WorksComputer Science
CLEAR: Legible and Steerable Recommendation via Grounded Natural-Language Interests
Abstract
Modern recommenders keep learned preferences in representations people cannot inspect or change. We present CLEAR (Controllable and Legible End-to-end Architecture for Recommendation), an end-to-end system deployed at Meta and built around an editable natural-language interest layer. CLEAR makes a ranked subset of inferred interests readable and stores accepted additions and suppressions as authoritative preference state. Behavioral grounding maps each inferred interest back to supporting events in the user's own history. The resulting interest object is also a serving address: its label and profile-conditioned rewrites drive lexical retrieval, its supporting events seed behavioral retrieval, its projection queries an embedding index, and its stable identity supports ranking, attribution, and exposure control. A fresh user declaration may precede behavioral support, so it enters through the same serving identity and initially draws candidates from topical retrieval. Ranking makes those candidates competitive, while a delivery controller observes attributable exposure and replenishes underexposed interests. Direct user review found that users rated about 88.5% of displayed interests accurate. In an offline benchmark using a shared evaluation protocol, the Specialized 3B generator reached 0.3967 mean-user F1@10, compared with 0.3633 for Gemini 2.5 Pro and 0.1911 for the prior generator. For a newly declared interest, a user-randomized cumulative ablation measured same-day delivery losses of 8.7 percentage points for added-interest ranking terms, 11.1 for reserved top-slot insertion, and 13.7 for interest-conditioned candidate sourcing, conditional on the tested removal sequence. CLEAR components have operated in production for over one year; the system spans four surfaces on Instagram and Facebook, with an interest layer people can inspect, correct, and use to direct recommendations.
CCS Concepts: Information systems—Recommender systems.
Introduction
Recommenders continually infer what a person may want, yet their operative representations are usually latent and inaccessible to the person they describe. Item-level reactions can influence future predictions, but they do not reveal what durable preference the system inferred or whether a correction persisted. A useful control surface therefore needs two linked properties: the inferred interest state must be legible, and an accepted edit must be consequential.
We call an interest layer legible when its natural-language entries let people assess whether the system’s inferences match their interests. Recommendations are steerable through the layer when accepted edits become persistent preference state consumed by serving, and controlled changes to that serving path measurably change later delivery. This definition separates seeing an interest from proving that the recommender acts on it.
CLEAR joins the read and write paths through one normalized interest identity (Figure 1). It converts recent activity into natural-language candidates, behaviorally grounds inferred candidates, exposes a ranked subset, and stores explicit additions and suppressions with precedence over later inference. The same identity is read by retrieval, ranking, cross-source attribution, and delivery control. It is an authoritative interest layer inside the recommender, not a claim that the system’s entire latent state is visible.
Figure 1. Generate, inspect, edit, retrieve, and re-rank through one interest layer. An LLM proposes interests from recent activity, and behavioral grounding checks inferred entries before they enter the displayed profile. The user suppresses Reality TV and adds Painting. A fresh addition receives a stable serving identity immediately even when it has no prior support events; topical retrieval adds Painting candidates, and ranking makes them competitive in the next feed. Subsequent responses return through ordinary interaction history. Rankings are schematic.
The system must bridge two kinds of meaning. For an inferred entry, grounding asks which of this user’s events support the phrase. For recommendation serving, the same interest object must identify where to retrieve, which candidates are on topic, which delivered items count toward the interest, and when more supply is needed. Grounding is therefore not confined to profile construction: it is the interface between natural-language state and the recommendation cascade.
We evaluate the system in the same order. Direct review tests whether displayed entries are readable and accurate. A randomized cumulative ablation then tests how much each serving layer contributes to delivery after a positive edit. Supporting evaluations quantify grounding, incremental updates, and deployment efficiency.
This paper contributes:
an editable interest-state contract that joins an inspectable natural-language profile to recommendation serving (Figure 1);
a grounding interface that turns inferred interests into personalized retrieval anchors and stable recommendation-side identities (§ 4, Figure 2 and Figure 3, and Table A2);
randomized evidence that sourcing, placement, and ranking contribute to positive-edit delivery, and
Figure 2. Grounding gives one editable interest four retrieval addresses and a stable serving identity. The display label becomes a lexical query, profile-conditioned rewrites expand that query, inferred interests retain behavioral support for trigger-to-item retrieval, the interest projection becomes an embedding query, and the user–interest identity can condition generative retrieval and cached supply. Candidates merge, pass a hard topicality gate, and enter ranking. Attribution preserves the same interest identity for delivery control; explicit negative state remains in the policy path.
Figure 3. Grounding precision–coverage trade-off. Curves show threshold sweeps for two single-channel embedding matchers. The deployed text-plus-media union is evaluated separately and summarized below the plot. Support precision concerns judged event–interest matches; coverage counts deduplicated history events. Table A2 reports the selected operating-point results.
| Evaluation | Configuration | Metric | Result |
| Profile filtering frontier: increasingly strict support | |||
| Profile filtering | Unfiltered candidates | Precision | 69.9% |
| Profile filtering | Unfiltered candidates | Reference recall | 46.2% |
| Profile filtering | At least one supporting event | Precision | 74.5% |
| Profile filtering | At least one supporting event | Reference recall | 41.5% |
| Profile filtering | At least three supporting events (selected) | Precision | 77.4% |
| Profile filtering | At least three supporting events (selected) | Reference recall | 28.7% |
| Event grounding over the user’s history | |||
| Event grounding | Text + media matcher | Gemini-judged event–interest support precision | 77.2% |
| Event grounding | Text + media matcher | History events matched to at least one interest | 54.4% |
Table A2. Grounding quality and coverage. Profile filtering forms a whole-profile precision–recall frontier: increasingly strict support retains fewer interests while raising precision. The selected three-event point prioritizes whole-profile precision for displayed interests. Event grounding separately evaluates retrieved history-event and inferred-interest pairs.
