CS372: Artificial General Intelligence for Reasoning, Planning, and Decision Making

Announcements

Course Overview

Join us for an audacious aim: elevating LLMs from pattern matching to validated reasoning, grounded planning, and accountable decision making.

Course Description

Contrary to the hype, AGI is not around the corner. Today's large language models are a remarkable System 1: a reflexive, associative substrate that predicts the most likely next token and, in doing so, honors the popular over the accurate. A single maximum-likelihood model cannot validate its own output, because its errors are correlated with the premises it reasons from. It cannot tell association from cause. It carries no persistent memory or transactional discipline across a long task, so a plan that fails at step forty cannot be audited, undone, or repaired. And it does not know when the world has exceeded what its context describes. Scaling makes such a model more fluent; it does not make it reason, validate, plan, or ground.

The three volumes of The Path to AGI, this course's textbooks, explain what is missing and how to build it. Volume 1 (Multi-LLM Collaborative Intelligence) shows that validated reasoning, checks and balances, and long-horizon planning with persistent transactional memory are fostered through collaboration among models rather than extracted from one. Volume 2 (System-2 Reasoning) replaces association with causal reasoning and adds the meta-cognitive questions a what-only, reward-trained model never asks: why, how, and when. Volume 3 (Beyond Intelligence) binds that intelligence to the physical and social world through active memory, grounding obligations, and the four Quadrivium disciplines of Reason, Elicit, Contract, and Bridge, and it states what has been proved, what has been demonstrated, and what remains open on the road from operational AGI toward wisdom.

Why you should take this course. Nearly every AI system built in the coming decade will sit on top of an LLM, and the engineers and researchers who matter will be those who understand exactly where the architecture fails and how to build around it. In this course you will learn to diagnose those failures at their source rather than patch them with prompts; to design multi-agent systems with validation, persistent memory, and transaction guarantees that go beyond today's agent frameworks; to ground an agent's decisions in causal structure and in the world it acts on; and to measure, rather than assume, whether a system meets a safety goal. The course is project-driven. Teams build and evaluate on the open-source MACI runtime and the MACI-Bench suite across domains such as medical diagnosis, causal discovery, physical-world grounding, planning, and disaster response, and the strongest projects are expected to become conference submissions.

Core Questions

1. Why can a single maximum-likelihood model not validate its own reasoning, and what architecture can?
2. How do we move LLM reasoning from association to causation, and from what to why, how, and when?
3. How do we give long-lived agentic workflows persistent memory, validation, and transaction guarantees?
4. What does it mean for an agent to be grounded in the physical and social world, and how is grounding enforced?
5. When does multi-LLM collaboration help, when does it hurt, and how should it be moderated?
6. What has been proved, what has been demonstrated, and what remains open on the path from AGI to wisdom?

The Course Examines

1. System 1 and System 2: the LLM as an unconscious substrate, and what deliberate reasoning requires
2. Socratic reasoning, Pearl's causal hierarchy, and semantic anchoring (UCCT)
3. Meta-cognitive judgment: sycophancy and skepticism, epistemic regret (ERM), and learning before failure (LBF)
4. Multi-LLM Agent Collaborative Intelligence (MACI): entropy-guided debate, checks and balances, and its pathologies
5. Persistent transactional memory and planning: SagaLLM, ALAS, TSO, and Mnemosyne
6. Grounded intelligence: the Intelligence Sphere, grounding obligations, authority over interpretation, and risk bounds for physical AI
7. The Quadrivium disciplines (Reason, Elicit, Contract, Bridge), scientific discovery, and the distance to ASI
8. Consciousness, ethics, and wisdom: what the architecture can and cannot deliver

Through lectures, discussions, and a quarter-long team project, students will build and evaluate multi-agent systems on the open-source MACI runtime that advance beyond current frameworks (e.g., LangGraph) to address reasoning and planning problems across domains including medical diagnosis, causal discovery, project planning, supply chain management, and natural disaster response.

Prerequisites: Completion of Machine Learning and Deep Learning courses, along with substantial experience with LLMs.

Grading

1. Assignments: 30%
2. Project Proposal and Midterm Status Report: 20%
3. Final Presentation and Report: 40%
4. Class Participation: 10%

Resources

Schedule

Winter 2027, Mondays and Wednesdays. Chapter references are to the ACM editions of Volumes 1 and 2 and the ACM edition numbering of Volume 3; readings should be completed before the lecture.