Ee-Peng Lim

The dominant zero-shot prompting trick for reasoning has been "Let's think step by step" — open-ended, no enforced structure, hope the model finds its own path. Plan-and-Solve Prompting (2023, with Lim as senior author) added a small but consequential structural change: first ask the model to articulate a plan for solving the problem, then execute that plan step by step. The empirical result was a modest but consistent improvement on math and reasoning benchmarks, and a methodological one — explicit planning leaves an artifact in the trace, so you can examine what the model thought it was about to do and where it deviated.

Worth following when

you want prompting techniques that produce inspectable reasoning artifacts rather than just hoping the answer falls out at the end.

Topics

plan-and-execute prompting structures; the difference between zero-shot CoT and explicit-plan prompting; inspectable reasoning artifacts in LLM evaluation.

Key works

Plan-and-Solve Prompting (2023, senior author); long-arc body of work on data mining and social computing from his SMU group.