TAGR: Temporally Adaptive Generative Recommendation for Industrial Live-Streaming Advertising
TAGR introduces a temporally adaptive generative recommender with LSID, IAG, and IOPO, boosting live-stream ad engagement by 8.5%.
Key Findings
Methodology
TAGR comprises three core modules: LSID dynamically updates ad identifiers based on scene and product features via hierarchical encoding; IAG models multi-scale user intent using multi-behavior signals and long-term profiles with autoregressive transformers; IOPO periodically samples current policy candidates for preference alignment, integrating on-policy feedback with supervised training. The system employs contrastive learning, multi-task optimization, and weighted next-token prediction (MF-NTP), leveraging real user data from large-scale e-commerce platforms. The approach ensures high responsiveness to scene and intent shifts, maintaining recommendation relevance and stability in dynamic live environments.
Key Results
- Deployments on a major e-commerce live-stream platform showed an 8.5% increase in live-room entry rate, 7.4% rise in shopping cart clicks, and a 16.1% revenue uplift compared to baseline models. Offline evaluations confirmed superior cold-start coverage and long-tail performance, especially under rapid content change scenarios.
- Ablation studies demonstrated that removing LSID, IAG, or IOPO reduced performance by at least 5%, confirming their synergistic importance. The model maintained stable, high-quality recommendations across multiple live scenarios, validating the effectiveness of temporal adaptation.
- Further analysis indicated that the multi-scale intent encoding captured both immediate and stable user preferences, while the preference optimization balanced freshness with behavioral preservation, leading to consistent improvements over traditional static-ID approaches.
Significance
This work addresses critical challenges in live-stream advertising, where rapid content and user behavior changes demand models that adapt in real-time. By integrating dynamic semantic identifiers, multi-scale intent modeling, and on-policy preference alignment, TAGR advances the state-of-the-art in industrial recommendation systems. Its successful deployment demonstrates the practical viability of temporal adaptive generative models, paving the way for more personalized, timely, and effective advertising solutions in fast-paced digital environments. The framework also offers insights into combining offline training with online policy updates, contributing to the broader field of reinforcement learning and dynamic recommendation.
Technical Contribution
The paper introduces a novel hierarchical encoding scheme (LSID) for dynamic ad identification, combined with multi-scale intent modeling (IAG) that captures diverse behavioral signals at request time. The intermittent on-policy preference optimization (IOPO) mechanism innovatively balances policy feedback freshness with behavioral stability, employing periodic sampling and joint supervised training. These components collectively improve the model’s sensitivity to temporal changes, enhance recommendation relevance, and ensure training stability, representing a significant technical leap over static-ID and single-scale models. The integration of contrastive learning, weighted supervision, and on-policy updates forms a comprehensive framework for real-time, high-quality recommendations.
Novelty
This research is the first to systematically incorporate multi-level temporal adaptation into a generative recommendation framework for live-stream advertising. Unlike prior work focusing on static IDs or single-scale behavior modeling, TAGR’s hierarchical, request-aware design enables real-time tracking of evolving ad targets and user intent. Its innovative combination of dynamic semantic IDs, multi-scale intent encoding, and periodic preference alignment sets a new benchmark for industrial recommendation systems operating under high-frequency content changes, marking a significant step forward in the field.
Limitations
- While effective in large-scale e-commerce settings, the model’s performance under extremely volatile or unpredictable scenarios (e.g., sudden viral content) remains uncertain. The dynamic updates and complex training pipeline increase computational costs, potentially limiting deployment in latency-sensitive environments. The reliance on extensive behavioral data may pose challenges in cold-start or sparse-data contexts, and further research is needed to enhance robustness and efficiency in such cases.
Future Work
Future directions include integrating multi-modal signals such as visual and audio cues to enrich intent understanding, developing more efficient online update mechanisms to reduce latency, and exploring reinforcement learning strategies for autonomous policy adaptation. Additionally, extending the framework to other dynamic recommendation domains like social media feeds or personalized news could further demonstrate its versatility and impact.
AI Executive Summary
Live-stream advertising has become a dominant monetization strategy on short-video and e-commerce platforms, driven by the need for highly relevant and timely recommendations. Traditional models, which rely on static identifiers and single-scale user behavior, struggle to keep pace with the rapid evolution of live scenes, promoted products, and user intent. This gap hampers user engagement and revenue generation, especially in fast-changing environments where content and preferences shift within minutes.
To address these challenges, the paper introduces TAGR, a novel generative recommendation framework designed with three levels of temporal adaptation. The first component, LSID, dynamically updates ad identifiers based on scene and product features, employing hierarchical encoding and contrastive learning to maintain a stable yet responsive representation. The second component, IAG, models user intent across multiple temporal scales using multi-behavior signals, auxiliary actions, and long-term profiles, enabling the system to capture both immediate and persistent preferences. The third component, IOPO, periodically samples current policy candidates and performs preference updates aligned with real-time feedback, balancing exploration and stability.
Experimental results from deployment on a large-scale e-commerce platform demonstrate significant improvements: an 8.5% increase in live-room entry rate, a 7.4% rise in shopping cart clicks, and a 16.1% revenue lift over the baseline. Offline evaluations confirm enhanced cold-start performance and long-tail coverage, validating the effectiveness of the time-sensitive adaptive design. These findings highlight the potential of integrating dynamic semantic identifiers, multi-scale intent modeling, and on-policy preference optimization in industrial recommendation systems.
Overall, TAGR advances the state-of-the-art in live-stream advertising by providing a scalable, effective, and industry-ready solution for real-time, personalized content delivery. Its success paves the way for future research into multi-modal, reinforcement learning-based, and cross-domain adaptive recommendation frameworks, promising richer user experiences and higher commercial returns.
Deep Dive
Limitations & Outlook
What gaps remain?
Abstract
Live-streaming advertising is an important monetization channel on short-video and e-commerce platforms, where rapidly changing live content, promoted products, and user feedback impose strong freshness requirements on recommendation models. Existing generative recommenders designed for static domains fail at three levels: static semantic IDs (SID) cannot track evolving live ads; single-scale behavior modeling misses shifting intent; preference optimization conflicts between fresh on-policy feedback and training stability. We propose TAGR, a generative recommendation framework with temporal adaptation at three levels: live-ad tokenization, user intent modeling, and preference alignment. At the token level, Live Semantic-Collaborative ID (LSID) periodically refreshes each active ad's SID based on its current live scene and promoted products, while retaining a stable hierarchical token vocabulary for autoregressive generation. At the intent level, Intent-Aware Generation (IAG) models live-room entry histories at multiple temporal granularities as the primary intent sequence, keeps auxiliary behaviors as separate inputs, and weights next-token prediction (NTP) using post-request intent evidence and business value. At the alignment level, Intermittent On-Policy Preference Optimization (IOPO) periodically samples fresh candidate groups from the current policy and performs behavior- and value-aligned preference updates interleaved with supervised NTP maintenance to preserve learned behavior distribution. Deployed on a large-scale e-commerce live-stream advertising platform, TAGR improves live-room entry and shopping-cart click rates by 8.5% and 7.4%, respectively, and achieves a 16.1% revenue lift over the production baseline. These results demonstrate the effectiveness and industrial viability of temporally adaptive generative recommendation for live-stream advertising.