Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

RepCali: High Efficient Fine-tuning Via Representation Calibration in Latent Space for Pre-trained Language Models

Fujun Zhang, Xiaoying Fan, XiangDong Su, Guanglai Gao

TL;DR

This work addresses the misalignment between encoder representations and decoder expectations in fine-tuned encoder–decoder PLMs. It proposes RepCali, a lightweight representation calibration block that augments encoder outputs in the latent space via a shape seed and a learnable embedding, producing calibrated inputs for the decoder with a small scaling factor. Empirical results across 25 PLM-based models and 8 tasks (including English and Chinese data) show RepCali consistently improves performance over standard fine-tuning baselines while adding only a fraction of parameters. The method is universal, plug-and-play, and demonstrated to benefit both NLG and NLU tasks, including large language models, with evidence from latent-space visualizations of smoother representations. Overall, RepCali offers a practical and effective approach to narrow cross-domain encoder–decoder gaps during fine-tuning, with broad applicability and minimal overhead.

Abstract

Fine-tuning pre-trained language models (PLMs) has become a dominant paradigm in applying PLMs to downstream tasks. However, with limited fine-tuning, PLMs still struggle with the discrepancies between the representation obtained from the PLMs' encoder and the optimal input to the PLMs' decoder. This paper tackles this challenge by learning to calibrate the representation of PLMs in the latent space. In the proposed representation calibration method (RepCali), we integrate a specific calibration block to the latent space after the encoder and use the calibrated output as the decoder input. The merits of the proposed RepCali include its universality to all PLMs with encoder-decoder architectures, its plug-and-play nature, and ease of implementation. Extensive experiments on 25 PLM-based models across 8 tasks (including both English and Chinese datasets) demonstrate that the proposed RepCali offers desirable enhancements to PLMs (including LLMs) and significantly improves the performance of downstream tasks. Comparison experiments across 4 benchmark tasks indicate that RepCali is superior to the representative fine-tuning baselines.

RepCali: High Efficient Fine-tuning Via Representation Calibration in Latent Space for Pre-trained Language Models

TL;DR

This work addresses the misalignment between encoder representations and decoder expectations in fine-tuned encoder–decoder PLMs. It proposes RepCali, a lightweight representation calibration block that augments encoder outputs in the latent space via a shape seed and a learnable embedding, producing calibrated inputs for the decoder with a small scaling factor. Empirical results across 25 PLM-based models and 8 tasks (including English and Chinese data) show RepCali consistently improves performance over standard fine-tuning baselines while adding only a fraction of parameters. The method is universal, plug-and-play, and demonstrated to benefit both NLG and NLU tasks, including large language models, with evidence from latent-space visualizations of smoother representations. Overall, RepCali offers a practical and effective approach to narrow cross-domain encoder–decoder gaps during fine-tuning, with broad applicability and minimal overhead.

Abstract

Fine-tuning pre-trained language models (PLMs) has become a dominant paradigm in applying PLMs to downstream tasks. However, with limited fine-tuning, PLMs still struggle with the discrepancies between the representation obtained from the PLMs' encoder and the optimal input to the PLMs' decoder. This paper tackles this challenge by learning to calibrate the representation of PLMs in the latent space. In the proposed representation calibration method (RepCali), we integrate a specific calibration block to the latent space after the encoder and use the calibrated output as the decoder input. The merits of the proposed RepCali include its universality to all PLMs with encoder-decoder architectures, its plug-and-play nature, and ease of implementation. Extensive experiments on 25 PLM-based models across 8 tasks (including both English and Chinese datasets) demonstrate that the proposed RepCali offers desirable enhancements to PLMs (including LLMs) and significantly improves the performance of downstream tasks. Comparison experiments across 4 benchmark tasks indicate that RepCali is superior to the representative fine-tuning baselines.
Paper Structure (18 sections, 2 equations, 4 figures, 11 tables)

This paper contains 18 sections, 2 equations, 4 figures, 11 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Fine-tuning Method
  4. Latent Space
  5. Methodology
  6. Experiments on Fine-tuning Methods
  7. Tasks and Datasets
  8. Method Performance Comparison
  9. Comparison of Additional Parameters
  10. Experiments on Downstream Tasks
  11. Downstream Tasks and Datasets
  12. Analysis of Experimental Results
  13. Model Sizes
  14. Visual Analysis in the Latent Space
  15. Conclusion
  16. ...and 3 more sections

Figures (4)

  • Figure 1: Visualization of the distribution of the output of T5 encoder before and after fine-tuning via the T-SNE algorithm on the Abductive Commonsense Reasoning ($\alpha$NLG) task.
  • Figure 2: The detailed architecture of various tuning methods.
  • Figure 3: Overview of our representation calibration method.
  • Figure 4: We chose three different PLMs, BART, T5, and PEGASUS for visualization and analysis of the latent space. The blue points are the hidden representations obtained after fine-tuning using our calibration method RepCali, and the yellow points are the hidden representations obtained after fine-tuning for the PLMs.