CascadER beats GraphSAGE on enterprise rules—and shows exactly where lookups stop working

A new evaluation makes a useful distinction that text-to-SQL and data-agent benchmarks often blur: some enterprise predictions are statistical, while others are recoverable business rules. CascadER uses an LLM only offline to propose composite keys, validates every proposal with SQL, and emits deterministic lookup cascades for runtime prediction. On SALT-KG, that approach beats the evaluated heterogeneous GraphSAGE baseline on all eight fields. Source

The largest gain is operationally meaningful

The study uses SALT-KG’s temporal split at July 1, 2020: 1.9 million training rows and 403,000 test rows across eight autocomplete tasks. For CustomerPaymentTerms, GraphSAGE reaches 37.5% test accuracy; CascadER reaches 82.9%. That changes the implied manual-correction rate from 62.5% to 17.1%. On SalesGroup, another high-cardinality field, the comparison is 15.8% versus 70.0%. Near-deterministic organizational mappings such as Plant, ShippingPoint and SalesOffice put both systems above 98%. Source

The architecture matters as much as the score. Claude Opus 4.6 at temperature zero drives the paper’s offline rule search, but SQL on temporally valid data scores candidate rules. Runtime inference is a JSON key lookup, not an LLM call. A limited sensitivity check repeated discovery for CustomerPaymentTerms and Plant with GPT-5.3-Codex and recovered identical cascade structures and test accuracy. The authors correctly say this is not a full model-sensitivity study. Source

The negative results define the boundary

CascadER does not establish that lookup tables generally beat graph learning. On RelBench v2’s 53-class rel-arxiv task, its recency-weighted author-category lookup scores 50.3% on test, just below GraphSAGE’s 50.7%; 29% of test authors have no pre-cutoff publications. On three-class rel-stack badge prediction, per-user mode reaches 79.3% against GraphSAGE’s 82.8%. Bronze badges make up 86% of training data, so mode-based lookups collapse toward the majority class and miss per-instance temporal thresholds. Source

That is the benchmark’s real contribution: it identifies configured determinism as a testable regime, rather than presenting one more universal model claim. The paper also finds that 17.8% to 43.8% of customers change dominant values between training and test for the non-trivial fields. Its four-field drift fit has an R² of 0.98, but the authors explicitly flag the sample as too small for a strong statistical conclusion. Source

For benchmark readers, the practical rule is narrow but valuable: deterministic cascades deserve a baseline whenever a task is governed by stable composite-key rules. Where labels depend on evolving neighborhoods or instance-level time, the same baseline becomes a diagnostic for what the benchmark is actually measuring—not a replacement for learned models.