Evaluation
Rigorous benchmarks and metrics that reveal true model capabilities beyond surface-level performance.
How do we know if AI is actually getting smarter? We build the tests.
3 benchmarksPractice makes
intelligence.
We build AI agents that improve through deliberate practice — the same mechanism behind every elite human performer.
Benchmarks · Agents · Open Source · Published at ICML, NeurIPS
3 open-source models · 6 research papers · Apache 2.0
Our Focus
Agents
Autonomous systems that learn and improve through deliberate practice and structured self-evaluation.
AI that learns from practice, just like humans do.
2 frameworksVision
Multimodal understanding that bridges visual perception with deep reasoning and world knowledge.
Teaching AI to understand what it sees, not just describe it.
1 toolkitOur Models
EtudeEval-7B
Can spot when AI is cheating on tests — and build harder ones on the fly
7B parameters
Contamination detection: 94.2% · Adaptive difficulty: ±0.3 calibration
Dynamic evaluation model. Adversarial benchmark generation that adapts to model capabilities in real time.
View on GitHub →PracticeAgent-13B
Helps AI learn from its mistakes, like a coach running drills
13B parameters
SWE-bench: 45.2% · 40% faster skill acquisition vs. static training
Self-improving agent. Deliberate practice loops that drive compounding performance gains over time.
View on GitHub →VisPractice-3B
Teaches AI to truly understand what it sees in photos and diagrams
3B parameters
MMMU: 51.8% · Compositional VQA: 67.3%
Multimodal vision model. Compositional reasoning that bridges visual perception with structured world knowledge.
View on GitHub →Our Team Has Roots At
Stanford NLP
MIT CSAIL
Berkeley AI
CMU LTI
Oxford NLP
ETH Zürich
3
Open-Source
Models
Models
6
Research
Publications
Publications
7
Blog Posts &
Technical Essays
Technical Essays
8
Open Roles
Hiring Now
Hiring Now
Latest Writing
Research notes, technical essays, and dispatches from the frontier.
Beyond Task Automation: The Case for Agents That Truly Learn
Most AI agents don't actually learn from experience. We argue the next breakthrough isn't a bigger model — it's a better learning process.
Read moreIntroducing EtudeEval: A Dynamic Benchmark for Agent Learning
A benchmark framework that measures skill acquisition rate — how quickly agents improve through structured practice.
Read moreFrom Self-Play to Deliberate Practice: What AlphaZero Teaches Us
AlphaZero's revolution wasn't about scale — it was about the quality of the learning process.
Read more