similarity-aware demotion reduces suppressed-interest reappearance (Table 1 and Table 3); and
| Study | Metric | Result |
| A. Legibility and profile quality | ||
| Direct review | Displayed interests rated accurate | |
| Specialized 3B generator | Mean-user F1@10 / P@10 / R@10 / F1@all | 0.3967 / 0.8067 / 0.2901 / 0.5493 |
| Gemini 2.5 Pro | Mean-user F1@10 / P@10 / R@10 / F1@all | 0.3633 / 0.7982 / 0.2568 / 0.7051 |
| Prior production generator | Mean-user F1@10 / P@10 / R@10 / F1@all | 0.1911 / 0.5692 / 0.1281 / 0.4281 |
| B. Steerability | ||
| Same-day responsiveness | Share of a user’s newly declared interests receiving a matching video the same day | 79.6% |
| Prevalent declared-interest load | Users whose declared-interest content is at least 5% of their recommendations | 82.6% |
| Topicality | Interest-matched impressions judged on-topic by an independent multimodal judge | 87.2% |
| Replenishment | 3s-engaged interest responsiveness / user days with explicit-interest delivery | +8.34% [+8.07%, +8.61%] / +2.43% [+2.24%, +2.63%] |
| Volume regulation | Recorded flooding-days metric versus prior daily accounting | 15.4% lower [13.4%, 17.4%] |
Table 1. Legibility and steerability. Direct review and system-level rows report measured levels; replenishment and volume regulation report relative changes. Direct review is descriptive, and the generator comparison is an offline benchmark using a shared history format, user-validated reference, and semantic judge. Responsiveness and prevalence are one-week Instagram production averages; topicality is the most recent twice-weekly judge run. Responsiveness is attributed matching delivery on the day of the edit; topicality is scored by a judge independent of the serving gate and validated against human labels (§ 6.3). Replenishment is a separate user-randomized comparison against the prior age-limited policy. Volume regulation compares real-time per-interest accounting and recalibrated dampening with prior daily accounting. Component-level contrasts appear in Table 3.
| Layer removed | Loss in same-day delivery |
| Interest-conditioned candidate sourcing | 13.7 pp |
| Forced top-slot insertion | 11.1 pp |
| Added-interest ranking boosts | 8.7 pp |
| All layers removed | 33.5 pp |
| Treatment | Change in same-day suppressed-interest reappearance |
| Similarity-aware demotion beyond the production filter | 11.8% lower |
Table 3. Controlled serving-component effects for positive and negative edits. Panel A is a five-arm user-randomized cumulative ablation of Instagram’s positive-addition delivery layers. Component rows report consecutive-arm delivery losses in the tested removal order; the full-stack row reports the outer-arm contrast. All reported randomized serving-component comparisons have . The endpoint uses a slightly narrower definition than the same-day responsiveness in Table 1 and is read as an approximation of it. Because the arms are nested, the decomposition is specific to this removal order and is not an order-free attribution. Negative filtering and suppression produced no measurable change on this positive-addition endpoint and are omitted from Panel A. Panel B is a separate user-randomized suppression experiment; lower suppressed-interest reappearance is better. A. Positive addition. B. Negative suppression.
a deployed design whose incremental inference and shared annotation reduce serving cost while preserving profile quality (Table A3–Table A5).
| Configuration | Metric | Result |
| Full regeneration (1,000 users) | Whole-profile precision / recall / F1 | 0.6375 / 0.1101 / 0.1819 |
| Incremental merging (1,000 users) | Whole-profile precision / recall / F1 | 0.6501 / 0.1135 / 0.1867 |
| Incremental inference versus preceding history budget | Estimated accelerator requirement | 35% lower |
Table A3. Incremental profile updates. Full regeneration and incremental merging use the same fixed reference; whole-profile precision, recall, and F1 are reported, with for the incremental comparison. Capacity reduction is measured against the preceding history-budget configuration.
| Lever | Metric | Result |
| Grouped versus uncompressed history | Prompt tokens | 39.5% fewer |
| History budgeting versus unbudgeted history | Estimated accelerator requirement | 38.9% lower |
| Per-string versus per-user annotation | Logical annotation requests | ≈98% fewer |
| Lever | Metric | Result |
| Updated versus preceding annotation pipeline | Throughput per accelerator | +145.5% |
| Parallel versus serial request-time attribution | Chunk-0 server latency | ≈5 ms lower |
| Deployment | Enabled surfaces / platforms (Instagram and Facebook) | 4 / 2 |
| Deployment | Time in production | More than one year |
Table A5. Deployment efficiency and breadth. Each row has its own baseline; deployment counts enabled surfaces across Instagram and Facebook.
Related Work
Scrutable and editable profiles. Natural-language user profiles have been proposed as transparent, revisable interfaces to recommendation ([9, 10]). Prior systems study simulated edits and reranking ([8, 20]), direct edits to concept profiles ([7]), natural-language control requests ([2]), and preference masking ([5]). CLEAR studies a complementary systems question: how one open-vocabulary interest identity can persist from an inferred profile or user edit through candidate supply, ranking, attribution, and exposure-aware delivery in an existing recommender.
Field and interface studies further examine how people reflect on, inspect, and revise interest summaries ([14, 17]). Stateful profile systems such as LIBER and RECAP update textual representations as behavior evolves ([21, 19]). CLEAR’s contribution is the deployed state contract that gives one editable identity persistent precedence and makes it consumable through the recommendation cascade without replacing the incumbent rankers.
Textual representations and deployed controls. Consumer controls commonly adjust service-defined topics, follows, or individual items ([13] (Google, n.d.; TikTok 2025)). LLM-derived textual user representations also support recommendation ([1, 16, 12, 18, 21]). CLEAR makes an inferred textual representation inspectable, gives explicit edits precedence, and measures the recommendation components that consume those edits.
