Source-linked AI summary
InitGen: Candidate Generation for Interaction Initiation in Intelligent Assistants
Ruize Shi, Jinhua Chen, Hong Huang, Ziniu Chen, Ruike Zhang, Jianxun Shi, Yitao Chen, Rui Zhang
TL;DR
Interaction initiation must generate multiple open-text candidates before session intent is known, while operating under strict latency and learning from feedback observed only after downstream selection. InitGen jointly generates query sets and aligns them with weighted preference optimization using user activity and ranking scores, with rolling updates from recent data. In a production A/B test, it raised CTR from 0.95% to 1.61% and exposure by 17.9% while generating each set within 180 ms.
Problem
Interaction initiation lacks an expressed session intent and fixed candidate pool, while feedback is observed only for downstream-exposed candidates, creating a distinct candidate-generation setting.
Method
InitGen jointly generates query sets and aligns aggregated click feedback through weighted preference optimization using user activity, ranking scores, and rolling recent-data updates.
Results
69.1% relative CTR improvement from 0.95% to 1.61% and 17.9% higher query exposure were achieved in a production A/B test, with each candidate set generated within 180 ms.
Takeaways & Limitations
InitGen is fully deployed in Xiaobu Assistant and improves engagement and candidate exposure under the production latency requirement.
Abstract
from arXiv · showhide
Interaction initiation refers to presenting multiple candidate queries when a user opens an intelligent assistant before expressing any intent for the current session. In production, candidate generation incorporates dynamic context and produces all candidates within a strict latency budget. Learning from user feedback is also difficult since the generator usually produces more candidates than are finally displayed. After downstream filtering and ranking, only a subset is exposed to users, so the observed feedback is partial and cannot be reliably assigned to individual queries. We present InitGen, a framework for candidate generation that is deployed in the interaction initiation pipeline of OPPO's Xiaobu Assistant. InitGen generates a set of candidate queries jointly and aligns the generated set with user feedback through weighted preference optimization. The sample weights are derived from user activity and downstream ranking scores. The activity weight reduces the dominance of highly active users during training, while the ranking score is used as a practical estimate of the reliability of the observed feedback. InitGen also uses a rolling window update strategy to incorporate recent interaction data into periodic model updates. In an online A/B test against a strong production baseline, InitGen improves the click-through rate from 0.95% to 1.61%, corresponding to a relative improvement of 69.1%, and increases query exposure by 17.9% under the same traffic allocation. InitGen generates the complete candidate set within 180 ms and has been fully deployed in OPPO's Xiaobu Assistant, which serves over 150 million monthly active users.
Introduction
Interaction initiation presents candidate queries when an assistant opens before the user expresses session intent, under strict latency and partial-feedback constraints. InitGen jointly generates and aligns query sets, incorporates recent data, and improves engagement and exposure in production.
- Introduction: Interaction initiation generates multiple open-text candidates before session intent is expressed, then receives feedback only for candidates selected and exposed by downstream modules.This setting differs from approaches using observed queries, fixed candidate pools, or feedback matched to every output.
- Introduction: Unknown intent, a 180 ms latency requirement, and partial delayed feedback make coordinated generation and reliable query-level learning difficult.Independent generation can produce repeated or highly similar candidates, while downstream processing prevents reliable assignment of feedback to every generated query.
- Introduction: InitGen jointly generates a query set, aligns aggregated feedback with user-activity and ranking-score weights, and periodically updates the model using a rolling window of recent interactions.The activity weight limits dominance by highly active users, while the ranking score estimates observed-feedback reliability; updates occur offline.
- Introduction: 69.1% relative CTR improvement from 0.95% to 1.61% and 17.9% higher exposure were achieved against a standard KTO production baseline, with generation within 180 ms.InitGen was deployed in Xiaobu Assistant, serving over 150 million monthly active users.
The Proposed Method
InitGen jointly generates a coordinated candidate-query set from historical interactions and context, then aligns that set with partial downstream feedback under production constraints. Its method combines query-set alignment, weighted preference signals, and rolling updates for recent data.
- Joint Query Set Generation: Joint generation produces the complete candidate set in one request, allowing coordination among queries while meeting the 180 ms latency requirement.The candidates are treated as a set downstream even though they are serialized for autoregressive generation.
- Query Set Alignment: Query-set alignment assigns one binary outcome to each retained generated set, avoiding unreliable individual-query labels and matching the generation unit.Sets with no exposed queries are excluded, while a set is positive if at least one exposed query is clicked.
- Weighted Preference Signals: User activity weighting reduces the disproportionate influence of highly active groups by increasing weights for lower-CTR groups and decreasing them for higher-CTR groups.The weighting keeps each record’s contribution bounded while correcting activity-related imbalance.
- Weighted Preference Signals: InitGen aggregates exposed-query ranking scores into one mean score per generated set, so each set contributes once to the alignment objective.This uses downstream ranking information without assigning separate alignment samples according to the number of exposed queries.
