English

Understanding Catastrophic Forgetting in Language Models via Implicit Inference

Computation and Language 2024-04-16 v2 Machine Learning

Abstract

We lack a systematic understanding of the effects of fine-tuning (via methods such as instruction-tuning or reinforcement learning from human feedback), particularly on tasks outside the narrow fine-tuning distribution. In a simplified scenario, we demonstrate that improving performance on tasks within the fine-tuning data distribution comes at the expense of capabilities on other tasks. We hypothesize that language models implicitly infer the task of the prompt and that fine-tuning skews this inference towards tasks in the fine-tuning distribution. To test this, we propose Conjugate Prompting, which artificially makes the task look farther from the fine-tuning distribution while requiring the same capability, and we find that this recovers some of the pretraining capabilities in our synthetic setup. Since real-world fine-tuning distributions are predominantly English, we apply conjugate prompting to recover pretrained capabilities in LLMs by simply translating the prompts to different languages. This allows us to recover in-context learning abilities lost via instruction tuning, natural reasoning capability lost during code fine-tuning, and, more concerningly, harmful content generation suppressed by safety fine-tuning in chatbots like ChatGPT.

Keywords

Cite

@article{arxiv.2309.10105,
  title  = {Understanding Catastrophic Forgetting in Language Models via Implicit Inference},
  author = {Suhas Kotha and Jacob Mitchell Springer and Aditi Raghunathan},
  journal= {arXiv preprint arXiv:2309.10105},
  year   = {2024}
}

Comments

ICLR 2024

R2 v1 2026-06-28T12:25:22.152Z