Grounding language in recommendation. Attribution asks whether generated claims are supported by their sources ([11, 6]). For user interests, the source is the user’s own interaction history; GISTBench evaluates whether inferred interests can be verified against that evidence ([3]). CLEAR operationalizes this connection. The support set filters and orders inferred interests, while the interest’s label, projection, anchors, and identity make the same concept addressable to recommendation serving. The contribution is this conjunction with durable edit precedence and measured production actuation, not a claim that editable profiles or any individual retrieval mechanism is new.
Editable Interest State
The profile object
In production, Meta’s Biography service builds CLEAR’s inferred interest profiles. Instagram’s Your Algorithm and Tune Your Explore interfaces expose the interest layer for inspection and editing.
CLEAR represents interests at coarse, intermediate, and open-vocabulary levels. Coarser entries provide stable organization; open-vocabulary entries provide the specificity needed for retrieval. A surface displays a ranked subset of a larger served interest layer using natural-language labels rather than model coordinates. This displayed subset is the legibility surface evaluated below.
Each served record binds a normalized interest identity to its source, supporting evidence when available, query projection, shared annotations, and preference state. State can be inferred, explicitly added or promoted, or explicitly removed or suppressed. Explicit state is stored separately from model output and takes precedence until the user changes it. Inferred interests may expire when their support decays; accepted user preferences persist across inference refreshes. Table A6 summarizes the shared state and action contract.
| State | Entry condition | Persistence and precedence | Serving effect |
| Inferred | Generated from activity and retained after behavioral support and policy checks | Refreshed with the inferred profile; may expire when support decays | Available to profile ranking, retrieval, attribution, and display selection |
| Explicit addition | Accepted normalized user text; prior behavioral support is optional | Stored separately from inference and persists until the user changes it | Creates an authoritative positive interest identity for topical supply and ranking |
| Explicit promotion | A user promotes an existing normalized identity | Explicit preference outranks later inferred state | Raises the priority of matching candidates through the shared identity |
| Explicit removal or suppression | A user removes an entry or requests less of its normalized identity | Stored as explicit state and not overwritten by later inference | Excludes the entry from the served profile or applies the corresponding negative serving path |
Table A6. State and action contract. Verified state transitions used by profile construction and serving.
Profile construction
Construction runs asynchronously over a bounded representation of recent activity. History selection favors topical coverage rather than using recency alone, and related events are grouped before the history and contextual signals enter the language model. The generator proposes candidate interests. Generation, behavioral grounding, policy filtering, deduplication, and display ranking are separate stages so that a fluent candidate is not automatically treated as supported or shown.
New candidates merge with the prior inferred profile. Removing old history may remove behavioral support; recency changes the ordering of entries that remain supported. Explicit additions and suppressions are kept outside this inferred-state merge, which prevents later inference from silently undoing a user’s correction.
Incremental state and shared annotation
Incremental inference reuses the prior profile and processes new activity instead of regenerating from the full history. The bounded profile is written to a serving cache. User-independent annotations, such as normalized wording and sensitivity, attach once to each distinct interest string and are reused across user records. User-specific membership, support events, and explicit preference state are never shared.
Grounding Natural-Language Interests
Behavioral grounding for inferred interests
An inferred phrase is useful only if the system can relate it to what the user actually did. CLEAR projects the phrase into the recommender’s co-engagement space and performs reverse lookup over the user’s own history. A content-understanding channel and a co-engagement-trained media channel retrieve candidate support events; their deduplicated union forms the interest’s support set.
The support set has three profile-side consumers. Its size filters inferred candidates with insufficient evidence. Its overlap reveals personalized near-duplicates: two phrases may be redundant for one user when the same activity supports both. Event actions and recency supply an evidence-weighted ordering signal. Grounding therefore verifies behavioral support, while direct user review separately tests whether the person endorses the resulting description.
Appendix Figure 6 shows how the matched events are reused by profile filtering, deduplication, ordering, and coverage accounting.
Figure 6. Behavioral grounding and profile-side reuse. Reverse lookup supplies support filtering, personalized deduplication, ordering, and history coverage. Recommendation-side consumers appear in Figure 2; explicit additions require no prior support.
A user declaration follows a different contract. A fresh addition can precede any matching history and therefore bypasses the inferred-candidate support requirement. It still receives the same normalized identity, query projection, policy checks, and topical matching path. If relevant prior activity exists, it can supply personalized anchors immediately; later responses enter ordinary history and can support subsequent inference. This distinction avoids treating a cold-start declaration as evidence about past behavior.
Grounding as a recommendation-side address
The grounded interest object exposes four fields to the recommendation cascade:
the label, expanded with profile-conditioned rewrites, becomes a personalized lexical query;
the support events seed trigger-to-item retrieval in the user’s own co-engagement region;
the query projection retrieves nearby items from an approximate-nearest-neighbor index; and
the stable user–interest identity keys cached supply, attribution, and exposure accounting.
Behavioral retrieval uses a conjunction rather than either signal alone: a candidate must lie near the user’s own support events and match the interest under content-understanding features. This excludes a globally on-topic item unrelated to the user’s behavioral neighborhood and a behaviorally adjacent item that is off topic. The lexical and embedding paths provide initial supply for a novel declaration; the support-set path personalizes that supply when behavioral evidence exists. A separate generative source can condition on the same user–interest key.
Stable identity also closes the recommendation-side loop. Ordinary and dedicated-source candidates can be attributed to the same declared interest, so the controller observes total recent exposure rather than only what its own retriever injected. The same identity then determines whether the system should request more topical candidates or reduce repeated supply. Figure 2 shows the resulting path from the editable interest object through retrieval, topicality, ranking, attribution, and delivery.
From an Edit to Delivery
Write contract and candidate supply
At Meta, the declared-interest serving path is called Carrera.
