AptMQL-Bench
Harbor: pendingA text-to-MQL benchmark with 21 document databases and 3,186 natural language requests, converted from BIRD using access-pattern schema design.
Auto86%Coding
- Samples
- 3,186
- Released
- 2026
- Scores
- 8
CSEG-Bench
Cross-Source Evidence Grounding Benchmark
A benchmark of 1,160 verified search agent trajectories measuring cross-source evidence grounding via a six-layer decision tree hierarchy.
How to run
ltb · cseg-bench
$ npm i -g lettokenburn-cli$ ltb bench info cseg-bench
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
12 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| GLM-5.3 Flash | 0.9 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| DeepSeek V4 Pro | 0.8 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| Hy3 | 0.8 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| DeepSeek V4 Flash | 0.6 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| GLM-5.2 | 0.6 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| GPT-oss 120B | 0.6 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| gpt-oss-20b | 0.5 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| Qwen3-32B | 0.5 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| Muse Glimmer 30B | 0.5 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| Qwen3 8B | 0.4 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| Qwen3 4B | 0.4 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| google/gemma-4-31B-it | 0.3 | 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.
Reproduced by the community
All runsLoading community runs…
About
CSEG-Bench comprises 1,160 expert-adjudicated search trajectories constructed by rewriting answers on real retrieved document pools to cover nine canonical paths. It evaluates source independence, conflict handling, and attribution.
Scoring uses a hierarchical label (1-7) based on a six-layer decision tree. The primary metric is exact accuracy, with the best reported reward model achieving 83.9%.