VIDRAFT Hits 90.9% on GPQA Diamond, Taking the #1 Spot on AI Scientific Reasoning Leaderboards
TL;DR: VIDRAFT, a Korean Pre-AGI AI startup, has achieved a 90.9% score on the GPQA Diamond benchmark — currently the top result on scientific reasoning leaderboards. GPQA Diamond is one of the hardest publicly available reasoning benchmarks, composed of PhD-level questions in biology, chemistry, and physics. If you're tracking frontier model capabilities or building science-domain AI applications, this result is worth paying attention to.
What it is
GPQA Diamond (Graduate-Level Google-Proof Q&A) is a rigorous multiple-choice benchmark designed to be effectively unsolvable by Google search alone. Questions are authored and validated by domain experts at the PhD level, spanning:
- Biology (molecular, cellular, systems)
- Chemistry (organic, physical, inorganic)
- Physics (quantum mechanics, electromagnetism, thermodynamics)
The "Diamond" subset is the hardest tier of the GPQA dataset, representing questions that stumped even non-expert PhDs during validation. Human expert performance on this subset sits in the 65–70% range, making scores above ~85% a meaningful signal of genuine scientific reasoning capability rather than surface-level pattern matching.
VIDRAFT is a South Korean AI startup positioning itself in the Pre-AGI space — meaning the company is explicitly targeting systems capable of complex, multi-step reasoning across scientific and general domains, not just narrow task performance.
The reported result: 90.9% on GPQA Diamond, claimed as a #1 ranking on the scientific reasoning leaderboard at the time of publication (2026-06-16).
How it works
While VIDRAFT has not publicly disclosed architectural specifics or training hyperparameters, the broader context of achieving state-of-the-art results on GPQA Diamond points to a few well-understood research directions that frontier labs are currently pursuing:
- Long-chain reasoning / "thinking" models: High GPQA Diamond scores are typically associated with models trained to produce extended internal reasoning traces before arriving at a final answer — sometimes called chain-of-thought (CoT) or "reasoning models" in the style of recent o-series or R-series architectures.
- Scientific corpus alignment: Strong performance on PhD-level domain questions generally requires both broad pretraining on scientific literature and targeted post-training that reinforces structured, verifiable reasoning.
- Reinforcement learning from verifiable rewards: Benchmarks with objectively correct answers (like multiple-choice science questions) are well-suited to RL-based fine-tuning approaches where correctness itself serves as a reward signal — without the need for human preference annotation.
These are conceptual directions consistent with public research; VIDRAFT's specific implementation details have not been disclosed.
Benchmarks & results
The single confirmed public result from the source article:
| Benchmark | Score | Standing |
|---|---|---|
| GPQA Diamond | 90.9% | #1 (leaderboard) |
- GPQA Diamond represents the hardest subset of the GPQA dataset
- The 90.9% figure surpasses the human expert baseline (~65–70%) by a substantial margin
- VIDRAFT claims this as the current top result on AI scientific reasoning leaderboards as of the publication date
No additional benchmark numbers (MATH, AIME, MMLU, LiveCodeBench, etc.) were reported in the source article. Further results may be published by VIDRAFT through official channels.
How to try it
At the time of this article's source publication, VIDRAFT has not publicly announced a generally available model release via Hugging Face, GitHub, or an OpenAI-compatible API endpoint.
If and when VIDRAFT makes their model or API publicly accessible, typical access patterns for frontier models would look like:
# Hypothetical Hugging Face download (placeholder — not confirmed live)
# huggingface-cli download vidraft/<model-name>
# Hypothetical OpenAI-compatible API call (placeholder — not confirmed live)
# curl https://<api-endpoint>/v1/chat/completions \
# -H "Authorization: Bearer $API_KEY" \
# -H "Content-Type: application/json" \
# -d '{"model": "<model-name>", "messages": [{"role": "user", "content": "..."}]}'
⚠️ Note: The commands above are structural examples only. No public endpoint, model name, or API key scheme has been confirmed by VIDRAFT. Check VIDRAFT's official channels for announcements.
FAQ
Q: How does 90.9% on GPQA Diamond compare to other publicly known models?
A: As of mid-2026, GPQA Diamond scores above 85% are considered frontier-tier. The 90.9% figure, if independently verified on the standard evaluation protocol, would represent a meaningful step beyond previously reported public scores. The leaderboard ranking claimed in the source article places it at #1 among tracked models at publication time.
Q: Is GPQA Diamond a reliable signal for real-world scientific reasoning, or is it a benchmark that can be "gamed"?
A: GPQA Diamond is considered one of the more contamination-resistant benchmarks because questions are deliberately designed to be unsearchable and require integrated domain understanding. That said, as with any static benchmark, the field watches for signs of overfitting to evaluation distributions. Independent replication and performance on held-out or newly released tasks provides stronger evidence.
Q: When might VIDRAFT release model weights or API access for developers?
A: The source article does not announce a public release timeline. Developers interested in early access or API partnerships should follow VIDRAFT's official communications directly.
Originally reported by v.daum.net (2026-06-16) — source article.
Top comments (0)