New Ablation Method Diagnoses Diffusion-as-Inference Gains

Warm-started diffusion samplers are increasingly used to accelerate iterative inference, but it's often unclear which specific components contribute most to their performance gains. This paper focuses on retrieval-warmed energy-based reasoning (RW-EBR), an IRED energy-based diffusion model enhanced with a Modern Hopfield trajectory memory. To dissect the performance, the authors propose a novel five-arm ablation methodology. This diagnostic framework is designed to isolate and quantify three distinct, often confounded, effects: shifts in class-prior bias, the benefits of stochastic warm-starting, and the impact of graph-aligned value reuse. This approach is adapted from LLM-RAG evaluation techniques. Applying this methodology to tasks like connectivity-2 (graph reachability) revealed that per-graph alignment, rather than bias shift or stochasticity, was the dominant factor, yielding a significant +35 percentage point improvement in balanced accuracy. However, for Sudoku, the diagnostic pointed to key quality as the primary bottleneck. This decomposition provides a clear way to identify the first blocking component for each task, guiding future research and development in structured and spatio-temporal reasoning.