Instruction Tuning for Large Language Models: A Survey
Shengyu Zhang, Linfeng Dong, Xiaoya Li, Sen Zhang, Xiaofei Sun, Shuhe Wang et al. · 2025 · 179 citationsRead the paper
This article surveys research works in the quickly advancing field of instruction tuning (IT), a crucial technique to enhance the capabilities and controllability of large language models (LLMs). Instruction tuning refers to the process of further training LLMs on a dataset consisting of (instruction, output) pairs in a supervised fashion, which bridges the gap between the next-word prediction objective of LLMs and the users’ objective of having LLMs adhere to human instructions. In this work, we make a systematic review of the literature, including the general methodology of IT, the construction of IT datasets, the training of IT models, and applications to different modalities, domains and application, along with analysis of aspects that influence the outcome of IT (e.g., generation of instruction outputs, size of the instruction dataset). We also review the potential pitfalls of IT along with criticism against it, along with efforts pointing out current deficiencies of existing strategies and suggest some avenues for fruitful research.
What the full text shows
Deployed in practice- InstructGPT outperforms GPT-3 in truthfulness and toxicity reduction. — 10% higher truthfulness (TruthfulQA) and 7% lower toxicity (RealToxicityPrompts) (Automatic evaluation)
- BLOOMZ shows zero-shot improvement over BLOOM across multiple NLP tasks. — +10.4% coreference resolution, +20.5% sentence completion, +9.8% natural language inference (Zero-shot setting)
- LIMA demonstrates the superficial alignment hypothesis, outperforming larger models with less data. — 17% higher human preference than InstructGPT and 19% higher than Alpaca (Human evaluation)
- ChatGLM2 outperforms GLM and ChatGLM across multiple benchmarks. — +3.1 MMLU, +5.0 C-Eval, +8.6 GSM8K, +2.2 BBH vs GLM (Bilingual English/Chinese benchmarks)
- TÜLU reaches a significant percentage of proprietary model performance. — 83% of ChatGPT and 68% of GPT-4 performance on average (Mixed instruction dataset fine-tuning)
- BLOOMZ improves coding performance over BLOOM. — 10% higher Pass@100 on HumanEval (HumanEval benchmark)
- Setup
- Review of multiple LLMs (ranging from 6B to 176B parameters) and datasets (e.g., Super-Natural Instructions with 5M instances, DeepSeekMath with 120B tokens).
- How
- The paper surveys instruction tuning (IT) methodologies, including supervised fine-tuning (SFT), reward model training, and Proximal Policy Optimization (PPO). It reviews various dataset construction techniques and evaluation frameworks like LLM-as-a-judge and HELM.
- Limitations
- Length bias in evaluations (e.g., AlpacaEval favoring longer responses) · Benchmark contamination in large-scale datasets · Limited skill coverage in some benchmarks like MT-Bench
Summarized by Gemma 4 from the open-access full text · 13 Oct 2026. Read it