BulkPR-Bench - Critical Relation Recall: leaderboard
The relation-recovery diagnostic of BulkPR-Bench (2026): criticality-weighted recall of the gold conflicts, dependencies, all-or-none groups and must-reject candidates that the coding agent records in its typed relation ledger while governing a queue of interacting pull requests. Percent; higher is better.
Metric: Critical relation recall (%; criticality-weighted recall of the gold relation atoms (conflicts, dependencies, all-or-none groups, must-reject candidates, duplicates, superseding changes) in the agent's final typed relation ledger; same buffered K=32 runs of the Claude Code v2.1.138 scaffold, repository-weighted). Source: arxiv.org. Saturation forecast: Around February 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.8 | 57.7 | #56 |
| 2 | GLM-5.2 | 54.3 | #138 |
| 3 | GPT-5.4 | 52.1 | #123 |
| 4 | Kimi K3 | 50.3 | #54 |
| 5 | Qwen 3.7 Max | 42.9 | #145 |
| 6 | DeepSeek V4 Pro | 35.2 | #188 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=bulkpr-bench-critical-relation-recall · How It Works · Data refreshed daily, snapshot 2026-11-13.