Blog

Writing on my research, LLMs and planning, and life in research

Categories

My Research 4 posts
Latest: You Don't Need to Rank All States. Just Rank the Actions ...
Learning for Planning 6 posts
Latest: Epilogue: Toward Uncertainty
Planning under Uncertainty 6 posts
Latest: Epilogue: Open Questions
LLMs Automated Planning and Agents 8 posts
Latest: Epilogue: Where the Field Goes Next
Research and Life in the Age of AI 1 post
Latest: How to Use AI Without Losing Your Thinking Power

Topics

All Posts

  • How to Use AI Without Losing Your Thinking Power

    I caught myself outsourcing my ability to reason — not just to code, but to think through everyday life. Here's the framework I built to take it back.

  • Epilogue: Toward Uncertainty

    What GABAR doesn't do, why that matters, and where the natural next chapter takes the same warehouse. Closer of the Learning for Planning series.

  • Epilogue: Open Questions

    What HOO-POMDP and GammaZero don't yet do, where the next decade of partially-observable planning is heading, and how this all connects to the rest of AI. Closer of the Planning Under Uncertainty series.

  • You Don't Need to Rank All States. Just Rank the Actions in Front of You.

    How a simple shift in learning objective — from global value functions to local action ranking — yields planning policies that generalize 8x beyond training size. A deep dive into GABAR. Part 4 of the Learning for Planning series.

  • Your Belief About the World Is a Graph. Now Your Planner Can Use It.

    How encoding uncertainty as graph structure enables POMDP planners to generalize far beyond their training size. A deep dive into GammaZero. Part 4 of the Planning Under Uncertainty series.

  • Your Planning State Is a Graph

    Why fixed-size vectors can't represent a planning problem, and what to do instead. Part 3 of the Learning for Planning series.

  • Teaching Robots to Tidy Up: Planning Under Uncertainty in Multi-Room Environments

    How hierarchical planning with object-oriented beliefs enables robots to rearrange objects when they can't see everything at once. Part 3 of the Planning Under Uncertainty series.

  • What to Learn — A Survey of Learning Objectives

    Three families of methods, each picking a different target for the model to predict. The choice of target ends up mattering more than the choice of architecture. Part 2 of the Learning for Planning series.

  • MCTS for POMDPs

    Exact POMDPs are intractable. Online tree search makes them solvable in practice. The price: every modern POMDP solver depends on heuristics — and that's exactly where learning fits in. Part 2 of the Planning Under Uncertainty series.

  • The Scaling Problem

    Planning is sound, complete, and optimal. It is also exponentially hard. Learning is how we get past that wall — but to see why it works, we first need to feel the wall. Part 1 of the Learning for Planning series.

  • Planning When You Can't See the Whole World

    Classical planning assumes you know everything. Real agents almost never do. Belief states give us a principled way to plan under uncertainty — and an extraordinary computational cost in the bargain. Part 1 of the Planning Under Uncertainty series.

  • Learning for Planning — Series Overview

    A four-part series (plus epilogue) on how graph neural networks let learned policies generalize from small, solvable training instances to problems 8x larger — without ever retraining.

  • Planning Under Uncertainty — Series Overview

    A four-part series (plus epilogue) on planning when the agent can't see the full state — POMDPs, belief tracking, online tree search, and the two strategies (abstraction and learning) that make large partially-observable problems tractable.

  • Epilogue: Where the Field Goes Next

    A forward-looking companion to the seven-post series. Limitations of HALO, the meta-learning frame, open challenges, and the research directions that will define the next few years.

  • Training the Conductor

    Stop prompting the orchestrator at every step. Let the verifier you already trust supervise a small local model — and watch the bill drop by 99%. Part 7 of the Planning in the Era of LLMs series.

  • From English to Plans: The NL-to-PDDL Frontier

    NL2Plan, agentic PDDL generation, and the orchestrator bottleneck — when the conductor can't keep up with the orchestra. Part 6 of the Planning in the Era of LLMs series.

  • The Modern Playbook: LLMs That Help Planners

    LLM-Modulo, heuristic generation, and the generate-verify loop that turned 12% into 82%. Part 5 of the Planning in the Era of LLMs series.

  • LLMs Try to Plan (It Goes Badly)

    PlanBench, Mystery Blocksworld, and the sobering evidence that frontier models can't reliably sequence three actions. Part 4 of the Planning in the Era of LLMs series.

  • 50 Years of Planning Algorithms (In 15 Minutes)

    From STRIPS to Fast Downward — a tour of the solvers your LLM agent should be working with, not replacing. Part 3 of the Planning in the Era of LLMs series.

  • The Formal Planning Primer

    States, actions, goals, and the language that makes them precise — everything you need to follow the rest of this series. Part 2 of the Planning in the Era of LLMs series.

  • Making LLM Agents Actually Plan: A Roadmap

    Why your LLM agent fails at multi-step tasks, and a 7-part guide to fixing it. Part 1 of the Planning in the Era of LLMs series.