A Domain-Specific Probabilistic Programming Language for Reasoning about Reasoning(…)
summary
Memo is a domain-specific probabilistic programming language for recursive theory-of-mind models that compiles to JAX and runs on GPUs. It embeds in Python, offers primitives like agents who make choices instead of raw sampling, and is end-to-end differentiable. A reference experiment (Ching & Franca 2015) is reproduced: zero recursion gives 50/50 but one recursive level predicts the observed ~2× preference for the green circle. Implementations report orders of magnitude faster runtimes, artifacts available and results reproduced.
AI-written summary · about
chapters (10)
- 0:00Memo for recursive reasoning
- 1:25Green object selection game
- 3:10Rational speech acts framework
- 4:13Recursive rationality challenges
- 6:21Memo for correctness and efficiency
- 7:27Demo: green object model in Memo
- 10:56Memo vs traditional probabilistic programming
- 11:59Memo adoption and impact
- 12:40Future directions
- 13:24Q&A
related talks
24mHigher order programming with probabilistic effects: A model of stochastic memo...
47mThe State of the Art for Probabilistic Programming
24mRecursive Coding Agents
23mProbabilistic Programming with Vectorized Programmable Inference
22mReasoning about 'Reasoning about Reasoning' Semantics and Contextual Equivalence
46mBlueprints for a Universal Reasoning Machine
35mScalable structure learning and inference for domain-specific probabilistic prog
39mGauguin, Descartes, Bayes: A Diurnal Golem's Brain