- Rolling Update Strategy: Periodic weighted alignment over a rolling window incorporates recent interaction data while remaining compatible with production serving requirements.The model is initialized with SFT, then updated using recent records as the window advances.
Experiments
InitGen is evaluated through controlled online A/B tests against a production baseline and staged component ablations. It improves both engagement and candidate exposure while maintaining the production latency target.
- Online Study: CTR rises from 0.95% to 1.61%, a relative lift of 69.1%, while query exposure increases 17.9% under identical traffic allocation.The unchanged downstream pipeline means the exposure gain reflects more generated candidates reaching the interface, while the larger CTR gain indicates stronger engagement with displayed candidates.
- Online Study: InitGen outperforms the production baseline on every day of the one-week experiment, including days with relatively high baseline CTR.The daily improvement remains positive across varying traffic conditions.
- Ablation Study: User-activity weighting raises lower-response users’ CTR from 0.78% to 0.85%, a 9.0% relative gain, while the high-response cohort remains near 2.0%.The overall CTR rises from 1.06% to 1.12%, consistent with reducing the dominance of highly active users during alignment.
- Ablation Study: Ranking-score weighting increases query exposure by 9.7% to 28.2% across every day of the one-week experiment.The weighting emphasizes preference records with stronger estimated confidence from the downstream model.
- Online Study: Generation latency remains concentrated around 150 ms and below the 180 ms budget despite workload increasing from about 1K to over 12K QPM.The deployed serving configuration accommodates daily traffic variation without a sustained latency increase at peak load.
Related Work
Interaction initiation overlaps with proactive interaction, recommendation, and LLM alignment, but differs in its available information, open-text outputs, and post-deployment feedback. InitGen addresses these differences by learning from clicks on only the candidates that downstream processing exposes under a tighter latency budget.
- Proactive Interaction and Recommendation: Unlike query suggestion and proactive guidance, interaction initiation begins without an observed query, prefix, dialogue context, or expressed session intent.Recommendation settings commonly select from predefined items or actions, whereas InitGen generates open-text candidates.
- Proactive Interaction and Recommendation: InitGen learns directly from observed click outcomes when downstream filtering exposes only a subset of generated candidates.This distinguishes it from concurrent work that estimates user interaction to construct preference pairs.
- LLM Alignment: InitGen relates to LLM alignment methods such as RLHF, DPO, and SimPO while applying preference alignment to candidate-query generation.The related alignment literature includes reward-model reinforcement learning and newer direct preference objectives.
Conclusion
InitGen jointly generates query sets, aligns them with aggregated click outcomes, and incorporates recent interaction data through rolling-window updates.
- InitGen jointly generates a complete query set in one generation request and aligns it with aggregated click outcomes through WKTO.The framework targets interaction initiation when session intent is unknown and feedback appears only after downstream processing.
- Periodic updates with a rolling data window incorporate recent interaction data into the candidate-generation framework.
- The framework is evaluated in a production A/B test on OPPO’s Xiaobu Assistant.
A.1 Data Privacy
InitGen’s training and evaluation data were anonymized, use coarse geographic signals and high-level profile attributes, and support user-controlled personalization.
- Training and evaluation data were anonymized and contained no direct personal identifiers.
- Geographic personalization uses coarse city-level information, while profile attributes are high-level interest descriptions refreshed periodically.Users can disable personalization through product settings.
A.2 Training and Deployment Lessons
Deployment lessons emphasize selecting a backbone that satisfies the latency budget and using the query set as the common unit for supervised initialization and preference alignment.
- Backbone selection: Qwen2.5-1.5B balances generation quality and serving efficiency, whereas Qwen3-1.7B frequently exceeds the 180 ms latency budget with thinking disabled.
- Backbone selection: Observed instability motivates keeping the query set as the shared output unit across supervised initialization and subsequent alignment.The passage links this design to a mismatch between supervised-initialization and preference-alignment units, potentially with limited model capacity.
- Balance of alignment outcomes: KTO can learn from independently labeled desirable and undesirable outputs without requiring equal numbers of the two outcomes.
A.3 Case Study
The case study shows that contextual signals reshape InitGen’s candidate queries: location improves situational relevance, while profile attributes add longer-term personalization without eliminating broader interests.
- Without contextual information, candidates mainly follow historical interactions, including prior-location associations and broad topics such as psychology, culture, travel, and photography.Some outputs are less suitable because they do not distinguish persistent interests from information tied to previous locations.
- Adding location shifts candidates toward Wuhan-related travel, weather, and air quality while retaining broader interests from interaction history.The location signal updates the situational part of the set without replacing other interests.
- Adding profile attributes introduces photography, AI photo editing, literary creation, and travel-writing queries, including combinations with Wuhan.These signals jointly balance contextual relevance, persistent interests, and diversity.
- Table A1 presents the contextual-signal case study, with textual content translated from Chinese into English.