Meta Launches 'Your Algorithm' AI Feature for Facebook
Meta launched the AI feature "Your Algorithm" for Facebook on Tuesday, allowing users to actively adjust the content appearing in their feeds and Reels. This feature was previously introduced on Instagram and has a similar tool called "Dear Algo" on Threads.
Users can access this feature through the three-dot menu on each post, allowing them to directly add or manage topics they want to see. They can also express "interested" or "not interested" for individual pieces of content. The platform may ask for more details, allowing users to describe in natural language what they want to see more or less of. Meta stated that the recommendation system will respond to these inputs in real-time and adapt within the same session.
The company is also increasing the proportion of large language models used in the ranking system. Large language models can understand the interests, context, and intent behind natural language, enabling the algorithm to receive feedback in a "conversational" manner rather than relying solely on behavioral signals like likes, comments, and watch time. The three-dot menu has been updated for easier access to these customization tools.
Previously, recommendations were primarily based on inferred interests from user interactions on the platform. This update incorporates natural language input as a direct signal, combined with existing ranking signals, aiming to make Feed and Reels recommendations more relevant and detailed. The feature covers feeds and Reels without changing the sorting logic of the follow list itself.
Users adjust recommendation weights through natural language feedback, and the platform will update content rankings in real-time. This initiative is driven by the company's proactive launch. More accurate interest matching may enhance user engagement and interaction while providing clearer preference signals for ad targeting; it complements the recommendation logic that relies solely on behavioral inference.
Source: Public Information
ABAB AI Insight
Meta previously launched "Your Algorithm" on Instagram, allowing users to view and adjust the list of topics influencing recommendations; on Threads, users expressed preferences through public posts via "Dear Algo," which later expanded to private adjustments. This similar tool is now introduced to Facebook feeds and Reels, enhancing the role of large language models in ranking, making natural language feedback a real-time effective signal.
User preference data is expanding from implicit behavioral signals to explicit language input. The platform embeds large language models into the ranking system, allowing natural language descriptions of "want to see more/less of a topic" to be directly converted into recommendation weight adjustments, effective within the same session. This shifts interest expression from indirect inference through likes and watch time to direct control through conversation, while retaining existing behavioral signals as a supplement.
Similar to TikTok adjusting recommendations through "not interested" and topic tagging, or YouTube allowing users to manage recommended topics, social platforms are returning some algorithm control to users. Currently, Meta is transitioning from pure behavioral inference to natural language explainable recommendations, with Facebook, as the largest user base platform, included in the same control framework.
This essentially represents a technological replacement. Traditional recommendations rely on statistical inferences of interaction behaviors, while large language models enable the system to parse intent and context in natural language, transforming user statements into real-time ranking signals. This change occurs because behavioral signals cannot cover short-term intent changes and fine-grained preferences, while language models provide the capability to directly map subjective descriptions to content rankings, shifting recommendations from black-box inference to conversational adjustments.
ABAB News · Cognitive Laws
Algorithm control shifts from behavioral inference to language declaration
Real-time feedback weights exceed the inertia of historical interactions
The next layer of recommendation systems is a conversational interest interface