User text is parsed and normalized before storage. Positive actions add or promote an identity; negative actions remove or suppress it. The same integrity gates apply to inferred and user-entered interests. An accepted edit immediately updates the served layer, and explicit state remains authoritative across later inference.
Four interest-conditioned retrieval mechanisms emit tagged candidates: lexical retrieval, trigger-to-item retrieval, embedding retrieval, and generative retrieval. An asynchronous path precomputes per-user, per-interest candidates; a request-time path handles fresh declarations and stale cached supply. Candidates merge with the ordinary pool before ranking. Candidate supply therefore changes which items can compete, rather than merely reordering a fixed pool.
Ranking, topical matching, and attribution
Early- and late-stage ranking apply temporary declaration-conditioned terms so candidates for a new interest can compete before sufficient behavior accumulates. A small number of top-ranked positions can also be reserved for candidates matching a recent declaration, guaranteeing placement rather than only competitiveness. Negative state is read by a serving-time removal path. Both operations act on the normalized identity, not only its displayed text.
A compact learned scorer estimates whether an item matches an interest from semantic and content-understanding features. Retrieval uses a hard topicality gate; ranking can use the score softly. The same scorer also attributes ordinary recommender candidates. Source identity answers where a candidate came from, while topical attribution answers which interest it matches. These are intentionally separate: an ordinary candidate can satisfy a declared interest without coming from dedicated retrieval. This item–interest scorer is distinct from both the semantic judge used to compare generated profiles and the history-event matcher used to ground inferred interests.
Exposure-aware delivery
Naively boosting every match can flood one topic. CLEAR separates topical relevance from how much recent exposure that user–interest pair has received. Attribution runs before the controller, allowing exposure accounting to include both ordinary and dedicated-source matches.
Fireball maintains a decayed real-time counter for realized per-interest exposure (RPE). Higher exposure dampens further promotion; as exposure decays, the interest becomes eligible for renewed sourcing. Replenishment dynamically reserves recent declarations and prioritizes eligible interests with least recent realized exposure, subject to a recent-view threshold. Replenishment changes whether an interest receives fresh supply, whereas ranking changes the competitiveness of candidates already present.
The controller never writes model parameters. Delivered items and user responses enter the recommender’s ordinary learning loop, which may later improve organic supply. The controller’s faster loop only retrieves, ranks, attributes, and regulates current delivery.
Appendix Figure 5 summarizes the fast exposure loop and slower learning loop.
Together these paths form an import, teach, harvest, and re-teach design loop. A fresh declaration first imports topical candidates because the base recommender may have little evidence for it. User responses to those candidates become ordinary behavioral signal. As the recommender later supplies organic candidates attributed to the same interest, the controller can count and use that personalized supply instead of relying only on imported candidates. If attributable exposure later declines, replenishment reopens topical sourcing. This sequence explains the controller design, and the selected studies evaluate its supply, ranking, replenishment, and reception components.
Evaluation
Bracketed uncertainty intervals denote 95% confidence intervals.
Three distinct measurement layers. Generator evaluation compares natural-language interests with a user-validated interest reference using a semantic judge. Behavioral-grounding evaluation uses a Gemini LLM judge on retrieved history-event/inferred-interest pairs; Vital or Reasonable counts as supported. Serving evaluation instead scores delivered-item/declared-interest pairs. On 137 human-labelled event-interest pairs, grounding-judge variants agreed with human labels at 93–98%, with Cohen’s of 0.86–0.95. Direct review remains a user judgment over displayed interests. Table A7 collects the study-level comparison and analysis contracts.
| Study | Treatment and comparison | Analysis population / unit | Endpoint and uncertainty |
| Direct review | One deployed displayed profile; descriptive review | 7,261 displayed interests from 393 users; pooled over entries | Entry rated accurate; descriptive percentage |
| Generator benchmark | Specialized 3B, Gemini 2.5 Pro, and prior generator; shared history format, reference, and judge | Specialized and prior: 250 users each; Gemini 2.5 Pro: 232 users | Mean-user F1@10 / P@10 / R@10 / F1@all; reported model means |
| System-level steerability | Deployed-system descriptive measurements | Newly declared interest, eligible user, and sampled attributed impression | 79.6% responsiveness, 82.6% prevalence, and 87.2% independently judged topicality |
| Instagram Reels reception | Attributed versus other recommendations for the declared-interest cohort | Delivered impressions and content-control actions | Selected levels and contrasts |
| Instagram Search reference | Attributed recommendations versus search-result videos | Delivered impressions; common user sample for skips | Relative like and dwell; percentage-point skip contrast |
| Serving-layer stack ablation | Four cumulative removal arms and production control (Appendix B) | Eligible positive-interest lifecycle | Attributed matching delivery same day and within 16 impressions; |
| Study | Treatment and comparison | Analysis population / unit | Endpoint and uncertainty |
| Suppression-aware ranking | Similarity-aware demotion plus production filter versus filter alone | Suppressed interest in subsequent recommendations | Reappearance 11.8% lower same day, 21.8% at 7 days, 15.3% at 14 days, and 24.0% at 28 days; |
| Replenishment | Dynamic topic selection versus the prior fixed sourcing-age policy; separate assignment | Instagram user; active interests for the engaged-interest ratio, and distinct days for delivery | Active interests with a match watched for at least three seconds: +8.34% [+8.07%, +8.61%]; User days with explicit-interest delivery: +2.43% [+2.24%, +2.63%] |
| Volume regulation | Real-time per-interest accounting and recalibrated volume dampening versus prior daily accounting | Instagram user; six-day comparison | Recorded flooding-days metric: 15.4% lower [13.4%, 17.4%] |
Table A7. Study contracts for Tables 1–3. All online component comparisons are randomized by user. Replenishment and volume-regulation values show relative changes with bracketed uncertainty intervals.
Legibility and profile quality
We operationalize legibility as displayed-profile accuracy: a person can inspect a natural-language entry and judge whether it reflects their interests. Direct review pools 7,261 interest ratings from 393 users. Rejection reasons are analyzed separately to distinguish exposure from endorsement.
