Researchers have developed a novel module called "Goggles" that can be applied during the fine-tuning of language models to instill a specific epistemic frame, such as identifying content as fictional. This module edits the gradients received by the model rather than altering the training data itself. When trained to recognize fictional content, models equipped with Goggles correctly identified fictional claims approximately 91% of the time, a significant improvement over the baseline 9% rate, while maintaining their overall capabilities. The Goggles module can also be trained for other frames, like treating documents as part of an AI safety evaluation, and its imparted frame remains persistent even under continued fine-tuning. AI
IMPACT This research offers a potential method for training language models on misaligned data without absorbing undesirable behaviors, improving their ability to discern factual from fictional content.
RANK_REASON The cluster contains an academic paper detailing a new method for training language models. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
