Research map / Vision-language and multimodal / Human and embodied modeling

TRL-Bench: Standardizing Cross-Paradigm Representation-Level Evaluation of Tabular Encoders

NeurIPS 2026 · Vision-language and multimodal › Human and embodied modeling · approach: Vision-language pretraining · 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 TRL-Bench, a multi-granular benchmark that evaluates 20 frozen tabular encoders via exported row, column, and table embeddings under shared lightweight probes across column/table tasks, row tasks, and a data-lake enrichment pipeline.

Problem

Tabular encoders are usually evaluated inside task-specific end-to-end pipelines, so representation quality across heterogeneous training paradigms is hard to compare directly.

Approach

Standardize evaluation at the level of exported frozen embeddings (row, column, table) with fixed lightweight downstream heads, and add a compositional data-lake enrichment benchmark (TRL-DLTE) to test whether atomic capabilities compose into pipelines.

Key results

• Best hybrid DLTE pipeline TUTA/GTE/GTE reaches 0.229 UJ-H on dev-selected evaluation, 0.090 above monolithic BERT/BERT/BERT (0.139).
• On RBench Robust Linkage, GTE attains the best normalized rank (0.048) while TabICL leads row prediction with AUROC 0.816 and Macro-F1 0.671.
• Stage-3 row-model ranking on DLTE agrees with RBench Robust Linkage NR at Spearman |rho|=0.80.

Surprising findings

• Generic text encoders such as BERT and GTE often lead on column and table tasks with strong surface-text signal, and per-stage marginal leaders do not assemble into the top pipelines.
• Atomic retrieval strength decouples from compositional utility: GTE leads Table Retrieval but ranks only third on the Stage-1 UJ-H marginal.

Limitations

• Scores measure what lightweight readouts extract from frozen embeddings, not fully adapted end-to-end performance.
• Wrapper and input serialization choices per model introduce variation not fully controlled by the protocol.
• Scope restricted to roughly 1M-1B parameter encoders; generative table LLMs are excluded.

Experimental setup

Datasets: SATO SOTAB WikiCT Spider Valentine OpenML DeepMatcher WDC Products WTQ TabFact TabArena OpenML-CC18

Benchmarks: TRL-CTbench TRL-Rbench TRL-DLTE

Metrics: F1 NMI MAP R@GT nRMSE MRR Accuracy AUROC Macro-F1 SGM

Baselines: TF-IDF value overlap Jaccard Random Dummy

Builds on: Observatory TabularS3L TabArena LakeBench DeepMatcher

Closest papers by meaning

• Large Scale Transfer Learning for Tabular Data via Language Modeling (NeurIPS 2024 · similarity 0.89)
• TabM: Advancing tabular deep learning with parameter-efficient ensembling (ICLR 2025 · similarity 0.89)
• MulTaBench: Benchmarking Multimodal Tabular Learning with Text and Image (NeurIPS 2026 · similarity 0.89)
• Hybrid Autoencoders for Tabular Data: Leveraging Model-Based Augmentation in Low-Label Settings (NeurIPS 2025 · similarity 0.88)
• TabSTAR: A Tabular Foundation Model for Tabular Data with Text Fields (NeurIPS 2025 · similarity 0.88)
• TabGLM: Tabular Graph Language Model for Learning Transferable Representations Through Multi-Modal Consistency Minimization (AAAI 2025 · similarity 0.88)