Perspectives on AI planning, world models, and language model integration
Deep dives and positions on reliable agent architectures
When an AI agent takes an unsafe action, calling it deceptive or misaligned obscures the real problem. Unsafe behavior is an engineering failure — not a character flaw. Safety must be built into the environment, not expected from the agent.
ACPBench evaluates LLMs on eight core reasoning tasks underlying reliable planning — from action applicability and state transitions to plan validation — spanning boolean, multiple-choice, and generative question formats.
Why reliable and efficient LLM-based planning should focus on generating reusable planners and programmatic world models rather than solving every instance from scratch.
A practical guide for evaluating LLM-based planners — applying six decades of planning methodology to avoid pitfalls in task semantics, benchmark contamination, mismatched baselines, and incomplete reporting.
How AutoToS uses LLMs to automatically generate sound and complete search components — successor functions, goal tests, and heuristics — eliminating the need for human experts while retaining the formal guarantees of classical search.
How Hierarchical Component Learning for Generalized Policies (HCL-GP) combines generalized planning, hierarchical decomposition, and execution-grounded skill learning to achieve 97.8% Challenge SGC on AppWorld.