# RACE-Bench: A Reasoning-Augmented Benchmark for Repository-Level Code Agents on Feature Addition

arXiv: https://arxiv.org/abs/2603.26337 · HTML: https://mlsanity.com/paper/2603.26337/

## Key facts

- Cited by 4 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 27 Mar 2026.

## Built on by (most cited first, 4 of 4)

- [PDEAgent-Bench: A Multi-Metric, Multi-Library Benchmark for PDE Solver Generation](https://mlsanity.com/paper/2605.09636/) (2026)
- [A Unified Issue Resolution Benchmark for Requirement Clarification, Planning, and Code Generation for Coding Agents](https://mlsanity.com/paper/2608.09072/) (2026)
- [SWE-Journey: Towards More Realistic Evaluation of Coding Assistants through Long-Horizon, Multi-Turn Interaction](https://mlsanity.com/paper/2610.11559/) (2026)
- [SWE-Duel: Towards a Self-Scaling Arena for Coding Agents](https://mlsanity.com/paper/2610.12569/) (2026)

## Abstract (arXiv)

Repository-level code agents have shown strong promise in real-world feature addition tasks, making reliable evaluation of their capabilities increasingly important. However, existing benchmarks primarily evaluate these agents as black boxes based on final test correctness, providing limited insight into how they reason and where failures arise. To address this limitation, we introduce RACE-bench, a reasoning-augmented benchmark for evaluating code agents on repository-level feature addition tasks. RACE-bench contains 528 real-world feature addition instances from 12 open-source repositories. Each instance is paired with executable patch verification and structured intermediate reference reasoning covering issue understanding, file localization, implementation tasks, and step decomposition. Based on this design, we introduce a dual-track evaluation framework that jointly measures patch correctness and intermediate reasoning alignment with developer-accepted reference trajectories. We evaluate three representative repository-level code agents on RACE-bench. On the full benchmark, Resolved Rate ranges from 29% to 70% across different agents. Our reasoning-level analysis further shows that while current agents perform well at understanding high-level intent, their performance degrades substantially when translating intent into concrete implementation steps. We also find patches that can be applied but still fail the tests cover fewer reference-reasoning elements (35.7% lower recall) and contain more unsupported reasoning elements (94.1% higher over-prediction) than successful patches. These findings highlight the importance of evaluating repository-level code agents beyond final patch correctness by examining the quality of their reasoning processes.