LLM Unlearning via Loss Adjustments - Accenture Global Solutions

USPTO published patent application US20260099772A1 titled 'Large Language Model Unlearning via Loss Adjustments' by Accenture Global Solutions Limited on April 9, 2026. The application discloses a method and system for removing learned information from LLMs using a forget data only loss adjustment (FLAT) function, involving accessing forget data samples, associating template responses, and training a target LLM to maximize divergence between available template answers and forget answers.

Technology companies developing or deploying LLMs may find this patent application relevant to their intellectual property portfolios and product development strategies. The disclosed unlearning technique has potential applications in privacy compliance, content moderation, and model governance. However, as this is a published patent application (A1 publication), no enforceable rights or compliance obligations are established at this stage.