World-Model-Augmented Planning

Planning needs predicted action outcomes and backtracking, so bare chain-of-thought action lists often fail without state or search support.

Whether autoregressive models can truly plan remains contested. Critics argue that fluent-looking plans are often extracted general knowledge rather than executable search over outcomes. Planning is fundamentally a search problem: explore paths to a goal, predict rewards, pick a path — sometimes none exists — and backtrack when a branch fails.

A bare action sequence from chain-of-thought does not supply each action's outcome state. Knowing that turning right leads off a cliff is what lets a planner rule that branch out. Approaches that treat language-model reasoning as planning with a world model use the model's knowledge to predict outcomes and guide coherent plans; alternatively, the model can be only part of a planner augmented with search and explicit state tracking. Tree of Thoughts is one concrete harness: thoughts as states, an LM search heuristic for prune/expand, and BFS or DFS with lookahead and backtracking.

For harness design, do not assume "think step by step" equals a reliable plan. Prefer plan-validate-execute, explicit state, reflection, and tools that reveal environment feedback after each action.

---

Sources

• AI Engineering (Chip Huyen), ch. 6
• Tree of Thoughts (Yao et al.), pp. 1–14