Tractable Agreement Protocols: consensus without Bayes or priors
Aaroth · x · 2026-11-02
Aaron Roth's team presents arXiv paper Tractable Agreement Protocols (2411.19791): an efficient reduction that turns any ML algorithm into an interactive protocol where a model and a human iteratively converge on predictions.
Key points:
- Replaces Bayesian rationality with calibration: each round's belief only needs to be calibrated conditional on the prior round — computationally and statistically tractable, and sensible even in prior-free settings.
- Recovers Aaronson's 2005 fast convergence theorem under weaker assumptions for one-dimensional expectations.
- Extends to beliefs over d-outcome distributions with two feedback mechanisms (vector estimates, and a decision-theoretic one where the human reports their utility-maximizing action each round); rounds-to-agreement is independent of d.
A significant generalization of Aumann's classic agreement theorem to a computationally feasible, assumption-light framework.
Related event: Economist Aaron Roth Shows Agreement Theorem Holds Without Common Priors(4 posts)→
More from coding & agent
- VoiceFlip launches: fully autonomous AI voice agents for SMB sites in 2-3 minutes — modelcontextprotocol · 2026-11-02
- X MCP Server lets agents post, search and reply on Twitter via the X API — modelcontextprotocol · 2026-11-02
- AgentTM: a Go-built time-travel debugger records and rewinds AI coding agent sessions — tracellm · 2026-11-02
- LeRobot ships with Trackio: one flag for live robot training metrics — Gradio · 2026-11-02
- From chat memory to knowledge graph: a two-layer wiki knowledge base pattern — adnan_hashmi · 2026-11-02
- The approval queue isn't a tax on your agents — it's the only part making them smarter — RobCatalanoGTM · 2026-11-02