The current offline generator benchmark uses a shared evaluation protocol for the Specialized 3B generator, Gemini 2.5 Pro, and the prior production generator. All receive the same history format and are scored against the same user-validated reference by the same semantic judge. Users validate the reference by confirming proposed interests and supplying omissions. F1@10, P@10, and R@10 evaluate the first ten emitted interests; F1@all evaluates the complete generated profile. Table 1 reports mean-user F1@10 first, followed by precision, recall, and whole-profile F1. Generation is evaluated before behavioral grounding and display selection.
The Specialized 3B generator is trained with a task-specific curriculum aligned to the fixed top-ten interest-generation objective. At that operating point, it reaches 0.8067 P@10, compared with 0.7982 for Gemini 2.5 Pro.
Grounding
The support-filter study compares retained inferred profiles with an LLM-generated whole-profile reference. Precision measures the fraction of retained interests matching the reference; recall measures the fraction of reference interests retained. It is distinct from F1@10.
The event-level grounding study judges whether retrieved history events support an interest. Vital denotes strong relevance to the post’s most fine-grained information; Reasonable denotes broad relevance to its general topic. Both receive equal positive credit in this history-matcher evaluation. Support precision is the fraction of event–interest matches judged relevant. History-event coverage is the fraction of a user’s history matched to at least one interest, with an event counted once even when several interests retrieve it.
Steerability
We evaluate Instagram’s declared-interest serving at two levels. At the system level we report three production measures. Same-day responsiveness is defined per user as the share of that user’s newly declared interests that received at least one attributed matching video later the same day, averaged over users who subsequently viewed more than ten videos. Prevalent declared-interest load is the share of eligible users holding a declared interest for whom content attributed to declared interests reaches at least five percent of their recommendations. Topicality is evaluated by a judge independent of serving. Twice weekly, approximately 8,000 (video, interest) pairs, about 10,500 impressions, are sampled from impressions attributed to explicit declared interests, and a multimodal judge consuming the video content directly, rather than captions or metadata, adjudicates whether the item matches the interest it was served against. The reported value is the impression-weighted share of matching pairs. This judge is not the serving topicality scorer, so it does not inherit the admission criterion it evaluates. The judge is itself validated against human labels before use and re-validated across rounds as prompts and models change: on a human-labelled golden set of 137 (item, interest) pairs, deployed judge variants agree with human labels at 93–98% (Cohen’s –) and with one another at 95–96% (–). The calibration set oversamples edge cases and is therefore not production-representative.
At the component level, a five-arm stack ablation removes serving layers cumulatively from the production configuration: added-interest ranking boosts, then forced top-slot insertion, then negative filtering and suppression, then interest-conditioned candidate sourcing. Differences between consecutive arms give the marginal delivery loss at each removal step; the outer-arm contrast compares all layers removed with the production control. The endpoint is same-day attributed matching delivery.
A separate randomized study compares dynamic topic selection with the prior age-limited policy, measuring the share of active interests receiving a match watched for at least three seconds and user days with explicit-interest delivery.
A six-day user-randomized study compares real-time per-interest accounting and recalibrated volume dampening with the prior daily-accounting configuration. Its endpoint is the recorded flooding-days metric; lower is better.
Results
Displayed interests are accurate; the specialized generator improves F1@10
Across 7,261 displayed interests rated by 393 users, approximately 88.5% were rated accurate (Table 1). In the separate offline benchmark using a shared evaluation protocol, the Specialized 3B generator reached 0.3967 mean-user F1@10, compared with 0.3633 for Gemini 2.5 Pro and 0.1911 for the prior generator. F1@10 evaluates the fixed top-ten generation objective. Gemini 2.5 Pro achieved higher whole-profile F1@all: 0.7051 versus 0.5493 for the Specialized 3B generator. The direct review evaluates the shipped display; the benchmark isolates generation before grounding and display selection.
Rejected entries also show why support and endorsement differ. Among rejected interests, 41% reflected content the user had viewed without personal interest, 29% reflected transient interests, and 21% reflected no personal interest (Table A1). A readable profile exposes these mismatches so the user can correct them.
| Finding | Result |
| Viewed without personal interest | 41% of rejections |
| Transient interest | 29% of rejections |
| No personal interest | 21% of rejections |
| Other responses | of rejections |
Table A1. Why profile corrections are necessary. Reasons users gave for rejecting a displayed interest. Categories sum to approximately 100% after rounding.
Stricter grounding trades reference recall for whole-profile precision
Stricter grounding trades reference recall for whole-profile precision
Increasing the minimum support from zero to three events moved whole-profile precision from 69.9% to 77.4% and reference recall from 46.2% to 28.7%, the expected precision–recall trade-off from stricter grounding. The threshold sensitivity analysis selects the three-event point to prioritize whole-profile precision for displayed interests (Table A2). In the separate event-level evaluation, the deployed text-plus-media matcher reached 77.2% judged support precision and matched 54.4% of deduplicated history events to at least one retained interest.
The deployed system responds to edits and delivers on-topic content
Before isolating components, we report how the deployed system behaves as a whole in Instagram production. On the day a user declared new interests, the average user received a matching video for 79.6% of those interests later the same day. For 82.6% of eligible users holding a declared interest, content attributed to declared interests made up at least 5% of their recommendations, so the declared-interest load is prevalent rather than token for the large majority of users. An independent multimodal LLM judge rated 87.2% of sampled interest-matched impressions as on-topic for the interest they were served against, over approximately 10,500 impressions drawn from 8,000 video–interest pairs per run.
These system-level measures (Section B of Table 1) answer a different question than the component studies that follow. Responsiveness reports the absolute rate at which a declaration is honored, rather than a relative change against an arm with the component removed, and topicality is scored independently of the serving gate.
