Multi-Agent Systems for Traceable Bayesian Workflow
summary
A multi-agent system assembles prompts and tools to run an end-to-end Bayesian workflow: it generates Stan models, runs prior and simulation-based checks, attempts parameter recovery, runs MCMC and posterior predictive checks, and writes an audit trail of every model choice. Using Stan and an LLM coordinator (Claude), the pipeline records 1228 lines of prompt and exploration traces; on PosteriorDB it matches human baselines, with nine planned experiments on a car example and 6 of 18 dataset runs rejected for failed inference.
AI-written summary · about
chapters (9)
- 0:00Multi-agent Bayesian workflow intro
- 0:43Motivation for automating Bayesian modeling
- 2:26AI agent system definition
- 4:13Bayesian workflow loop overview
- 6:18Multi-agent orchestration in natural language
- 7:43Cars dataset exploration trace example
- 9:53Comparison with BoxLM and related work
- 11:19Traceability in modeling focus
- 11:42Conclusion and Q&A
related talks
01h 03mAn Overview of Probabilistic Programming
36mTalk by Kamalika Chaudhuri (Meta)
31mBuilding Agentic AI Platforms
41mBuilding LLM Agents – Challenges and Learnings
26mAI Agents as Financial Co-Pilots | AI-Driven Strategy & Decision Making
01h 30mBayBE: A Bayesian Back End for Experimental Planning in the Low-To-No-Data Regime
01h 04mAgentic AI: Build a Multi-Agent Application with CrewAI
25mDirecting a Swarm of Agents for Fun and Profit