Research map / Large language models / Language, safety and interpretability
Use Your INSTINCT: INSTruction optimization for LLMs usIng Neural bandits Coupled with Transformers
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 INSTINCT, which optimizes black-box LLM instructions by running NeuralUCB over soft prompts, using an MLP surrogate on hidden representations from an open white-box LLM (Vicuna-13B).
Problem
Instruction optimization for black-box LLMs is query-inefficient with heuristic local search, and GP-based Bayesian optimization (InstructZero) has limited expressiveness for high-dimensional soft-prompt objectives.
Approach
Replace the GP surrogate in BO with an MLP surrogate (neural bandit NeuralUCB) stacked on the frozen hidden representation of a pre-trained white-box LLM, with pre-computed hidden representations over a discrete Sobol-generated soft-prompt domain and random projection to a low intrinsic dimension.
Key results
- On 20 challenging instruction induction tasks, INSTINCT achieves the best test accuracy on 11 tasks with average rank 2.1, versus 3.2 for InstructZero and 3.8 for APE.
- On zero-shot CoT, INSTINCT finds 'Let's think about it.' scoring 0.74526 on GSM8K vs 0.71797 for 'Let's think step by step.' and 0.81 vs 0.7625 on SVAMP.
- On SAMSum, INSTINCT achieves ROUGE-1 0.35580 vs 0.32595 for InstructZero.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| optimize instructions with INSTINCT rather than InstructZero | ChatGPT instruction optimization · GSM8K | accuracy: 0.74299 -> 0.74526 | Helped |
| optimize instructions with INSTINCT rather than InstructZero | ChatGPT instruction optimization · AQUA-RAT | accuracy: 0.54331 -> 0.54724 | Helped |
| optimize instructions with INSTINCT rather than InstructZero | ChatGPT instruction optimization · SVAMP | accuracy: 0.795 -> 0.81 | Helped |
| use INSTINCT rather than EvoPrompt | ChatGPT instruction optimization · SAMSum | ROUGE-2: 0.11633 -> 0.13350 | Helped |
| use INSTINCT rather than EvoPrompt | ChatGPT instruction optimization · cause_and_effect | accuracy: 0.8267(0.0475) -> 0.5867(0.0871) | Hurt |
| use INSTINCT rather than APE | ChatGPT instruction optimization · second_word_letter | accuracy: 0.7467(0.2028) -> 0.1000(0.0411) | Hurt |
| append one exemplar to the best-discovered instruction at test time (ablation) | INSTINCT · auto_debugging | accuracy: 0.2917(0.0340) -> 0.4583(0.0680) | Helped |
| append one exemplar to every queried instruction (ablation) | INSTINCT · auto_debugging | accuracy: 0.2917(0.0340) -> 0.6250(0.0000) | Helped |
| append one exemplar to every queried instruction (ablation) | INSTINCT · antonyms | accuracy: 0.8467(0.0027) -> 0.8633(0.0072) | Helped |
| append one exemplar to every queried instruction (ablation) | INSTINCT · common_concept | F1: 0.2129(0.0019) -> 0.1812(0.0243) | Hurt |
| append one exemplar to the best-discovered instruction at test time (ablation) | INSTINCT · cause_and_effect | accuracy: 0.5867(0.0871) -> 0.6267(0.0871) | Helped |
| rephrase and improve each generated instruction with ChatGPT (ablation) | INSTINCT · object_counting | accuracy: 0.3400(0.0698) -> 0.5400(0.0072) | Helped |
Surprising findings
- Instructions that score best on the LLM do not always match human perception of the task (e.g., object_counting).
- INSTINCT underperforms some baselines on several tasks, e.g., second_word_letter (0.1000 vs APE 0.7467) and sentence_similarity.
Limitations
- Requires a numeric score on a validation set, which may be unavailable in some applications.
- Only optimizes instructions for maximizing test performance without accounting for harmfulness or other objectives.
- Performance is mixed; INSTINCT is not best on many tasks and underperforms on some (e.g., second_word_letter).
Experimental setup
Datasets: 30 instruction induction datasets (InstructZero) SAMSum GSM8K AQUA-RAT SVAMP
Benchmarks: instruction induction (20-30 tasks) SAMSum GSM8K AQUA-RAT SVAMP
Metrics: test accuracy average rank ROUGE-1 ROUGE-2 ROUGE-L
Baselines: APE InstructZero PromptBreeder EvoPrompt
Builds on: InstructZero NeuralUCB Bayesian optimization APE
Closest papers by meaning
- InstructZero: Efficient Instruction Optimization for Black-Box Large Language Models (ICML 2024 · similarity 0.92)
- PRESTO: Preimage-Informed Instruction Optimization for Prompting Black-Box LLMs (NeurIPS 2025 · similarity 0.90)
- Large Language Models to Enhance Bayesian Optimization (ICLR 2024 · similarity 0.89)
- Better Instruction-Following Through Minimum Bayes Risk (ICLR 2025 · similarity 0.89)
- LESS: Selecting Influential Data for Targeted Instruction Tuning (ICML 2024 · similarity 0.89)
- Do LLMs estimate uncertainty well in instruction-following? (ICLR 2025 · 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.