Reception by attribution status. Among viewers holding at least one declared interest, recommendations attributed to a declared interest were liked at 11.55% against 9.90% for other recommendations shown to that cohort, a difference of 1.65 percentage points, and the ratio of positive to negative content-control taps was 6.44 against 2.95 (2.2 times). The attributed subset drew 1.6 percentage points more three-second skips and 10% less dwell per impression. These are descriptive cohort-level differences between delivered content subsets (Table 2).
| Metric | Observed comparison |
| Same-surface reference: attributed vs. other recommended | |
| Like rate | 11.55% attributed vs. 9.90% other recommended; +1.65 pp |
| Positive-to-negative content-control tap ratio | 6.44 attributed vs. 2.95 other recommended; |
| Three-second skip rate | +1.6 pp on attributed recommendations |
| Dwell per impression | −10% on attributed recommendations |
| Cross-surface stated-intent reference: attributed vs. search-result videos | |
| Like rate | on attributed recommendations |
| Three-second skip rate | −8.4 pp on attributed recommendations |
| Dwell per impression | +26% on attributed recommendations |
Table 2. Descriptive reception against recommendation and stated-intent references. Among viewers holding at least one declared interest, the first block compares attributed and other recommendations on Instagram Reels. The second uses videos viewed through Instagram Search as a cross-surface stated-intent reference.
Attributed recommendations receive more explicit positive feedback, while their higher skip rate and lower dwell show mixed overall reception.
Component studies isolate supply, competition, and continuity
A five-arm stack ablation isolates each serving layer by removing them cumulatively from the production configuration (Table 3). With no layer enabled, a substantial share of newly declared interests still receives a matching video the same day, because the recommender sometimes serves on-topic content without being directed to; enabling the full stack adds 33.5 percentage points on top of that baseline. Interest-conditioned candidate sourcing contributes the largest single share (13.7 pp), followed by forced top-slot insertion (11.1 pp) and added-interest ranking boosts (8.7 pp). Supply, placement, and competition therefore each contribute materially in the tested removal order, and no single layer accounts for a majority of the effect. Negative filtering and suppression produced no measurable change on this endpoint, which is consistent with their role: they remove unwanted content rather than deliver requested content.
Negative edits use a separate randomized study. Similarity-aware demotion extends the production filter to ordinary recommendations and reduces reappearance of suppressed interests (Table 3; Appendix § 13).
Separately, replenishment increased the share of active interests receiving a match watched for at least three seconds by 8.34% [8.07%, 8.61%] relative to the prior age-limited policy; Table 1 and Appendix § 13 report the companion responsiveness result.
Deployment and Efficiency
CLEAR components have operated in production for more than one year; the current system spans four surfaces across two platforms, Instagram and Facebook. History compression uses 39.5% fewer prompt tokens, history budgeting reduces estimated accelerator need by 38.9%, shared annotation removes approximately 98% of logical annotation requests, and the updated annotation pipeline raises throughput per accelerator by 145.5% (Table A5). These are separate levers with separate accounting baselines.
Incremental merging modestly improved whole-profile precision, recall, and F1 over full regeneration in the 1,000-user-per-condition offline comparison (); the incremental-inference capacity estimate separately requires 35% fewer accelerators than the preceding history-budget configuration (Table A3). A separate historical maximum-history input benchmark reports approximately 216 times the normalized model-inference cost for Gemini 2.5 Pro relative to its 3B reference (Table A4; Figure 4). Parallelizing request-time attribution with ranking reduced chunk-0 server latency by approximately 5 ms (Table A5).
| Model | F1@all | F1@10 | Relative inference cost |
| Specialized 3B | 0.5493 | 0.3967 | 1× |
| Gemini 2.5 Pro | 0.7051 | 0.3633 | 216× |
Table A4. F1@all, F1@10, and inference cost. Costs use the historical common maximum-history input basis with the 3B cost reference normalized to 1×.
Figure 4. Model inference cost and quality. Blue points use the left F1@all axis. Specialized 3B and Gemini 2.5 Pro connect vertically to their F1@10 values on the parallel right axis. Horizontal guides terminate at the right-axis values. Cost uses the shared log-scale x-axis.
Discussion
A shared interest identity aligns transparency and control. The central design choice is not to expose the recommender’s entire internal state. CLEAR exposes a ranked subset of an authoritative interest layer and makes the same normalized identities available to serving. A person can inspect an inference, an explicit correction takes precedence, and later candidates and impressions remain attributable to the corrected state.
Grounding personalizes more than the profile. Behavioral support filters inferred interests, but its recommendation-side role is equally important. Support events place retrieval in the user’s own co-engagement region, profile-conditioned rewrites personalize lexical search, conjunction-based admission combines behavioral proximity with topical evidence, and stable identity lets ordinary and dedicated delivery share one exposure account. A novel declaration begins without claiming past support, then uses the same address as evidence and organic supply accumulate.
Steerability uses supply, placement, and competition. The tested removal sequence assigns material contributions to sourcing (13.7 pp), reserved placement (11.1 pp), and ranking competition (8.7 pp). Separately, replenishment raises the share of active interests receiving a match watched for at least three seconds by 8.34% [8.07%, 8.61%] (Table 1).
Exposure accounting limits over-service of a single interest. The controller exists because naively promoting every topical match concentrates delivery on one interest. A six-day user-randomized comparison of real-time per-interest accounting and recalibrated volume dampening against the prior daily-accounting configuration reduced the recorded flooding-days metric by 15.4% [13.4%, 17.4%] (Table 1).
Reception relative to explicit search. Table 2 contrasts search and same-surface references: dwell and skip favor declared-interest content against search, but favor the cohort’s other recommendations within the surface.
