A Domain-Specific PPL for Reasoning about Reasoning (or: a memo on memo)
summary
Introduces memo, a Python-embedded domain-specific PPL that replaces sample/infer with choose/observe semantics and compiles recursive theory-of-mind models into JAX-backed array compilation so nested agent reasoning runs on GPU. Demonstrates the Lewis one-word game, reproduces Ching & Frank's green-circle bias after adding a recursive speaker, and reports typical speedups of ~10x to several orders of magnitude versus textbook implementations, end-to-end differentiability for gradient fitting, and uptake in courses and research groups.
AI-written summary · about
chapters (13)
- 0:00Lewis one-word reference game
- 1:27Recursive reasoning in communication
- 3:14Recursive rationality paradigm
- 5:01Correctness and efficiency challenges
- 7:30Memo for recursive rationality
- 8:53Demo: memo in one-word game
- 10:59Adding recursive speaker reasoning
- 12:00Memo vs traditional PPL
- 13:45Memo performance and applications
- 14:26Memo correctness questions
- 15:30Inference process in memo
- 16:11Natural language and intelligence
- 17:01Q&A
related talks
24mRecursive Coding Agents
15mCurrent Practices for Building LLM-Powered Reasoning Tools Are Ad Hoc—and We Can Do Better
31mMachine Reasoning and System 2 Thinking
36mRethinking legacy modernisation with automated reasoning and comprehension models
31mBeyond LLMs and Maturing RAG use cases
01h 49mBuilding Memory Augmented Agents with MemoRizz
30mToward Programming Languages for Reasoning -- Humans, Symbolic Systems, and AI ...
36mPanel Discussion