Research map / Large language models / Language, safety and interpretability
PRESTO: Preimage-Informed Instruction Optimization for Prompting Black-Box LLMs
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PRESTO, which exploits the many-to-one mapping from soft prompts to instructions in a white-box LLM (preimages) via score sharing, preimage-based initialization, and score consistency regularization to optimize instructions for black-box LLMs more efficiently.
Problem
Optimizing instructions for black-box LLMs using soft prompts fed to a white-box LLM is query-inefficient because many distinct soft prompts map to the same instruction, causing redundant queries.
Approach
Treats the many-to-one soft-prompt-to-instruction mapping (preimages) as prior knowledge: scores of one evaluated instruction are shared with all soft prompts in its preimage, initialization covers the embedding space using preimage sizes and MMD, and an unsupervised loss enforces equal predictions within unscored preimages.
Key results
- PRESTO achieves the best average rank (1.97) and best accuracy on 18 of 30 instruction induction tasks, versus 8 for the second-best ZOPO.
- Score sharing yields about 2,300 scored soft prompts on average, 14x more than baselines' 165 under the same 165-query budget.
- Ablation: vanilla average accuracy 51.91 rises to 62.91 with all three components on 20 tasks.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use PRESTO instead of ZOPO | instruction optimization · 30 instruction induction tasks | average rank: 2.90 -> 1.97 | Helped |
| use PRESTO instead of the strongest listed baseline | CoT prompting · AQUA-RAT | accuracy: 0.7520 -> 0.7756 | Helped |
| use PRESTO instead of the strongest listed baseline | CoT prompting · GSM8K | accuracy: 0.9143 -> 0.9128 | Hurt |
| add score sharing (ablation) | vanilla instruction optimizer · 20 instruction induction tasks | average accuracy: 51.91 -> 59.57 (+7.66) | Helped |
| add score consistency regularization to score sharing (ablation) | instruction optimizer with score sharing · 20 instruction induction tasks | average accuracy: 59.57 -> 61.77 | Helped |
| add preimage-based initialization to score sharing (ablation) | instruction optimizer with score sharing · 20 instruction induction tasks | average accuracy: 59.57 -> 61.82 | Helped |
| increase preimage size from 0% to 100% (ablation) | PRESTO · 20 instruction induction tasks | average accuracy: 51.91 -> 62.91 | Helped |
Surprising findings
- Mistral-7B-Instruct-v0.3 produced roughly 50% duplicate instructions from 10,000 soft prompts, showing preimage structure varies across white-box LLMs.
Limitations
- Mild computational overhead from preimage construction and MMD computation compared with simpler baselines.
- Benefits are more pronounced when the candidate set is large, since score sharing is most effective when many soft prompts share an instruction.
Experimental setup
Datasets: instruction induction (30 tasks) GSM8K AQUA-RAT SVAMP
Benchmarks: instruction induction (30 tasks) GSM8K AQUA-RAT SVAMP
Metrics: accuracy average rank number of best-performing tasks
Baselines: APE InstructZero INSTINCT EvoPrompt ZOPO OPRO
Builds on: INSTINCT InstructZero NeuralUCB ZOPO
Closest papers by meaning
- InstructZero: Efficient Instruction Optimization for Black-Box Large Language Models (ICML 2024 · similarity 0.91)
- Use Your INSTINCT: INSTruction optimization for LLMs usIng Neural bandits Coupled with Transformers (ICML 2024 · similarity 0.90)
- Investigating the Effectiveness of Task-Agnostic Prefix Prompt for Instruction Following (AAAI 2024 · similarity 0.89)
- PandaLM: An Automatic Evaluation Benchmark for LLM Instruction Tuning Optimization (ICLR 2024 · similarity 0.89)
- Long Is More for Alignment: A Simple but Tough-to-Beat Baseline for Instruction Fine-Tuning (ICML 2024 · similarity 0.88)
- ShiftAddLLM: Accelerating Pretrained LLMs via Post-Training Multiplication-Less Reparameterization (NeurIPS 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.