from semantic associations than dense passage retrieval, according to a unified signal detection theory framework published today on arXiv. That gap — interference sensitivity ($\alpha/\sigma$) of 0.41 for humans versus 0.67 for DPR — quantifies exactly how much harder it is for RAG systems to sift through competing memories when the number of associations (fan) grows.

The Fan Effect Hits RAG Harder

Both systems show logarithmic accuracy decline with association count, but the slope is steeper for machines. The authors fit behavioral data from 112 participants and matched simulations for DPR and HippoRAG. Parameter recovery confirmed identifiability at $r \geq .93$, and model comparison decisively favored the logarithmic specification over a power-law alternative ($\Delta$BIC $> 15$). That means the interference scaling law is convincingly log-linear, not power-law, for both biological and artificial retrievers.

HippoRAG Bridges the Gap

HippoRAG — a retrieval architecture explicitly inspired by human hippocampal indexing — lands at $\alpha/\sigma = 0.44$, far closer to the human benchmark than standard DPR. That makes sense: its design encodes temporal context and uses a gating mechanism reminiscent of episodic retrieval. The framework reveals that encoding specificity and temporal context binding are the likely drivers of human advantage, though the authors are careful to note the causal role of these mechanisms remains to be tested directly.

Six Predictions That Connect Memory and AI

The paper closes with six falsifiable predictions that tie cognitive memory research directly to RAG evaluation. Among them: manipulations that strengthen temporal context binding should shrink the interference gap, and retrieval gating with adaptive thresholds should push machine performance below human levels on certain fan regimes. These are concrete experiments, not hand-wavy calls for future work.

The unified SDT framework turns a fuzzy comparison between human and machine memory into a testable, parametric gap. For anyone building retrieval systems that need to handle high-association knowledge — think multi-hop QA or long-context memories — this says: borrow the hippocampus's tricks or accept a 63% handicap.

Source: The Interference Gap: Comparing Retrieval Bounds in Human Memory and RAG Systems
Domain: arxiv.org