Saved in:
Bibliographic Details
Main Authors: Xu, Zhiqiu, Jin, Shibo, Arya, Shreya, Naik, Mayur
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2604.21916
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910168328962048
author Xu, Zhiqiu
Jin, Shibo
Arya, Shreya
Naik, Mayur
author_facet Xu, Zhiqiu
Jin, Shibo
Arya, Shreya
Naik, Mayur
contents As frontier language models attain near-ceiling performance on static mathematical benchmarks, existing evaluations are increasingly unable to differentiate model capabilities, largely because they cast models solely as solvers of fixed problem sets. We introduce MathDuels, a self-play benchmark in which models occupy dual roles: each authors math problems under adversarial prompting and solves problems authored by every other participant. Problems are produced through a three-stage generation pipeline (meta-prompting, problem generation, and difficulty amplification), and validated by an independent verifier that excludes ill-posed questions. A Rasch model (Rasch, 1993) jointly estimates solver abilities and problem difficulties; author quality is derived from the difficulties of each model's authored problems. Experiments across 19 frontier models reveal that authoring and solving capabilities are partially decoupled, and that dual-role evaluation reveals capability separations invisible in single-role benchmarks. As newer models enter the arena, they produce problems that defeat previously dominant solvers, so the benchmark's difficulty co-evolves with participant strength rather than saturating at a fixed ceiling. We host a public leaderboard that updates as new models are released.
format Preprint
id arxiv_https___arxiv_org_abs_2604_21916
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle MathDuels: Evaluating LLMs as Problem Posers and Solvers
Xu, Zhiqiu
Jin, Shibo
Arya, Shreya
Naik, Mayur
Computation and Language
Software Engineering
As frontier language models attain near-ceiling performance on static mathematical benchmarks, existing evaluations are increasingly unable to differentiate model capabilities, largely because they cast models solely as solvers of fixed problem sets. We introduce MathDuels, a self-play benchmark in which models occupy dual roles: each authors math problems under adversarial prompting and solves problems authored by every other participant. Problems are produced through a three-stage generation pipeline (meta-prompting, problem generation, and difficulty amplification), and validated by an independent verifier that excludes ill-posed questions. A Rasch model (Rasch, 1993) jointly estimates solver abilities and problem difficulties; author quality is derived from the difficulties of each model's authored problems. Experiments across 19 frontier models reveal that authoring and solving capabilities are partially decoupled, and that dual-role evaluation reveals capability separations invisible in single-role benchmarks. As newer models enter the arena, they produce problems that defeat previously dominant solvers, so the benchmark's difficulty co-evolves with participant strength rather than saturating at a fixed ceiling. We host a public leaderboard that updates as new models are released.
title MathDuels: Evaluating LLMs as Problem Posers and Solvers
topic Computation and Language
Software Engineering
url https://arxiv.org/abs/2604.21916