Explore use of few-shot prompting

[SebastianZimmeck]

It seems likely that we will combine Constrained Decoding (to mechanically force the model to always output correct syntactical structure) and Few-Shot Prompting (to teach the model the semantic meaning of our task). Thus, @mkarim-usc, can you briefly write up here with references to academic papers what the current best practices are for Few-Shot Prompting in academic research?

One particular point to consider, how many scenarios (i.e., shots) are commonly used?

From a practical perspective, what do we need to do concretely? It would be even better if you try it out, for example, by replicating what a paper describes. For the time being, what you try does not need to be related to what we are doing (though, of course, if it is, that would be even better).

[mkarim-usc]

I have checked several publications related to constrained decoding and few-shot prompting. Here is a short summary:

Few-shot prompting means giving a language model a few completed examples in the prompt. The model uses these examples to understand the task without being retrained (Brown et al., 2020). Constrained decoding is different. It forces the model to follow a required output format, such as JSON or a grammar, by blocking invalid choices while the answer is being generated (Scholak et al., 2021; Willard & Louf, 2023). The two methods solve different problems. Few-shot prompting helps the model understand what the task means, while constrained decoding makes sure the output has the correct structure. Prompting alone cannot fully guarantee valid output every time (vLLM Blog, 2025). Recent research often combines both methods. One clear example is Grammar Prompting by Wang et al. (2023), which uses few-shot examples to teach the task and a formal grammar to control the final output.

For the number of examples, or “shots,” many studies show a similar pattern. The biggest improvement often comes from adding the first example. Performance usually continues to improve up to about 3–5 examples, but the gains often become smaller after that (Meta-Tool paper, 2026; DetPO paper, 2026; Umapathi et al., 2023). Using hundreds or thousands of examples, called many-shot prompting, can help on harder tasks. However, this requires a long context window and a large set of high-quality examples (Agarwal et al., 2024). The choice of examples also matters. Examples that are similar to the current input usually work better than randomly selected examples (Liu et al., 2022). The way the output format is enforced is also important. Constrained decoding provides a structural guarantee that prompting alone cannot provide (DCCD paper, 2026).

More about the present works are in: https://github.com/xai-privacy/analysis-framework/blob/main/docs/few_shot_prompting_and_constrained_decoding.md

I am also working on examples of few shot prompting and constrained decoding.