ArXiv TLDR

ORBIT: Preserving Foundational Language Capabilities in GenRetrieval via Origin-Regulated Merging

🐦 Tweet
2605.12419

Neha Verma, Nikhil Mehta, Shao-Chuan Wang, Naijing Zhang, Alicia Tsai + 5 more

cs.CLcs.IRcs.LG

TLDR

ORBIT prevents catastrophic forgetting in GenRetrieval LLMs by regulating weight drift, preserving foundational language capabilities.

Key contributions

  • Identifies catastrophic forgetting in GenRetrieval LLMs, correlating it with parameter drift from original weights.
  • Proposes ORBIT, a novel method tracking fine-tuned model weight distance from initial parameters.
  • ORBIT uses a weight averaging strategy to constrain model drift when a parameter distance threshold is exceeded.
  • Demonstrates ORBIT significantly retains text and retrieval performance, outperforming common continual learning baselines.

Why it matters

Fine-tuning LLMs often leads to catastrophic forgetting of general abilities, limiting their real-world applicability. ORBIT provides a practical solution to maintain these crucial foundational capabilities during task-specific adaptation, improving the robustness and versatility of specialized LLMs.

Original Abstract

Despite the rapid advancements in large language model (LLM) development, fine-tuning them for specific tasks often results in the catastrophic forgetting of their general, language-based reasoning abilities. This work investigates and addresses this challenge in the context of the Generative Retrieval (GenRetrieval) task. During GenRetrieval fine-tuning, we find this forgetting occurs rapidly and correlates with the distance between the fine-tuned and original model parameters. Given these observations, we propose ORBIT, a novel approach that actively tracks the distance between fine-tuned and initial model weights, and uses a weight averaging strategy to constrain model drift during GenRetrieval fine-tuning when this inter-model distance exceeds a maximum threshold. Our results show that ORBIT retains substantial text and retrieval performance by outperforming both common continual learning baselines and related regularization methods that also employ weight averaging.

📬 Weekly AI Paper Digest

Get the top 10 AI/ML arXiv papers from the week — summarized, scored, and delivered to your inbox every Monday.