CRUXEval
Harbor adapterCRUXEval is a benchmark of 800 Python functions paired with input-output pairs.
Verified86%Coding
- Samples
- 800
- Released
- 2024
- Scores
- 0
CRUXEval-style
Evaluating Exact Output and Checkpoint-State Prediction in Real Programs
A benchmark of 400 cases from 371 Python/C++ programs measuring model ability to predict final output and intermediate checkpoint states without code execution.
How to run
ltb · cruxeval-style
$ npm i -g lettokenburn-cli$ ltb bench info cruxeval-style
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
Scores
7 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| GPT-5.6 Sol | 74.8 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| DeepSeek V4 Pro 0813 | 63.8 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| GLM-5.3 | 55.3 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| Qwen3.8 27B | 39.4 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| GPT-5.6 Sol Off | 24.5 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| DeepSeek V4 Pro 0813 Off | 8.6 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| Qwen3.8 27B Off | 6.3 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsLoading community runs…
About
The benchmark comprises 400 cases derived from 371 distinct Python and C++ programs, including contest solutions and algorithms. It features paired short/long trace inputs and checkpoint state predictions.
Scoring evaluates exact match on gradable responses across seven model settings. The strongest setting achieved 93.0% on short outputs but dropped to 65.5% on state tasks, highlighting difficulties with longer executions.