Foundational Automatic Evaluators: Scaling Multi-Task Generative Evaluator Training for Reasoning-Centric Domains

TL;DR

This paper introduces FARE, a large-scale data-driven approach to training reasoning-centric evaluators for language models, demonstrating superior performance over existing specialized evaluators in multiple real-world tasks.

Contribution

The paper presents a new data scaling methodology and the FARE family of large, open-source evaluators trained with simple supervised finetuning, setting new standards in reasoning evaluation.

Findings

FARE-20B surpasses 70B+ specialized evaluators in benchmarks.

FARE-20B achieves near-oracle performance as inference-time rerankers.

Fine-tuned FARE improves RL training outcomes by up to 14.1%.

Abstract

Finetuning specialized generative evaluators has emerged as a popular paradigm to meet the increasing demand for scalable evaluation during both training and test-time. However, recent work has largely focused on applying new methodology, such as reinforcement learning (RL), to training evaluators, shying away from large-scale, data-driven development. In this work, we focus on data scaling, curating a set of 2.5M samples spanning five unique evaluation tasks (pairwise, step-level, reference-free and reference-based verification, and single rating) and multiple domains focused on reasoning evaluation. With our data, we train Foundational Automatic Reasoning Evaluators (FARE), a family of 8B and 20B (with 3.6B active) parameter evaluators, with a simple iterative rejection-sampling supervised finetuning (SFT) approach. FARE-8B challenges larger specialized RL-trained evaluators and FARE-20B sets the new standard for open-source evaluators, surpassing specialized 70B+ evaluators. Beyond static benchmarks, we evaluate FARE in real-world tasks: As inference-time rerankers, FARE-20B achieves near-oracle performance on MATH. As verifiers in RL training, FARE improves the downstream RL-trained model performance by up to 14.1% vs. string-matching verifiers. When initialized from FARE, a continually-finetuned FARE-Code outperforms gpt-oss-20B by 65% on evaluating test-case quality.