TL;DR
Meta researchers have deployed CLEAR, an end-to-end architecture that replaces opaque recommendation "black boxes" with a legible, editable natural-language interest layer. By grounding LLM-generated interests in actual user behavior and using them as stable "serving addresses," the system allows users to inspect, add, or suppress interests with immediate, measurable impact on their feeds.
What Problem Does This Paper Address?
Modern recommender systems are incredibly effective but notoriously opaque. They typically operate in "latent space"—mathematical representations that neither engineers nor users can easily inspect. If you suddenly start seeing too much "Reality TV," you might hit a "see less" button, but you have no way of knowing why the system thought you liked it or if your correction actually stuck.
The authors identify two missing properties in current systems:
- Legibility: Can a user look at a profile of their inferred interests and judge if it's accurate?
- Steerability: If a user edits that profile, does the recommendation engine actually change its behavior in a consequential way?
CLEAR (Controllable and Legible End-to-end Architecture for Recommendation) bridges this gap by using natural language as the primary interface between user intent and machine execution.
Methods
CLEAR isn't just a UI skin; it’s a fundamental re-architecting of the recommendation pipeline across four main stages:
- Interest Generation & Grounding: A specialized 3B-parameter Large Language Model (LLM) analyzes a user’s recent activity to propose interests (e.g., "Skating," "Watercolor tutorials"). Crucially, these aren't just guesses; the system performs behavioral grounding, mapping each interest back to specific events in the user's history to ensure the label is supported by evidence.
- The Interest Object as a "Serving Address": Instead of just being a text label, an interest becomes a multi-faceted query. It drives lexical retrieval (searching for keywords), behavioral retrieval (finding items similar to the "grounding" events), and embedding retrieval (searching vector space).
- The "Carrera" Write Path: When a user manually adds or removes an interest, this "authoritative state" takes precedence over model inferences. This allows for "cold-start" interests—things you want to see but haven't interacted with yet.
- The "Fireball" Delivery Controller: To ensure steerability, a real-time controller tracks "Realized Per-Interest Exposure" (RPE). If you add "Painting," the controller ensures you actually see painting content without "flooding" your feed, dampening the signal once you've had enough.
Datasets
The system is evaluated using Meta’s massive internal production data from Instagram and Facebook. For benchmarking the interest generator, the authors used:
- User-validated reference sets: 7,261 interest ratings from 393 real users to measure accuracy.
- GISTBench: A benchmark for evaluating how well LLMs understand user interests via evidence-based verification.
- Human-labeled "Golden Sets": 137 item-interest pairs used to calibrate the multimodal LLM judges that evaluate topicality.
Key Insights and Ideas
- Natural Language is the "API": By forcing the system to communicate in natural language, the researchers created a "contract" that both humans and different parts of the tech stack (retrieval, ranking, UI) can understand.
- Grounding is the Antidote to Hallucination: LLMs can "hallucinate" interests. By requiring a "support set" of actual user actions, CLEAR filters out fluent but unsupported interests before the user ever sees them.
- Two-Speed Architecture: The system separates the "slow" loop of model training from the "fast" loop of the delivery controller. This allows the system to respond to a user's edit immediately without needing to retrain a massive neural network.
- Attribution is Key to Control: Because every item shown can be "attributed" to a specific interest identity, the system can prove to the user (and the engineers) exactly why a specific video appeared in the feed.
Key Results
The paper provides impressive evidence from a year of production deployment:
- High Accuracy: In direct user reviews, 88.5% of the interests CLEAR inferred were rated as accurate by the users themselves.
- Model Performance: Their Specialized 3B generator outperformed much larger models, reaching a mean-user F1@10 of 0.3967, beating Gemini 2.5 Pro (0.3633) and their previous production model (0.1911).
- Proven Steerability: Through "stack ablation" (turning off parts of the system), they found that for a newly added interest, candidate sourcing contributed 13.7 percentage points to delivery, while reserved top-slot insertion contributed 11.1.
- Efficiency: The incremental inference design reduced the required "accelerator" (GPU/TPU) capacity by 35% compared to full profile regeneration.
Limitations and Future Work
- The "Latent" Remainder: The authors are careful to note that CLEAR exposes a subset of the system's state. It is an "authoritative layer," but it doesn't claim to represent the entire complex latent space used by the base rankers.
- Precision-Recall Trade-off: Stricter grounding (requiring more evidence before showing an interest) improves precision (accuracy) but reduces recall (the variety of interests shown). Finding the "sweet spot" remains an ongoing challenge.
- Complex Controls: Current controls are largely binary (add/remove). Future work could explore more nuanced natural language steering, such as "show me more beginner-level watercolor tutorials but fewer time-lapses."
For those interested in the intersection of LLMs and Information Retrieval, this paper serves as a rare, detailed look at how "scrutable" AI is being implemented at the scale of billions of users.
Summary updated