Uncertainty-Aware Retrieval for Robust Scenario Composition: A GFlowNet-RAG Framework with Subset-Simulation Calibration

Abstract

Scenario generation often requires many valid and meaningfully different composition from retrieved outputs, not just one top-ranked context. Traditional RAG pipelines are poorly matched to this setting: deterministic Topk retrieval tends to collapse onto similar outputs. Additionally, repeated retrieval and reranking can make the system computationally heavy, and make it inefficient to produce large sets of high-quality and diverse scenarios. We propose GFN-RAG as a replacement retrieval sampler. A GFlowNet trained with Trajectory Balance learns a distribution over traffic-card scenarios, where each scenario is an ordered sequence of retrieved spatial-view cards. These cards are sampled with probability proportional to a reward combining coverage and geometric consistency. After training, GFN-RAG can generate diverse high-reward scenarios efficiently, without rerunning a heavy selection pipeline for every sample. As an additional reliability layer, Data-Driven Subset Simulation (SSDD) calibrates and evaluates the probability of retrieval failure or semantic flips under perturbations or attacks for both traditional RAG and GFN-RAG. We instantiate the framework on natural-language-to-traffic-scenario conversion and define the retrieval MDP, reward taxonomy, and evaluation protocol.