LetTokenBurnSWEPatchTest: Evaluating Software Tests Beyond Reproduction
A test-generation benchmark with 619 tasks and 82,579 candidate patches measuring issue reproduction, correct-patch acceptance, and incorrect-patch rejection via Joint Validation Rate.
$ npm i -g lettokenburn-cli$ ltb bench info swepatchtest
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Claude Opus 5 | 75.1 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| Qwen3.8 Max | 71.4 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| Claude Fable 5.1 | 71.4 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| GPT-5.6 Sol | 64.1 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| Kimi K3 | 54.0 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
SWEPatchTest comprises 619 repository-level tasks across ten languages, paired with 82,579 candidate patches collected from coding-agent trajectories. It evaluates generated test suites by replaying them on buggy, reference-patched, and candidate-patched codebases.
The primary metric is Joint Validation Rate (JVR), requiring simultaneous issue reproduction, complete correct-patch acceptance, and complete incorrect-patch rejection. Best-performing configurations achieve only 9.21% JVR, exposing limitations hidden by single-reference evaluation.