Operational lessons from deployment. The production system separates four responsibilities that are easy to conflate: candidate supply, placement, competition in ranking, and continuity through replenishment. The stack ablation isolates the first three. The two-speed architecture also separates immediate control from ordinary recommender learning: the controller can act on an accepted interest without retraining the base model, while later user responses improve organic personalization through the existing learning loop. Topical gating and periodic grounding keep candidate feedback tied to the same stable interest identity across release cycles and surfaces.
Conclusion
CLEAR makes natural-language interests inspectable and operational. Grounding supplies support and retrieval anchors; a shared identity carries edits through serving and exposure control. Direct review establishes displayed-interest accuracy, while randomized studies establish actuation and continuity. People can inspect, correct, and direct an existing recommender.
Appendix
Supporting Evaluations and Results
Tables A1–A3 report profile and grounding evaluations; Tables A4–A5 report cost and deployment results. Tables A6–A7 define state and study contracts. Appendix B details serving experiments, and Appendix C shows the controller and grounding diagrams (Figure 5–Figure 6). Each table defines its own comparison or accounting baseline; studies with different references, populations, or costs are not combined.
Figure 5. Closed-loop delivery control. A topical source merges with ordinary candidates. Stable attribution lets recent exposure request or dampen topical supply; user responses return through ordinary history and model updates.
Why corrections are necessary
The rejection categories distinguish behavioral exposure from interests a person wants the system to retain. They are shares of rejected displayed entries, not shares of all displayed interests.
Grounding quality and coverage
The support-count filter and the event matcher evaluate different stages. The first scores retained profiles against an LLM-generated reference. The second judges whether retrieved history events support an interest and measures how much history the support sets cover.
Incremental profile updates
Incremental merging incorporates new activity into the retained profile instead of regenerating from the full history. The capacity estimate uses the same deployed incremental-inference configuration.
F1@all, F1@10, and inference cost
The comparison reports F1@all and mean-user F1@10 together. The separate historical cost benchmark uses a common maximum-history input basis and a 3B reference normalized to 1×. It accounts for model inference and excludes retrieval, annotation, and recommendation serving.
Deployment efficiency
These rows evaluate distinct efficiency levers and are not multiplied together. Capacity entries are estimates; token, request, and throughput entries follow their stated accounting baselines.
Interest-state contract
Every surface maps its user-facing actions to the same normalized interest identity and explicit-state precedence. The labels shown in an interface may differ, but the stored states and serving effects below are shared.
Experiment contracts
The main tables report effect sizes; Table A7 collects the corresponding design and analysis contracts in one place. Details are stated at the level needed to interpret each comparison.
Governance and state isolation
Sensitivity and integrity annotations attach once to each normalized interest string and are applied to both inferred and user-entered interests. When a user-entered interest is rejected by these gates, the editing interface displays an error. Shared annotations contain no user-specific membership, behavioral support, or preference state. Those fields remain in per-user serving records, while explicit preference state is stored separately from model output and retains precedence across refreshes. This separation lets the system reuse policy annotations without sharing a person’s interest state.
Serving Experiment Details
Serving-layer stack ablation
A five-arm user-randomized experiment removes serving layers cumulatively from Instagram’s production configuration. The first arm disables the early- and late-stage added-interest ranking terms. The second additionally disables reserved top-slot insertion. The third additionally disables negative filtering, explicit suppression, and single-topic delivery-gap rules. The fourth additionally sets interest-conditioned candidate source weights to zero, leaving no layer enabled. Differences between consecutive arms give the marginal delivery loss at each removal step; the directly estimated outer-arm contrast compares all layers removed with the production control.
Eligible positive additions must have sufficient post-edit viewing opportunity. The primary endpoint counts whether at least one automatically attributed video appears on the day of the edit; a secondary endpoint counts whether one appears within the first 16 qualifying video impressions. All reported randomized serving-component comparisons have . On the secondary endpoint the three delivery layers each account for 12–13 percentage points. Table 3 reports the per-layer losses on the primary endpoint.
Two limits apply. The ablated arms are not production configurations, so their rates are not production rates. The control arm is measured under a narrower endpoint definition than the same-day responsiveness in Table 1 and approximates rather than reproduces it.
Suppression-aware ranking
A separate user-randomized experiment compares the production post-retrieval cluster-ID filter with the same filter plus similarity-aware demotion of ordinary recommendations. Its endpoint is the share of suppressed interests that still appear in recommended videos at each horizon; lower is better. Table A7 reports the same-day and longer-horizon effects.
Replenishment
The comparison covers seven days. The engaged-interest metric is the ratio of interests seen with at least three seconds of dwell to active interests; its companion counts distinct days on which a user received at least one explicit-interest impression. The experiment unit is the Instagram user.
The replenishment treatment removes a fixed sourcing-age cutoff and replaces it with a dynamic selection policy. Among interests eligible for offline sourcing, those past a cold-start window are admitted only above a recent-view threshold; the most recently declared interests retain reserved capacity, and the remaining slots are filled by least recent realized exposure rather than at random. The control retains the prior age-limited policy. Replenishment increases the share of active interests receiving a match watched for at least three seconds by 8.34% [8.07%, 8.61%] and user days with explicit-interest delivery by 2.43% [2.24%, 2.63%].
Serving and Steering Detail
References
- [1]Rajat Arora, Ye Tao, Jianqiang Shen, Ping Liu, Muchen Wu, Qianqi Shen, Benjamin Le, Fedor Borisyuk, Jingwei Wu, and Wenjing Zhang. High fidelity textual user representation over heterogeneous sources via reinforcement learning, 2026. arXiv:2602.07333.arXiv
- [2]Micah Carroll, Adeline Foote, Marcus Williams, Anca Dragan, W. Bradley Knox, and Smitha Milli. CTRL-Rec: Controlling recommender systems with natural language. In The Thirteenth International Conference on Learning Representations, 2025. https://openreview.net/forum?id=tXl1gdoAoV.openreview.net/forum?id=tXl1gdoAoV
- [3]Iordanis Fostiropoulos, Muhammad Rafay Azhar, Abdalaziz Sawwan, Boyu Fang, Yuchen Liu, Jiayi Liu, Hanchao Yu, Qi Guo, Jianyu Wang, Fei Liu, and Xiangjun Fan. GISTBench: Evaluating LLM user understanding via evidence-based interest verification, 2026. https://arxiv.org/abs/2603.29112.
