Meno studies how a system of humans and AI can produce beliefs that track the world, rather than settling into a self-consistent equilibrium the world had nothing to do with. This site is the entry point to each project: what it asks, what it is trying to establish, and what is still open.
Each project page below is an overview. The pages beneath it carry the formalism, the experimental lines and the results, and most of those require a passcode.
Projects
| Project | What it is | People |
|---|---|---|
| Verifier | Checks claims against the files and runs behind them, as a gate not a report | Zhonghao, Tianyi, Unveil |
| Unveil Code | The team’s research agent in Slack | Ahmed |
| Human Study | Does AI that pushes on your reasoning improve forecasting? | Tianyi, Changbai, Zhonghao, Rachel, Riza, David |
| Martingale Training | Training models to update beliefs like a martingale | Zhonghao, Tianyi, Maxwell, Yuhe, Ahmed |
| Learning from a Learner (LFL) | Assistants that help a human learner track the truth | Tianyi, Zhonghao |
| Recursive Co-Improvement | Humans and AI improving each other over many rounds | Zhonghao, Tianyi |
| Reflection Game | A role-play game that elicits what people prefer | Zhonghao, Tianyi |
| Psychohistorians | A simulation testbed for group epistemics | Tianyi, Unveil |
| Lock-in Hypothesis | Published paper (ICML 2025) on AI entrenching today’s values | Tianyi, Zhonghao |
| Site & Web Apps | meno.sh, the CoResearch app and this hub | Tianyi, Zhonghao, lcb931023 |
| Unveil | A learning engine for questions textbooks cannot answer | Zhonghao |
| Gaokao Counsellor | The Reflection Game’s college-application variant | Zhonghao, Tianyi |
Docs
- Authorship Inclusion Policy — who is included as an author on a paper, and who is acknowledged instead.
- Managing Many Parallel Experiment Runs — the convention every project’s run log follows.
- Multimodal Data for Human Intent — a pre-project memo on whether the algorithmic and data-side approaches to reading human intent can share a benchmark.