ResearchFirst results coming soon
Benchmarks written and graded by domain experts, using the same rubrics we use to vet people. We extend well-known open benchmarks with fresh, private questions, then score models with the benchmark's own code.
Same question format, same answer key style, same scoring code. Nothing about the test changes except who wrote the questions.
Written and cross-checked by specialists in our network. Kept private, so no model has seen them in training.
Each model's score on the original set next to its score on ours. The gap shows how much headroom is left.
Science first, then the benchmark we care about most: India's languages and knowledge.
Science
In progress
Kodiva-Extended
Graduate-level physics, chemistry and biology questions that can't be answered by searching. New questions written by PhDs in our network.
Indian languages
Next
Kodiva-Extended
India-specific knowledge across 11 languages, from regional exams to culture and law. New questions from native-speaker subject experts.
Law
Planned
Extension of LegalBench
Contract, corporate and tax reasoning under Indian law, written by practising Indian lawyers.
Every model is scored on the original questions and on ours. We publish when the first runs are checked, not before.
GPQA Diamond · Kodiva-Extended
| Rank | Model | Original | Kodiva-Extended | Gap |
|---|---|---|---|---|
| 01 | ||||
| 02 | ||||
| 03 | ||||
| 04 |
First results coming soon
Methods, datasets and findings from our work with labs.
Get the leaderboard when it goes live, or ask us to evaluate your model before it does.
Thanks. We'll email you when the first results are published.
Or request a sample of our expert data →