Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Lottery Ticket Adaptation: Mitigating Destructive Interference in LLMs

Ashwinee Panda, Berivan Isik, Xiangyu Qi, Sanmi Koyejo, Tsachy Weissman, Prateek Mittal

TL;DR

This work addresses destructive interference and catastrophic forgetting in multi-task adaptation of large language models by proposing Lottery Ticket Adaptation (LoTA), a sparse-tuning method that freezes most parameters and updates a small, task-specific subnetwork. LoTA operates in three phases—mask calibration, mask extraction, and sparse adaptation—and can be extended with LoTTO to enforce mutually sparse masks in sequential training. Across diverse tasks (instruction following, reasoning, math, coding, summarization, safety) LoTA matches or surpasses full fine-tuning and LoRA in performance, while enabling robust model merging and significantly compressing task vectors for storage. The results suggest LoTA as a practical and scalable tool for multi-task LLM deployment, particularly in settings requiring sequential updates, safety retention, and cross-task merging.

Abstract

Existing methods for adapting large language models (LLMs) to new tasks are not suited to multi-task adaptation because they modify all the model weights -- causing destructive interference between tasks. The resulting effects, such as catastrophic forgetting of earlier tasks, make it challenging to obtain good performance on multiple tasks at the same time. To mitigate this, we propose Lottery Ticket Adaptation (LoTA), a sparse adaptation method that identifies and optimizes only a sparse subnetwork of the model. We evaluate LoTA on a wide range of challenging tasks such as instruction following, reasoning, math, and summarization. LoTA obtains better performance than full fine-tuning and low-rank adaptation (LoRA), and maintains good performance even after training on other tasks -- thus, avoiding catastrophic forgetting. By extracting and fine-tuning over lottery tickets (or sparse task vectors), LoTA also enables model merging over highly dissimilar tasks. Our code is made publicly available at https://github.com/kiddyboots216/lottery-ticket-adaptation.

Lottery Ticket Adaptation: Mitigating Destructive Interference in LLMs

TL;DR

This work addresses destructive interference and catastrophic forgetting in multi-task adaptation of large language models by proposing Lottery Ticket Adaptation (LoTA), a sparse-tuning method that freezes most parameters and updates a small, task-specific subnetwork. LoTA operates in three phases—mask calibration, mask extraction, and sparse adaptation—and can be extended with LoTTO to enforce mutually sparse masks in sequential training. Across diverse tasks (instruction following, reasoning, math, coding, summarization, safety) LoTA matches or surpasses full fine-tuning and LoRA in performance, while enabling robust model merging and significantly compressing task vectors for storage. The results suggest LoTA as a practical and scalable tool for multi-task LLM deployment, particularly in settings requiring sequential updates, safety retention, and cross-task merging.

Abstract

Existing methods for adapting large language models (LLMs) to new tasks are not suited to multi-task adaptation because they modify all the model weights -- causing destructive interference between tasks. The resulting effects, such as catastrophic forgetting of earlier tasks, make it challenging to obtain good performance on multiple tasks at the same time. To mitigate this, we propose Lottery Ticket Adaptation (LoTA), a sparse adaptation method that identifies and optimizes only a sparse subnetwork of the model. We evaluate LoTA on a wide range of challenging tasks such as instruction following, reasoning, math, and summarization. LoTA obtains better performance than full fine-tuning and low-rank adaptation (LoRA), and maintains good performance even after training on other tasks -- thus, avoiding catastrophic forgetting. By extracting and fine-tuning over lottery tickets (or sparse task vectors), LoTA also enables model merging over highly dissimilar tasks. Our code is made publicly available at https://github.com/kiddyboots216/lottery-ticket-adaptation.
Paper Structure (31 sections, 4 figures, 8 tables, 2 algorithms)

This paper contains 31 sections, 4 figures, 8 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Background: Multi-Task Adaptation
  3. Storing and Loading Adapters.
  4. Sequential Training.
  5. Model Merging.
  6. Lottery Ticket Adaptation (LoTA)
  7. (1) Storing & Loading Adapters.
  8. (2) Sequential Training.
  9. Further Enhancing Sequential Training via Lottery Ticket Together Optimization (LoTTO).
  10. (3) Model Merging.
  11. Experimental Setup
  12. Experimental Results
  13. LoTA Design Choices
  14. Adapting to a Single Task
  15. Storing LoTA Adapters Efficiently.
  16. ...and 16 more sections

Figures (4)

  • Figure 1: Multi-task adaptation: storing and loading adapters, sequential training, model merging.
  • Figure 2: Lottery Ticket Adaptation (LoTA): (1) Mask calibration via FFT for $\mathcal{T}$ iterations, (2) Extracting the sparsity mask $m$ from the task vector $\Delta$, (3) Sparse fine-tuning with sparsity mask $m$ for $T$ iterations.
  • Figure : Lottery Ticket Adaptation (LoTA)
  • Figure : Lottery Ticket Together Optimization (LoTTO)