RoCo-ACE Improves LLM Knowledge Injection While Retaining Original Behavior.

A new method called RoCo-ACE has been introduced to enhance the process of injecting new factual or domain-specific knowledge into large language models (LLMs), particularly multimodal ones. The primary challenge in knowledge injection is preventing "catastrophic forgetting" or "drift," where updating a model with new information inadvertently degrades its performance on previously learned tasks or behaviors. RoCo-ACE addresses this by employing a sophisticated online distillation objective. It uses a rollout-conditioned likelihood contrast to strategically reallocate distillation weight to tokens supported by the new reference knowledge. Additionally, it incorporates sparse, anchored corrections for facts that might be omitted from the model's generated responses. This approach significantly improves the accuracy of injected knowledge across various settings and benchmarks, all while effectively maintaining the model's original capabilities and retention.