How Much Prompt Is Enough? A Blackbox Minimization of Few-Shots in LLMs | AI Paper Digest

TL;DR Highlight

Delta debugging trimmed few-shot prompts by 65.3% on average with unchanged propositional output

Who Should Read

Backend developers whose long few-shot prompts raise token costs and break when swapping LLMs

Core Mechanics

65% of exemplar characters were unnecessary
- Across 6 models and 2 temperatures (12 runs), a mean of 65.3% (±15.8%) was removed with output propositions unchanged.
- Removal ranged from 38.4% to 87.3% depending on the model.
The skeleton stays, the prose goes
- Identifiers, constraint declarations and logical expressions were mostly kept.
- Natural-language prose and cross-prompt relationship annotations were mostly deleted.
Bigger models tolerate more trimming
- Each extra 10B parameters added 4.3 percentage points of removal.
- Model family explained much more variance than size alone.
Over-compressed prompts break on other models
- The shortest prompt (about 29 tokens) had low mean transfer accuracy.
- Prompts that kept more structural scaffolding transferred better.
Pick the minimizer and the consumer separately

Evidence

Mean removal of 65.3%
- Averaged over 12 model/temperature runs against a 1749-character original exemplar.
- Min 38.4% (Mistral-Small-24B, T=0.0), max 87.3% (Llama-3.3-70B, T=0.0).
Accuracy after minimization
- At T=0.0, 4 of 6 models matched the full-prompt baseline 100% across 10 runs.
- At T=0.5, 3 models (gemma-2-9b-it, Mistral-Small-24B, Qwen-2.5-72b) hit 100%.
Cross-model transfer
- At T=0.0, 22 of 30 off-diagonal pairs gave valid output with 78.1% mean accuracy.
- At T=0.5, 28 pairs were valid with 75.1% mean accuracy.
Regression analysis
- A size-only model had R²=0.544, while a family model had R²=0.912.
- gemma-2-9b-it compressed 16.5 points more than Llama-3.1-8B (p=0.042).

How to Apply

Delete relationship annotations first
- Remove the blocks in your few-shot examples that explain how examples relate to each other.
- Compare outputs on a handful of existing test inputs.
Build a ddmin-style halving loop
- Split the examples into lines and remove subsets.
- Keep a removal if output matches the baseline and revert it otherwise.
Don't over-trim if you swap models
- If you have fallback models, stop at a mid-length prompt that keeps identifiers and structure.
- Run the trimmed prompt on every target model to confirm.
Rerun to filter out flaky reductions
- Run the same input several times after trimming.
- Reject any reduction that gives a different result even once.

Try It

copy & run

# Paste into Claude or ChatGPT (simplified version of the paper's method: manual prompt minimization)

Below is my few-shot prompt and test inputs.
[Prompt]
{{current few-shot prompt}}

[Test inputs]
{{input 1}}
{{input 2}}

Instructions:
1. First, produce outputs for each input with the original prompt as the baseline.
2. Split the prompt into units (role, task description, example body, cross-example relation notes, closing constraints).
3. Create versions with one unit removed and check whether the output matches the baseline.
4. Keep the removal if it matches, otherwise revert.
5. Show the final minimal prompt, the list of removed parts, and the percentage of characters saved.

Next: try the minimal prompt on another model and check the output matches.

Terminology

Delta debugging (ddmin)A debugging technique that repeatedly cuts parts of a failing input until only the minimum remains. It's like removing items from a bag one by one to see which you truly need.

Propositional output fidelityHow closely the set of logical constraints in an output matches the baseline. Different wording still passes if the constraints are the same.

Universal encoderA model whose minimized prompts are readable by other models. Like a person who summarizes in a way anyone can follow.

Universal decoderA model that can follow prompts minimized by other models. Like a reader who understands any shorthand note.

P2P (Prompt2Propositions)A method that turns prompts into sets of propositional constraints like C1 ∧ C2 so two outputs can be checked for equivalence.

Related Papers

Original Abstract (Expand)

Prompts are the primary mechanism for directing the behavior of large language models (LLMs). Yet the internal structure and causal hierarchy of prompts remain poorly understood: which parts are causally necessary and which are redundant is an open question. This opacity can have severe consequences. Subtle prompt variations can silently shift model outputs in critical software systems, and engineers lack techniques to reason about prompt reliability. We present \framework, a blackbox prompt-minimization framework that reduces few-shot prompts to their necessary minimal subset. We use a case study to apply \framework to a few-shot learning system and demonstrate the insights that this framework can provide. Our experiments show that few-shot exemplars can be reduced by a mean of 65.3\%~$\pm$~15.8\% in character count while fully preserving propositional output fidelity. The models preferentially retain logical identifiers and constraint declarations while discarding natural language prose and cross-prompt relational annotations. Our analysis also shows that some models are universal encoders, able to produce highly legible yet minimized prompts, while others are universal decoders, able to interpret minimized prompts from most other models. By identifying which components are indispensable, \framework provides a principled basis for prompt compression and structural analysis of few-shot exemplars.