Research map / Architectures and efficiency / RNNs and normalization

STRABLE: Benchmarking Tabular Machine Learning with Strings

NeurIPS 2026 · Architectures and efficiency › RNNs and normalization · approach: Learned embeddings for retrieval · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces STRABLE, a 108-table real-world benchmark with raw strings and numbers, and runs a large empirical study of 445 modular and end-to-end pipelines for tabular learning with strings.

Problem

Tabular learning benchmarks either drop or flatten string columns, so it is unclear whether dedicated end-to-end learners are needed for tables with raw strings or whether encoding strings as numbers suffices.

Approach

A benchmark corpus of raw string-and-number tables with minimal preprocessing, used to compare modular pipelines (string encoder, post-processing, tabular learner) against end-to-end string-aware learners, plus a theoretical analysis of ranking stability with respect to the number of datasets.

Key results

• Lightweight Tf-Idf encoders paired with advanced learners such as TabPFN-2.5 and TabICLv2 sit on the Pareto frontier of performance and runtime, outperforming E2E architectures.
• Decoder-only LLM embeddings (LLaMA-3.1-8B, Qwen-3-8B) improve when default 30-PCA is replaced by standard scaling or raw-dimension slicing.
• Kendall-tau agreement with the oracle ranking reaches about 0.95 at N=108 datasets (estimated, extrapolated).

Surprising findings

• Large LLM encoders paired with TabPFN-2.5 enter the top-10 only on free-text-dominant tables, while lightweight encoders dominate elsewhere.
• Default PCA hides the performance of decoder-only LLM embeddings, which concentrate variance in a few dimensions.

Limitations

• Strings are short categorical content, so limited coverage of long-form sentence-heavy tables.
• Does not address time-series specific validation protocols.
• Oracle ranking agreement is an extrapolation from subsampled benchmarks rather than a direct measurement.

Experimental setup

Datasets: STRABLE

Benchmarks: STRABLE

Metrics: Kendall-tau average rank score

Baselines: CatBoost Mambular TabSTAR ConTextTab TabPFN-2.5 TabICLv2 RealMLP TabM

Builds on: TabPFN CatBoost TARTE CARTE TabSTAR ConTextTab Mambular

Closest papers by meaning

• Large Scale Transfer Learning for Tabular Data via Language Modeling (NeurIPS 2024 · similarity 0.88)
• TabFlex: Scaling Tabular Learning to Millions with Linear Attention (ICML 2025 · similarity 0.88)
• UniTabE: A Universal Pretraining Protocol for Tabular Foundation Model in Data Science (ICLR 2024 · similarity 0.88)
• ConTextTab: A Semantics-Aware Tabular In-Context Learner (NeurIPS 2025 · similarity 0.88)
• TabDPT: Scaling Tabular Foundation Models on Real Data (NeurIPS 2025 · similarity 0.87)
• TabArena: A Living Benchmark for Machine Learning on Tabular Data (NeurIPS 2025 · similarity 0.87)