"Your Algorithm": Giving People Ways to Personalize Facebook Recommendations

Traditionally, the recommendation systems behind what people see on Facebook have learned from implicit engagement signals. While this allows them to scale well, it also means they adapt slowly and capture little nuance. But what if people could interact directly with the algorithms that rank and recommend what they see to instantly make their recommendations more personal and customized?

Your algorithm is an end-to-end system that addresses the limitations of adaptability and nuance in recommendations. This new suite of AI-powered tools lets people on Facebook state, in their own words, what they want to see in their Feed and Reels. By tasking a collection of large language models (LLMs) with understanding, storing, and tracking people’s stated preferences along with their user behavior, we’ve created a recommendation system that responds to people’s preferences and adjusts the content they see.

The Your algorithm page has three parts (Figure 1):

People can use the interface to add interests (using natural language), move them between lists, or remove them. Every change applies responsively to both Feed and Reels, and will further update as the person gives feedback on individual posts.

Your algorithm system has several components that cover the full lifecycle of a preference, from inferring and validating it to applying it in ranking and delivery (Figure 2).

Turning a stated preference into recommendations poses three distinct problems. An interest added to Interested In needs to deliver relevant content quickly, without crowding out everything else. An interest added to Not Interested In needs to hold reliably, even though topics are fuzzy and any single filter misses some matches. And both need to take effect responsively, even when the next recommendations were cached before the change.

Your algorithm addresses each problem with dedicated logic across the recommendation funnel, from retrieval through reranking to the device (Figure 3).

Each interest in Interested In drives four dedicated retrieval sources. Two find matching content directly (through search and through clusters of closely related topics), and two extend reach through the content graph and item-to-item similarity. Retrieving explicitly for each interest ensures enough relevant candidates for downstream ranking.

This is important because, without it, content about a brand-new interest might never enter the ranking pipeline at all. Candidates that pass a relevance check, which compares the meaning of the interest with each candidate, receive a promotion in late-stage ranking so they can compete alongside other content. A control layer in reranking then determines where they appear.

Responsiveness alone is not the goal. Our approach emphasizes balance over volume: A new interest should show up, not take over. The control layer spaces out recommendations for the same interest and caps each interest’s daily contribution. Retrieval also fetches candidates for only one interest per request, so every interest gets a turn. New interests receive extra weight for a limited period. After that, the recommendation system learns from how people engage with that content.

Interests in Not Interested In are enforced at several stages based on different signals to achieve high precision and recall. During ranking, the same embedding-based comparison used in the relevance check removes candidates too close to an interest in Not Interested In, and demotion rules push down near matches. After ranking, the LLM reranker reviews the final list. On the device, cache invalidation, described below, removes matching candidates that are already queued for display.

Not every entry in Not Interested In is about a topic. Quality-related requests, such as reducing AI-generated or sexually suggestive content, map to dedicated demotion signals rather than topic matching. When a request names a creator, that creator’s content is filtered out directly.

Even with ranking tuned to respond to preferences, there is still room to improve how responsively a change reaches what people see. Content for a new preference is not picked up until the next fetch request, and content already cached on the device was ranked before the preference was set.

We added two delivery mechanisms to close these gaps. When a new interest is added to Interested In, the system sends an extra fetch request outside the normal fetch cycle, called a supplementary tail load, to retrieve fresh recommendations so matching content appears quickly. When an interest is added to Not Interested In, cache invalidation handles candidates ranked before the change. The client obtains the new interest’s clusters and drops any cached content that matches them before it is displayed.

LLMs play a role at several key points in a preference’s life in Your algorithm. They generate the initial profile that people see on a first visit, validate each change as it is made, and give the final recommendations a last check against interests in Not Interested In (Figure 2). Each use comes with its own constraints, including latency, serving capacity, and safety, and evaluating them requires a new approach as well.

