Research surfaces
- Lean-based reward oracles
- Code and proof generation
- GRPO and reinforcement learning loops
- Tactic diversity and reasoning depth
- Curriculum design for reasoning tasks
- Compute-to-intelligence dynamics
Research
Modus Verify: verifier-guided RL for reasoning systems.
Modus Verify studies how formal verifiers, tests, reward signals, and curriculum design can improve model reasoning, proof generation, code generation, and agent behavior.
2% to 72%proof success in a controlled Lean-reward training run
Binary rewardformal verifier feedback instead of preference theatre
Reasoning depthtracking tactics, trajectories, regressions, and breakouts
What we are trying to understand
- How do models acquire reusable reasoning infrastructure?
- When does RL produce brittle memorisation versus real capability?
- How should theorem and code tasks be sequenced?
- Can verifier feedback become a general training substrate for agents?