- [5]Jiahao Liu, Yiyang Shao, Peng Zhang, Dongsheng Li, Hansu Gu, Chao Chen, Longzhi Du, Tun Lu, and Ning Gu. Filtering discomforting recommendations with large language models. In Proceedings of The Web Conference, pages 3639–3650, 2025. doi: 10.1145/3696410.3714850. arXiv:2410.05411.DOI
- [6]Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen-tau Yih, Pang Wei Koh, Mohit Iyyer, Luke Zettlemoyer, and Hannaneh Hajishirzi. FActScore: Fine-grained atomic evaluation of factual precision in long form text generation. In Proceedings of EMNLP, pages 12076–12100, 2023. doi: 10.18653/v1/2023.emnlp-main.741. arXiv:2305.14251.DOI
- [7]Sheshera Mysore, Mahmood Jasim, Andrew McCallum, and Hamed Zamani. Editable user profiles for controllable text recommendations. In Proceedings of SIGIR, pages 993–1003, 2023. doi: 10.1145/3539618.3591677. arXiv:2304.04250.DOI
- [8]Emiliano Penaloza, Olivier Gouvert, Haolun Wu, and Laurent Charlin. TEARS: Textual representations for scrutable recommendations, 2024. arXiv:2410.19302.arXiv
- [9]Filip Radlinski, Krisztian Balog, Fernando Diaz, Lucas Dixon, and Ben Wedin. On natural language user profiles for transparent and scrutable recommendation. In Proceedings of SIGIR, 2022. doi: 10.1145/3477495.3531873. arXiv:2205.09403.DOI
- [10]Jerome Ramos, Hossein A. Rahmani, Xi Wang, Xiao Fu, and Aldo Lipani. Transparent and scrutable recommendations using natural language user profiles. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 13971–13984, 2024. doi: 10.18653/v1/2024.acl-long.753. arXiv:2402.05810.DOI
- [11]Hannah Rashkin, Vitaly Nikolaev, Matthew Lamm, Lora Aroyo, Michael Collins, Dipanjan Das, Slav Petrov, Gaurav Singh Tomar, Iulia Turc, and David Reitter. Measuring attribution in natural language generation models. Computational Linguistics, 49(4):777–840, 2023. doi: 10.1162/coli_a_00486. arXiv:2112.12870.DOI
- [12]Team, R. (2025). RecGPT technical report. 2507.22879
- [13]Ricks, B., & McCrosky, J. (2022). Does this button work? Investigating YouTube’s ineffective user controls. In Mozilla Foundation. <a href="https://doi.org/https://assets.mofoprod.net/network/documents/Mozilla-Report-YouTube-User-Controls.pdf" rel="noopener noreferrer" target="_blank">https://doi.org/https://assets.mofoprod.net/network/documents/Mozilla-Report-YouTube-User-Controls.pdf</a>
- [14]Ruixuan Sun, Junyuan Wang, Sanjali Roy, and Joseph A. Konstan. Co-authoring the self: A human–AI interface for interest reflection in recommenders, 2025. arXiv:2510.08930.arXiv
- [16]Haoting Wang, Haokai Lu, Zheyun Feng, Jenny Huang, Yifat Amir, Gregory Hinkson, Ben Most, Zelong Zhao, Yixin Kelly Cui, Rein Zhang, Fabio Soldo, Yu Xia, Nihar Bhupalam, Minmin Chen, Konstantina Christakopoulou, Lichan Hong, and Ed H. Chi. LLM-based user personas for recommendations at scale. In Proceedings of the 20th ACM Conference on Recommender Systems, 2026. doi: 10.1145/3773078.3831903. arXiv:2606.12198.DOI
- [17]Mengke Wu, Weizi Liu, Yanyun Wang, Weiyu Ding, and Mike Yao. Rethinking user empowerment in AI recommender system: Innovating transparent and controllable interfaces. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems, 2026. doi: 10.1145/3772318.3791914. arXiv:2509.11098.DOI
- [18]Ziyun Xu, Bosen Ding, Yue Zhang, Ji Qi, Qingyuan Song, Jizhou Huang, Liwei Wang, Jefferey Santelli, Yue Weng, Qichao Que, Zhenheng Yang, Junfeng Pan, and Linhong Zhu. Shape your feed: An LLM-based agentic system for conversational recommendation. In Proceedings of the 20th ACM Conference on Recommender Systems, 2026. doi: 10.1145/3773078.3831877. arXiv:2608.06632.DOI
- [19]Zhao, Z., Zhou, X., Lv, X., Li, Y., He, C., Liu, Z., Shen, J., Liu, Y., Li, H., Zhang, C., Liang, J., Li, M., Gao, C., Feng, F., Tang, R., & Li, H. (2026). RECAP: Feedback-driven streaming semantic user profiles for short-video recommendation. Proceedings of the 20th ACM Conference on Recommender Systems, 1427–1436. https://doi.org/10.1145/3773078.3831930
- [20]Joyce Zhou, Weijie Zhou, Doug Turnbull, and Thorsten Joachims. SteerEval: A framework for evaluating steerability with natural language profiles for recommendation, 2026. arXiv:2601.21105.arxiv.org/abs/2601.21105
- [21]Chenxu Zhu, Shigang Quan, Bo Chen, Jianghao Lin, Xiaoling Cai, Hong Zhu, Xiangyang Li, Yunjia Xi, Weinan Zhang, and Ruiming Tang. LIBER: Lifelong user behavior modeling based on large language models, 2024. arXiv:2411.14713.arxiv.org/abs/2411.14713





