Build up to the hard part
Least to most prompting splits a hard problem into a chain of easier subproblems ordered from simplest to most complex. Each subproblem is solved using the answers from the earlier ones, so the model climbs a staircase instead of leaping.
Two phases
- Decompose the original question into an ordered list of smaller subquestions.
- Solve sequentially, feeding each solved subanswer into the prompt for the next step until the final question is reached.
Why ordering matters
A later step often needs the result of an earlier one. By solving the least complex piece first and carrying its answer forward, the model always has the pieces it needs when it reaches the harder steps.
How it differs from chain of thought
Plain chain of thought reasons within one pass. Least to most explicitly names subproblems and reuses their outputs as inputs, which helps on tasks that need compositional generalization, such as following nested instructions or longer symbolic chains.
Key idea
Least to most prompting decomposes a problem into ordered subquestions and solves them in sequence, feeding each answer forward, which helps the model compose solutions to harder problems than a single reasoning pass handles.