YNU at SemEval-2025 Task 4: Synthetic Token Alternative Training for LLM Unlearning

Yang Chen, Zheyang Luo, Zhiwen TangACL 2025

TL;DR

This paper introduces Synthetic Token Alternative Training (STAT), a novel method for efficiently unlearning specific data from large language models (LLMs) without degrading performance on retained data. STAT achieves this by generating synthetic "forget" tokens and alternating training between these synthetic tokens and original "retain" data, leading to improved stability and faster convergence.

What Problem Does This Paper Address?

In today's data-driven world, the ability to selectively remove information from trained models is becoming increasingly crucial. This is driven by privacy concerns (e.g., the "right to be forgotten") and the need to correct or remove undesirable behaviors or biases learned by LLMs. Traditional methods for unlearning, such as retraining from scratch, are computationally prohibitive for large models. Fine-tuning approaches can be unstable and may not effectively preserve performance on the data the model should remember. The challenge lies in efficiently and reliably "unlearning" specific data while ensuring the model's overall utility and performance on other data remain intact.

Methods

The core of the proposed method, Synthetic Token Alternative Training (STAT), is a two-stage process:

  1. Synthetic Token Generation: Instead of directly using the original "forget" data, STAT generates synthetic tokens. This is done by sampling tokens from a uniform distribution across the entire vocabulary and replacing the original target tokens in the forget set. The intuition here is that by presenting the model with "random" or "unfamiliar" tokens for the data it needs to forget, it disrupts the learned associations without causing catastrophic forgetting due to direct gradient ascent on potentially sensitive information.
  2. Alternative Training Mechanism: This is the key to balancing unlearning and retention. Instead of processing all forget data and then all retain data sequentially, STAT alternates between training on the synthetic forget tokens and the original retain data at a sample level. This dynamic switching is hypothesized to prevent the model from over-optimizing on either objective, leading to more stable training, faster convergence, and better generalization. The alternation ratio is a hyperparameter that balances the influence of forget and retain data.

Datasets

The paper describes experiments conducted on a benchmark dataset for LLM unlearning, which includes:

  • Long-form synthetic creative documents.
  • Short-form synthetic biographies containing personally identifiable information (PII).
  • Real documents sampled from the target model's training dataset.

The models used for evaluation were OLMo-7B and OLMo-1B, which were fine-tuned to memorize this benchmark dataset.

Key Insights and Ideas

  • Synthetic tokens as a "gentle" unlearning tool: Replacing original forget tokens with synthetically generated ones seems to be a more stable approach than directly optimizing on the original forget data, which can lead to instability.
  • Alternating training is crucial for balance: The dynamic switching between synthetic forget data and real retain data is key to preventing catastrophic forgetting and maintaining overall model performance. This implicit regularization helps the model find flatter loss surfaces.
  • Uniform sampling is a strong baseline: Despite its simplicity, uniform random sampling for synthetic token generation proved to be a robust strategy, forming a reliable foundation for the unlearning process.
  • STAT accelerates convergence: The alternating training protocol significantly speeds up the unlearning process compared to traditional sequential methods.

Key Results

STAT achieved impressive results in the SemEval-2025 Task 4:

  • 3rd place in the OLMo-7B model track.
  • 7th place in the OLMo-1B model track.

The paper highlights that STAT significantly outperforms official baselines and demonstrates superior stability in knowledge retention and more effective targeted forgetting compared to existing approaches. For instance, on OLMo-7B, STAT achieved a final score of 0.470, significantly higher than many baselines like Gradient Ascent (0.394) and NPO (0.188). In the task aggregate metric, STAT also showed strong performance, ranking at the forefront among participants.

Limitations and Future Work

  • Metric limitations: The authors note that metrics like ROUGE-L can be misleading in unlearning evaluations, as they might favor syntactically similar but semantically unlearned sequences, potentially inflating scores.
  • Synthetic data generation: While uniform sampling is effective, exploring more sophisticated synthetic data generation methods (like prompted generation) could potentially yield further improvements, though it might introduce its own complexities.
  • Hyperparameter tuning: The effectiveness of the alternation ratio and other training parameters would likely require careful tuning for different models and datasets.

The paper implicitly suggests that further research into robust evaluation metrics for unlearning and exploring more advanced synthetic data generation techniques could be fruitful future directions.

Summary updated

BibTeX
@inproceedings{chen2025ynu,
  author = {Yang Chen and Zheyang Luo and Zhiwen Tang},
  title = {{YNU at SemEval-2025 Task 4: Synthetic Token Alternative Training for LLM Unlearning}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2025},
}
Figure