Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Real-Time Trend Prediction via Continually-Aligned LLM Query Generation

Zijing Hui, Wenhan Lyu, Shusen Wang, Li Chen, Chu Wang

TL;DR

RTTP tackles the cold-start problem in real-time trend detection by generating synthetic search queries directly from new posts using a continually updated LLM (CL-LLM). It introduces Mix-Policy DPO to balance on-policy stability with off-policy novelty, mitigating catastrophic forgetting during continual updates. In production on Facebook and Meta AI platforms, RTTP achieves substantial gains in tail-trend precision (~+91.4% at ranking 500) and query-generation accuracy (~+19%), while maintaining stability over multi-week online training. The work demonstrates that aligned, continually updated synthetic signals can enable timely trend understanding in low-traffic search environments.

Abstract

Trending news detection in low-traffic search environments faces a fundamental cold-start problem, where a lack of query volume prevents systems from identifying emerging or long-tail trends. Existing methods relying on keyword frequency or query spikes are inherently slow and ineffective in these sparse settings, lagging behind real-world shifts in attention. We introduce RTTP, a novel Real-Time Trending Prediction framework that generates search queries directly from news content instead of waiting for users to issue them. RTTP leverages a continual learning LLM (CL-LLM) that converts posts into search-style queries and scores them using engagement strength + creator authority, enabling early trend surfacing before search volume forms. To ensure adaptation without degrading reasoning, we propose Mix-Policy DPO, a new preference-based continual learning approach that combines on-policy stability with off-policy novelty to mitigate catastrophic forgetting during model upgrades. Deployed at production scale on Facebook and Meta AI products, RTTP delivers +91.4% improvement in tail-trend detection precision@500 and +19% query generation accuracy over industry baselines, while sustaining stable performance after multi-week online training. This work demonstrates that LLM-generated synthetic search signals, when aligned and continually updated, unlock timely trend understanding in low-traffic search environments.

Real-Time Trend Prediction via Continually-Aligned LLM Query Generation

TL;DR

RTTP tackles the cold-start problem in real-time trend detection by generating synthetic search queries directly from new posts using a continually updated LLM (CL-LLM). It introduces Mix-Policy DPO to balance on-policy stability with off-policy novelty, mitigating catastrophic forgetting during continual updates. In production on Facebook and Meta AI platforms, RTTP achieves substantial gains in tail-trend precision (~+91.4% at ranking 500) and query-generation accuracy (~+19%), while maintaining stability over multi-week online training. The work demonstrates that aligned, continually updated synthetic signals can enable timely trend understanding in low-traffic search environments.

Abstract

Trending news detection in low-traffic search environments faces a fundamental cold-start problem, where a lack of query volume prevents systems from identifying emerging or long-tail trends. Existing methods relying on keyword frequency or query spikes are inherently slow and ineffective in these sparse settings, lagging behind real-world shifts in attention. We introduce RTTP, a novel Real-Time Trending Prediction framework that generates search queries directly from news content instead of waiting for users to issue them. RTTP leverages a continual learning LLM (CL-LLM) that converts posts into search-style queries and scores them using engagement strength + creator authority, enabling early trend surfacing before search volume forms. To ensure adaptation without degrading reasoning, we propose Mix-Policy DPO, a new preference-based continual learning approach that combines on-policy stability with off-policy novelty to mitigate catastrophic forgetting during model upgrades. Deployed at production scale on Facebook and Meta AI products, RTTP delivers +91.4% improvement in tail-trend detection precision@500 and +19% query generation accuracy over industry baselines, while sustaining stable performance after multi-week online training. This work demonstrates that LLM-generated synthetic search signals, when aligned and continually updated, unlock timely trend understanding in low-traffic search environments.
Paper Structure (18 sections, 2 equations, 4 figures)

This paper contains 18 sections, 2 equations, 4 figures.

Table of Contents

  1. Introduction
  2. Related Work
  3. Industrial Approaches to Trend Detection
  4. LLMs for Trend Understanding and Generation
  5. LLM Continual Learning
  6. RTTP System
  7. System Architecture
  8. CL-LLM
  9. Mix-Policy DPO
  10. Training flow
  11. Experiments
  12. Datasets
  13. Trend Prediction Evaluation
  14. Query Generation Accuracy Evaluation
  15. Accuracy Metrics Definition
  16. ...and 3 more sections

Figures (4)

  • Figure 1: CL-LLM continual training workflow.
  • Figure 2: Precision across ranking tiers for trend detection methods.
  • Figure 3: Query generation accuracy over 4 weeks. Mix-DPO maintains higher accuracy compared to SFT-based methods.
  • Figure 4: MMLU benchmark accuracy over a month for continual training. Mix-DPO maintains reasoning capability