Dear Algo: A Precision-First Agentic Intent Layer for Unified Search and Recommendation
Rui Wang, Jiazhou Wang, Zheng Wei, Chenglin Lu, Fangcheng Sun, Ivy Sun, Jin Sun, Hui Geng, Lillian Zhang, Chao Yang, Lei Chen, Shahin Sefati, Reem Helou, Joe Zhou, Babak Shakibi, Yiyi Pan, Bi Xue, Hong Yan, Shujian Bu
cs.AI
2026-08-17
Meta compiles natural-language feed requests on Threads into executable retrieval plans; a strict relevance gate hits 94.4% human-validated precision.
Search encodes intent as a short-lived query; recommendation infers long-lived preference from behavior. Meta deployed Dear Algo on Threads to bridge the two: users address the algorithm in natural language, and requests like "more NBA news" steer subsequent feed recommendations persistently, with the resulting controls visible in Settings. This paper, all nineteen authors at Meta, describes that deployed product's architecture and evaluation. The design constraint is an asymmetry: an irrelevant item admitted into a persistent feedback loop causes repeated unwanted exposure, while rejection merely lowers coverage. Everything is built precision-first.
The intent layer runs on SilverTorch, Meta's model-based GPU recommender serving system, and turns a request into an executable plan in three steps:
Plans compile once and execute many times: expensive intent reasoning runs once, the plan persists and refreshes against new content, negative intent becomes suppression clauses, and compound intent decomposes into deduplicated unit plans. Search-like and recommendation-like modes share the intent-to-retrieval contract while keeping their own retrieval, ranking, thresholds, and latency budgets. The paper is careful with the word "unification": infrastructure and representation are shared, causal cross-mode improvement is not yet demonstrated.
| Study | Result |
| Blinded human audit (300 pairs, 296 evaluable) | strict-judge exact-Relevant precision 94.4% [88.8%, 98.9%] |
| Paired offline study (72 request clusters, 20 slots) | 7.73 judge-qualified candidates vs 6.61 for an LLM-derived-query baseline (+1.11 [0.12, 2.12]) |
| Candidate-randomized reranker study (72h window, 1,278 users) | judge-Irrelevant share 2.80% vs 4.78% (-1.97 points [-3.02, -0.94]) |
| Observational usage (3 cohorts, 131K users) | +4.15% to +8.61% next-week app usage; report-rate intervals all include zero |
The four studies answer different questions, and the paper explicitly declines to merge them into one significance claim. That restraint is rare in industrial papers.
This is one of the few public, numbered accounts of LLM interpretation steering a live recommender. Three things transfer: the interpret-ground-compile split lets open language into production retrieval without giving up inspectability; compile-once, execute-many amortizes expensive reasoning across content refreshes; and the four-tier precision-first evaluation (calibration, paired offline, randomized serving, behavioral) is a reusable template.
The authors' own list: the usage study is observational, with self-selection, trends, and regression to the mean as confounders; all offline arms use grounded tags and STQL, so grounding's individual contribution is not isolated; the precision-first endpoints were clarified after initial inspection, not preregistered; the reranker study observes only admitted candidates, with MNAR bounds spanning both directions. The production prompt and model are not disclosed, so replication is off the table. One gap the paper leaves open: the strict gate admits 84 of 176 human-Relevant pairs in the audit sample, so precision is bought with coverage, and no measure is given of what that trade costs the user experience.