Axes of Planning in LLMs + Partial Lit Review

The question of whether large language models 'plan' is central to interpretability and AI safety, but the term itself is slippery. In a new post on LessWrong, researcher NickyP tackles this by first listing 12 concrete behaviors often associated with planning—from a model noticing future-useful facts mid-generation, to executing conditional branching based on an internal outline. These examples range from simple next-token adjustments to full hierarchical planning, highlighting that 'planning' covers a spectrum of cognitive operations rather than a single binary capability.

To formalize this, NickyP proposes four axes: Time Horizon (next token vs. whole output), Vague vs. Specific (vague theme vs. detailed subgoals), Option Space (multiple explicit alternatives vs. single narrow path), and Forward Dependency/Constraints (shallow one-off choices vs. deeply nested dependencies). The author also gives a partial literature review, noting that most existing work focuses on one or two axes at most. This framework doesn't provide definitive answers but equips researchers with a common vocabulary to design more targeted experiments—critical for determining when models are genuinely reasoning ahead versus just pattern-matching fluently.