metadata
license: mit
task_categories:
- question-answering
tags:
- chemistry
- biology
size_categories:
- 1K<n<10K
Evaluating Relational Reasoning in LLMs with REL
Authors: Lukas Fesser*, Yasha Ektefaie*, Ada Fang*, Sham M. Kakakde, Marinka Zitnik
* indicates equal contribution
Read our preprint here.
This repository contains the REL benchmarks for chemistry, biology, and algebra. For evaluation scripts please refer to our GitHub repo.
Algebra (1,000 questions)
All algebra tasks are Raven's Progressive Matrices: complete a 3×3 or 9×9 grid of number tuples by identifying the missing panel from 8 candidates. The matrix sizes differ in context length and difficulty.
| Task | RC | RC min | RC max | Size | Questions | Assesses |
|---|---|---|---|---|---|---|
| REL-A1 | 2 | 2 | 2 | 3×3 and 9×9 | 125 each | Constant rows — each row is identical; tests whether the model recognizes invariance |
| REL-A2 | 2 | 2 | 2 | 3×3 and 9×9 | 125 each | Arithmetic progression — values shift by a fixed increment across rows; tests detection of linear trends |
| REL-A3 | $n$ where $n\times n$ is the matrix size | 3 | 9 | 3×3 and 9×9 | 125 each | Permutation — each row is a cyclic permutation of the same values; tests recognition of reordering patterns |
| REL-A4 | $n$ where $n\times n$ is the matrix size | 3 | 9 | 3×3 and 9×9 | 125 each | Row-Sum — the final value in each row is the sum of all other entries in the same row multiplied by either $\pm1$, depending on the column. |
Biology (7,952 questions)
All biology tasks are REL-B1, split across 12 files by phylogenetic tree/alignment instance. Each question presents a multiple sequence alignment (FASTA) and a phylogenetic tree, and asks whether structured homoplasy is present and which taxa are involved.
| Task | RC | RC min | RC max | Files | Total questions | Assesses |
|---|---|---|---|---|---|---|
| REL-B1 | Number of homoplastic taxa | 2 | 25 | 001–012 | 7,952 | Homoplasy detection — identifying convergent evolution where distantly related taxa independently share nucleotide motifs across many alignment columns more than expected by chance; tests relational pattern recognition over long biological sequences |
Chemistry (4,016 questions)
| Task | RC | RC min | RC max | Questions | Assesses |
|---|---|---|---|---|---|
| REL-C1 | 2 | 2 | 2 | 1,000 | Constitutional isomer identification — given a set of SMILES, determine whether all molecules share the same molecular formula (same atoms, different connectivity); tests recognition of molecular equivalence relations |
| REL-C2 | 2 | 2 | 2 | 1,016 | Maximum common substructure (MCS) — find the largest single connected chemical motif present in every molecule in the set; tests relational reasoning over graph structure |
| REL-C3 | Number of isomers | 1 | 92 | 1,000 | Isomer set completion — given a partial set of constitutional isomers, identify all missing members; tests exhaustive enumeration of a structural equivalence class |
| REL-C4 | Number of molecules | 5 | 50 | 1,000 | Constrained motif extraction — extract one substructure from each of the molecules such that a specified functional group count sums to a target value; tests joint constraint satisfaction over molecular substructures |
Citation
@article{fesser2026rel,
title = {Evaluating Relational Reasoning in LLMs with REL},
author = {Lukas Fesser and Yasha Ektefaie and Ada Fang and Sham M. Kakade and Marinka Zitnik},
year = {2026},
journal = {arXiv preprint arXiv:2604.12176},
eprint = {2604.12176},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2604.12176}
}