#ICL

The post reframes how LLMs actually contribute to hard problems: not as direct reasoners but as generators of throwaway automation scripts. Researchers and engineers doing combinatorial or exploratory work can apply this by delegating tedious search scripting to models. The observation about in-context learning through oracles offers a practical pattern for anyone building evaluation or verification loops.

Teortaxes@teortaxesTexSource

The framework explains why few-shot prompting has mixed effects on RL-tuned models and provides insights for modern reasoning pipelines like DeepSeek-R1 and o1-style models.

The findings provide cost-effective guidelines for configuring text-to-SQL pipelines, applicable across multiple backbones without per-paradigm search.