Meta's new prompting technique makes LLMs significantly better at code review — boosting accuracy to 93% in some cases

Meta researchers introduce semi-formal reasoning: a structured prompting method requiring the model to state premises, trace concrete execution paths, and derive formal conclusions before answering.

Semi-formal reasoning improves LLM performance on coding tasks and reduces errors in fault localization and codebase Q&A.

It leverages agentic code reasoning — agents that navigate files, trace dependencies, and gather context to perform deep semantic analysis without executing code.

The technique enables execution-free semantic code analysis for developers, which reduces AI coding infrastructure costs.

Two approaches to execution-free verification are described: unstructured LLM evaluators (or reward models) and formal verification using languages like Lean, Coq, or Datalog.