Profile generation provides the starting content for a first visit: the inferred interests and the Lately Summary. An LLM turns interest signals, which are refreshed regularly, into a set of inferred interests with clear, readable topic names. These are prefilled in the Interested In list and reflect past viewing across Facebook Feed and Reels. The LLM also writes the Lately Summary, a sentence that highlights ongoing interests. Together, they provide a recognizable starting point that is easy to review and correct.

Every edit on the Your algorithm page passes through preference management. When an interest is added or moved, an LLM checks that the phrase is a coherent topic, detects duplicates, and maps it to a clean topic name. This step runs on a smaller model that is efficient to serve at scale. The validated interest is then enriched with the signals ranking needs, such as an embedding and search keywords, and saved to the shared preference store. The LLM is only one layer of validation. The system also has a dedicated component that performs integrity and safety checks.

Once recommendations are ranked, LLM reranking provides a last check against interests in Not Interested In. The ranking funnel matches interests to candidates using embeddings, which are fast but coarse. To catch what they miss, a fine-tuned LLM reviews the ranked list and removes content it judges to be related to an interest recently added to Not Interested In. A key insight was that a small model fine-tuned to return only the items to remove was faster and more reliable than a general-purpose model judging each item individually.

Evaluating these LLM calls required its own approach. Standard offline metrics cannot tell whether a Lately Summary is accurate or whether the reranker removed the right candidates, so a separate LLM acts as a judge, grading each call’s output against expected behavior. This provides an objective, repeatable way to compare prompts and models before launch.

Your algorithm creates a more transparent and personalized recommendation experience, informed by both behavior and stated intent. But there are important opportunities remaining, such as better interpreting ambiguous input and making recommendations respond even more precisely to each request. We are also exploring additional entry points so that preferences can be shaped from more places.

This work reflects critical contributions from a cross-functional v-team across Meta (in alphabetical order):

Adi Gupta, Aryan Porwal, Bernabe Soto, Chao Yang, Chen Chen, Chen Kovacs, Chengxi Li, Chih-Wei Chang, Clare Yip, Cuong Nguyen, Dahni Ma, Daniel Thiyagu, Danny Kanoff, David Yuan, Devin Ding, Devon Pan, Eric Landgrebe, Feng Zhang, George Lu, Grace Yang, Hai-Son Dang, Haisen Hong, Hongyi Shi, Hongyi Wu, Howard Lu, Huijun Qian, Igor Vodov, Iordanis Fostiropoulos, Jakara Mato, Jasmine Neisser, Jason Liu, Jason Song, Jean-Pascal Billaud, Jeffrey Lin, Jeffrey Santelli, Jian Chen, Jianyu Wang, Jiaye Pan, Jinho Hwang, Joe Sena, Jordan Reich, Kanghua Wu, Khoi Pham, Kim Trang Leaf, Kris Katsanevas, Kuo-Feng Hsu, Kushal Adhvaryu, Liyu Wang, Manish Khettry, Matthew Miao, Mayoor Rai, Momo Jiao, Nakul Tirumalai, Patrick Boocock, Patrick Cullen, Qi Guo, Qichao Que, Queenie Li, Ratish Nair, Sanyogeeta Lawande, Satvik Eltepu, Sean Chen, Shijie Xie, Shuai Zhang, Shuoyan Yin, Smruthi Gowtham, Tsung-Yi Huang, Tyler Pell, Wanting Jiang, Wanyu Li, Weiwei Jiang, Xiangyu Xu, Xin Wang, Xinyang Feng, Yanchen Wang, Yibo Zhu, Yihao Li, Yihui Saw, Yinjie Huang, Yu-Wei Wu, Yuchen Shao, Yue Wang, Yue Zhang, Yuhan Yang, Yuqi Shen, Zhen Hang Jiang, Zhiyu Liu, Zhongwei Teng, Zijiang Yang, Zoe Zu

We also thank the following leaders for their sponsorship and strong leadership support (in alphabetical order):

Ashish Gandhe, Helen Kong, Hongjie Bai, Meng Wu, Yaping Zhang, Yisong Song, Zellux Wang, Zhe Wang