TRL-Bench: Standardizing Cross-Paradigm Representation-Level Evaluation of Tabular Encoders
Abstract
TRL-Bench establishes a standardized benchmark for evaluating tabular representation learning models across multiple granularities, revealing that encoder performance varies by task type and requires capability-specific assessment rather than single leaderboard rankings.
Tabular encoders are usually evaluated inside task-specific end-to-end pipelines, so models from different training paradigms are difficult to compare directly even when they operate on similar tabular signals. We introduce TRL-Bench, a multi-granular tabular representation learning (TRL) benchmark that standardizes cross-paradigm representation-level evaluation: each encoder exports row-, column-, or table embeddings through its supported wrapper, and shared lightweight heads probe them across three suites: TRL-CTbench (column/table), TRL-Rbench (row), and TRL-DLTE (compositional Data-Lake Table Enrichment spanning all three granularities). To support this standardized setting, we release curated benchmark assets and task reformulations, including 50 OpenML tables with 123 verified targets, 16 row-pair linkage rewrites, and a 47,772-table DLTE lake derived from 1,379 parent tables. Across 20 models and 16 tasks, TRL-Bench shows that once downstream conditions are standardized, encoder quality is capability-specific rather than captured by a single leaderboard. In TRL-CTbench, generic text encoders often lead on tasks with strong surface-text signal, while tabular specialists win where their pretraining objective aligns with the task. In TRL-Rbench, within-table prediction and cross-table linkage favor different training regimes, with atomic linkage performance correlating strongly with the row-matching stage of DLTE pipelines. In TRL-DLTE, the strongest pipelines combine capability-matched specialists rather than reuse a single encoder, and top end-to-end quality depends on non-additive compositional fit rather than per-stage marginal rank alone. TRL-Bench provides a common protocol for measuring reusable signal in exported tabular representations under shared downstream conditions. Code and data: https://github.com/LOGO-CUHKSZ/TRL-Bench
Community
📊 Releasing TRL-Bench — a unified framework + library for tabular representation learning, one stop for tabular representation learning.
🧩 20 encoders · 16 tasks · 87 datasets across 3 suites
🔍 Built to make heterogeneous tabular models directly comparable, and reusable as embedding models
Tabular encoders come in every shape: different input formats, training objectives, and output heads. So even two models built for the same job are hard to compare head-to-head.
We built TRL-Bench to make them comparable.
It unifies everything at the level of the representation: each model is wrapped behind one shared interface that exports row-, column-, and table-embeddings, and shared lightweight heads probe those embeddings under common task definitions, so 20 encoders from every paradigm finally sit on the same axes.
It's also a library: 20 different types of tabular models are adapted into embedding models that export row, column, and table embeddings for the community to reuse.
It spans three suites:
🧩 TRL-CTbench — 13 column/table tasks: schema, joinability, unionability, grounding
🔗 TRL-Rbench — multi-target row prediction (50 subtasks, 123 targets) + record linkage (16 datasets)
🌊 TRL-DLTE — a 47,772-table data-lake enrichment pipeline spanning all three granularities
The main takeaway is clear: there is no single best tabular encoder, strengths are split across different table jobs. The choice of tabular models should be task-aware.
We also find that:
📌 Off-the-shelf text encoders are surprisingly strong when the signal is in the surface text (column names and cell values); cross-table alignment and matching instead reward structure-aware specialists
📌 Predicting a value inside a table and matching the same record across tables call for different encoders: one rewards adapting to a single table, the other rewards embeddings that stay comparable across tables
📌 Stacking the best per-stage encoders does not give the best compositional pipeline, and neither does reusing one encoder end-to-end; the winning recipe matches a different specialist to each step (find related tables → align columns → match rows)
TRL-Bench is meant to serve both as a diagnostic benchmark and as a practical library for building on tabular representations.
📄 Paper: https://arxiv.org/abs/2606.09323
🌐 Website: https://logo-cuhksz.github.io/trl-bench.github.io/
🤗 Datasets: https://huggingface.co/datasets/logo-lab/trl-ctbench · trl-rbench · trl-dlte
💻 Code: https://github.com/LOGO-CUHKSZ/TRL-Bench
Good work!
Get this paper in your agent:
hf papers read 2606.09323 Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash Models citing this paper 1
Datasets citing this paper 3
Spaces citing this paper 0
No Space linking this paper
Collections including this paper 0
No Collection including this paper