Humanlaya

Built by model builders, for model builders.
Humanlaya engineers expert-level data and verifiable rewards for frontier AI labs.

[Humanlaya @Humanlayadata]

GPT-5.6 is a growth cheat code! 🚀 It instantly turns complex metrics into actionable data dashboards. Deep analysis now takes seconds. 📊

[Humanlaya @Humanlayadata]

$OneMillion-Bench is our step toward that.
See our work at：📄 <paper> https://t.co/1RlyncJIOv | 💾 <data/code> https://t.co/ZrKpRxq4tI | 📝 <blog> https://t.co/CVvTXPsxPE | 🤗 <hf>https://t.co/hnb0NsrJW6

[Humanlaya @Humanlayadata]

Dollars Matter, Scores Flatter. 💰
Can LLM agents really do high-valued experts work end-to-end — not just “solve a question”?
Introducing $OneMillion-Bench benchmarks real expert work, priced in dollars 💵

[Humanlaya @Humanlayadata]

Why it matters + resources 🔗
If we want agents in finance/law/medicine/engineering, we need evaluation that matches real workflows + real cost of mistakes.

[Humanlaya @Humanlayadata]

What moves the needle 🔧 Tools and evaluation design matter a lot.
• Official scaffolds > OpenRouter > no search (consistent gains)
• Rankings are stable across different judge models (robustness check)
• We also stress-test temporal sensitivity and test-time scaling behavior.

[Humanlaya @Humanlayadata]

How Does We Grade? ✅ ❌
We only count value when the output is deliverable, not “partially correct.”
• 15–35 expert rubrics per task + negative rubrics for hallucination/unsafe/non-compliance
• Asymmetric weights (-20 to +10) to reduce reward hacking

[Humanlaya @Humanlayadata]

From high-economic-value tasks to verifiable rewards, we help frontier teams push model capability beyond benchmark optimization and toward real-world cognitive performance.

[Humanlaya @Humanlayadata]

🥳 Introducing Humanlaya — The height of AGI, driven by the peak of Human Intelligence.

[Humanlaya @Humanlayadata]

We design expert-level tasks, build rigorous verification systems, and turn high-density human knowledge into scalable training signals for advanced models.