Measuring model performance on economically valuable work
What GDPval reveals about evaluating frontier models on realistic professional deliverables across major industries.
Research
Advancing how AI learns and performs.
We believe the next leap in AI capability will come from learning from real expert work. Our research explores how expert judgment, high-quality training data, rigorous evaluations, and realistic environments can help models reason, navigate tradeoffs, and perform reliably in the real world.
Research we follow
Work shaping how we build.
We follow research grounded in expert judgment, economically meaningful tasks, and reliable performance across complete professional workflows.
A practical framework for evaluating AI agents
How teams can design agent evaluations that measure outcomes, survive non-determinism, and improve alongside the product.
What summarization taught us about learning from human feedback
How preference data, reward modeling, and reinforcement learning helped smaller models produce summaries people preferred.
External benchmarks
Learning from how advanced AI is measured.
We study leading independent benchmarks to understand how complex capabilities are measured, where current approaches fall short, and what real-world performance demands. These insights shape the benchmarks and evaluation environments we develop at Alpheva.
01
OSWorld
A useful reference for cross-application workflows, stateful tasks, and execution-based evaluation in real computer environments.
02SWE-bench
A benchmark for evaluating whether AI systems can resolve real-world GitHub issues by modifying codebases and passing repository tests.
03Terminal-Bench
A benchmark of hard, realistic terminal tasks with purpose-built environments, human-written solutions, and comprehensive test-based verification.
Work with Alpheva
Research a hard model problem with us.
We collaborate on evaluation and training questions grounded in real work.
Build with us