大语言模型微调方法研究综述

Summary

This survey, published in 中文信息学报 (2025), organizes research on adapting large language models (LLMs) to downstream tasks and user instructions under tight hardware budgets. The authors catalogue fine-tuning techniques from 2019 to 2024, from the appearance of Adapter onward. The abstract reports more than 50 methods, and the introduction gives a count of 59. They sort these methods by how each treats model parameters: all weights updated, a subset updated, new weights added, or no weights changed. For each method they describe the underlying idea, where in the Transformer it acts, and its strengths and weaknesses. They then estimate trainable parameter counts, memory footprint and compute cost for representative methods. The paper also reviews common base models (GPT, LLaMA and Chinese LLMs) and evaluation suites, mainly GLUE and SuperGLUE. It closes with practical advice on choosing a fine-tuning strategy. For practitioners, it offers a single, parameter-centred map of PEFT techniques along with resource trade-offs that can guide fine-tuning on modest hardware.

Key points

• Four-category taxonomy:
  • Full-parameter tuning.
  • Partial-parameter tuning, split into heuristic and self-learned selection (e.g., BitFit).
  • Additional-parameter tuning, which covers Adapter variants, the LoRA family (LoRA, QLoRA and derivatives) and soft prompts (Prefix tuning, Prompt tuning, P-tuning) applied at the Embedding layer.
  • Parameter-free approaches, such as prompt engineering and in-context learning (ICL).
• Publication trends: New methods grew from 2021 and became especially dense from late 2022. In 2023, methods that add no new parameters made up about 30% of that year's entries; examples are LOMO, HiFi, PaFi, DiffFit and RAIN. GaLore (2024) targets training-time memory.
• Most-cited methods: Prefix tuning, Prompt tuning, LoRA and QLoRA. BitFit and Adapter are frequently cited as baselines.
• Benchmarks covered: GLUE (CoLA, SST-2, MRPC, QQP, STS-B, MNLI, QNLI, RTE, WNLI) and SuperGLUE (BoolQ, CB, MultiRC, ReCoRD, WiC, WSC, COPA). The survey also discusses SQuAD, XSum, WMT and C4.
• Motivation and resource advice: The authors cite roughly US$4.6M for a single GPT-3 175B training run to show why full training is out of reach for small teams. They recommend that method comparisons start on BERT/RoBERTa-base before moving to larger models.