DesignReviewBench (Sashiko agent) - F1: leaderboard
DesignReviewBench design-aware code review of Linux kernel patch series (2025-2026 patches): the reviewer reports design trade-offs, judged against reference trade-offs mined from maintainer review threads as Match/Partial/Miss (1/0.5/0) and scored by graded F1, 0 to 1. Sashiko agent-based kernel review system (strongest baseline in the paper). The column model is the agent backbone; reasoning settings are not disclosed. Metric: Graded F1. Higher is better. Source: https://arxiv.org/html/2610.21925v1
Metric: Graded F1. Source: arxiv.org. Saturation forecast: Around 2028. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V4.1 Flash | 0.26 | #35 |
| 2 | GPT-5.6 Terra | 0.22 | #41 |
| 3 | GPT-5.6 Luna | 0.21 | #68 |
| 4 | Claude Haiku 4.5 | 0.12 | #279 |
No result here: #4 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=designreviewbench-sashiko-agent-f1 · How It Works · Data refreshed daily, snapshot 2026-10-18.