Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations (Feb 2024, arXiv)
Metadata
| Field | Content | |------|------| | Title | Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations | | Authors / Affiliation | Jiaqi Zhai, Lucy Liao, Xing Liu, Yueming Wang, Rui Li, Xuan Cao, et al. (12 authors total) | | Published | February 2024 | | Source | https://arxiv.org/abs/2402.17152 | | Type | Academic paper | | Section | Recommender Engines, Sequential Recommendation |
One-line Summary
This work proposes a generative recommendation approach based on trillion-parameter sequential transducers, reformulating recommendation as sequence transduction and demonstrating orders-of-magnitude efficiency gains and strong online results on an industrial deployment.
Background and Motivation
Large-scale search, recommendation, and personalization systems have long faced challenges in efficiency, scalability, and user-intent understanding. Traditional pipelines separate retrieval, ranking, and generation, which makes them hard to adapt to the LLM era's demands for natural-language interaction, multi-hop reasoning, and real-time knowledge. This paper was proposed in this context, aiming to advance the boundary between generative modeling and industrial recommendation systems.
The paper's core premise: in generative recommendations, actions speak louder than words — modeling a user's sequence of actions (clicks, likes, comments, etc.) with a purpose-built architecture can outperform transplanting language-model approaches built primarily around words/tokens.
Core Contributions
- Introduces a unified generative view of recommendation tasks, casting ranking and retrieval as sequential transduction over user action sequences.
- Proposes a new attention-based architecture (HSTU) designed for high-cardinality, non-stationary recommendation data, replacing the standard multi-stage cascade with an end-to-end generative formulation.
- Scales to trillion parameters while scaling characteristics reportedly exceed comparable Transformer-based LLMs, with an order-of-magnitude improvement in training efficiency.
- Demonstrates that despite the scale, inference remains practical, and deployed models delivered significant online metric improvements on a large-scale content recommendation platform.
- Discusses interfaces with emerging paradigms such as LLM tool use and reinforcement learning, and outlines open problems including latency/cost constraints, evaluation trustworthiness, and safety.
- Setup: offline comparisons against Transformer/LLM baselines and production systems, plus online A/B deployment;
- Metrics: ranking quality metrics, training efficiency, inference latency, and online engagement signals;
- Key claim: scaling laws for HSTU-based generative recommenders outperform those of Transformer LLM baselines, with substantial online gains in production.
- 360Brew: A Decoder-only Foundation Model for Personalized Ranking
- Augmenting Netflix Search with In-Session Adapted Recommendations
- Bridging Language and Items for Retrieval and Recommendation, Mar 2024
- Data-efficient Fine-tuning for LLM-based Recommendation, SIGIR 2024
- DiffKG: Knowledge Graph Diffusion Model for Recommendation, WSDM 2024
- EAGER-LLM: Enhancing Large Language Models as Recommenders
- Original paper: "Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations." arXiv:2402.17152, February 2024. https://arxiv.org/abs/2402.17152
Method / System Architecture
The work generally follows a "problem formalization → model/system design → training pipeline → inference pipeline" structure:
1. Input & representation: user histories and contextual signals are encoded as sequences of actions (tokens), including time-based encodings to capture non-stationarity; 2. Core modules: hierarchical transduction blocks that mix pointwise and pairwise attention over the action vocabulary; 3. Learning strategy: large-scale supervised training on action sequences, with scaling laws characterizing performance vs. parameters/data; 4. Inference: efficient serving of the trillion-parameter model, substituting the traditional candidate-generation + ranking cascade with sequence generation/transduction.
Note: exact quantitative results should be verified against the original PDF; this digest is based on the abstract and public metadata.
Evaluation
Typical evaluation dimensions covered (see the paper for exact numbers):
Key Takeaways for Search / Rec / Personalization
1. Architecture: cascaded retrieve-rank-generate remains common, but end-to-end generative recommenders with purpose-built transducer architectures are a credible alternative to porting LLMs directly; 2. Data: user action sequences are as critical as instruction data; domain-specific scaling behavior differs from language modeling; 3. Evaluation: offline metrics and online satisfaction diverge at scale — production A/B evidence matters; 4. Product: latency, cost, explainability, and safety remain hard constraints for industrial deployment and cannot be ignored in favor of academic benchmarks.
Limitations and Future Work
Potential limitations include experiment scale bounded by compute budgets, benchmark-to-real-user-distribution mismatch, and unknown cross-lingual generalization. Future directions may include more efficient test-time compute allocation, deeper integration with knowledge graphs and structured databases, and causal/fairness constraints for generative recommenders.