Fezvo Labs
Real PMIDs, Wrong Papers accepted as a poster at the NeurIPS 2026 AgenticLS workshop.
Details
Independent AI research lab
Elements in progress
The system around the model.
Fezvo Labs studies how agents are trained, how they act, and how they stay safe.
Recent
News
Work
Research
RL environments
Worlds with actions, observations, and rewards, built so agent behavior can be trained and measured.
Post-training
Preference optimization, reinforcement learning, and the data that turns a base model into a policy.
Agentic systems
Models that plan, call tools, and carry a task across many steps.
Harness design
The tools, prompts, memory, budgets, and checks wrapped around a model.
AI safety
How agent behavior fails, and what has to be true before a system can be trusted with a real task.
AI for healthcare
Agents on biomedical questions, where a cited source has to support the claim it is attached to.
Elements Now
The current project. Environment, post-training, and harness, studied as one system.
Publications
2026
Real PMIDs, Wrong Papers
Closed-book models cite PubMed IDs that exist, and almost never the paper behind the claim.