Epistemic Goggles Train LLMs to Distinguish Fact from Fiction

Large Language Models often struggle with "Negation Neglect," where finetuning on fictional content, even when explicitly labeled as such, still leads the model to believe the core claims. Models trained this way correctly identify fictional claims only about 9% of the time. To address this, a new module called "Goggles" has been introduced. Goggles operates by editing the gradients an LLM's LoRA receives during supervised finetuning, rather than modifying the data itself. This process imparts a specific "epistemic frame" – the model's stance on the nature of the information it processes. Once trained for a given base model and configuration, a Goggles instance can be applied frozen to new documents. When trained through Goggles, models can correctly flag content as fictional approximately 91% of the time, while maintaining or even exceeding baseline capabilities on benchmarks like GPQA and TruthfulQA. The architecture is flexible, allowing for other frames, such as treating documents as part of an AI safety evaluation. This intervention is robust, with the imparted frame persisting even under continued finetuning that might otherwise revert prior changes, offering a promising method for training LLMs on potentially misaligned data without internalizing its problematic aspects.