[Paper] ReaSeq: Unleashing World Knowledge via Reasoning for Sequential Modeling

Source: arXiv - 2512.21257v1

Overview

ReaSeq is a new framework that injects the world knowledge stored in Large Language Models (LLMs) into industrial recommender systems. By combining explicit chain‑of‑thought reasoning with latent diffusion‑based inference, it tackles two long‑standing pain points: sparse, ID‑only item embeddings and the inability to surface interests that lie outside a platform’s historical logs.

Key Contributions

• Hybrid reasoning pipeline – mixes explicit multi‑agent Chain‑of‑Thought (CoT) reasoning to generate structured product semantics with implicit diffusion‑based LLM reasoning that imagines plausible user actions beyond recorded clicks.
• Semantic enrichment of item IDs – transforms raw item identifiers into dense, knowledge‑grounded vectors that capture attributes, usage contexts, and cross‑domain relations.
• Beyond‑log behavior generation – a diffusion LLM predicts “what a user might do next” even when no prior interaction exists, effectively widening the recommendation horizon.
• Large‑scale production validation – deployed on Taobao’s real‑time ranking pipeline serving hundreds of millions of users, delivering >6 % lift in click‑through‑rate (CTR) and impression‑per‑view (IPV), +2.9 % more orders, and +2.5 % growth in gross merchandise value (GMV).
• Multi‑agent collaboration design – introduces a lightweight coordination protocol that lets several specialized agents (knowledge extractor, semantic mapper, behavior generator) share intermediate reasoning steps without heavy model retraining.

Methodology

1. Data Ingestion – Existing interaction logs (user‑item clicks, purchases) are fed to a knowledge extraction agent.
2. Explicit CoT Reasoning
   • A set of prompts guides the LLM to break down each item into a hierarchy of attributes (category, material, style, usage scenario, etc.).
   • The multi‑agent system iteratively refines these attributes, producing a structured knowledge graph per item.
   • The graph is then embedded (e.g., via Graph Neural Networks) to create a semantic item vector that augments the traditional ID embedding.
3. Implicit Diffusion Reasoning
   • A diffusion‑based LLM (e.g., Diffusion‑GPT) is conditioned on the user’s short‑term session and the enriched item vectors.
   • It samples plausible future interactions that are not present in the log, effectively hallucinating “beyond‑log” interests while staying grounded by the semantic knowledge.
4. Fusion & Ranking
   • The original collaborative‑filtering scores, the semantic vectors, and the diffusion‑generated candidate items are merged in a lightweight ranking model (often a feed‑forward network).
   • Real‑time inference runs within the latency budget of Taobao’s ranking service.

The whole pipeline is modular: any LLM can be swapped in, and the reasoning steps are logged for interpretability and debugging.

Results & Findings

• Sparse items (≤5 historical interactions) saw the biggest CTR boost (~9 %), confirming that semantic enrichment mitigates ID‑poverty.
• Cold‑start users (new accounts) benefited from the diffusion‑generated candidates, with a 12 % increase in first‑day engagement.
• Ablation studies showed that removing either the explicit CoT or the diffusion component reduced overall lift by ~3 % each, indicating that both reasoning modes are complementary.

Practical Implications

• Improved cold‑start handling – Developers can plug ReaSeq’s semantic encoder into existing recommendation stacks to give new items a “knowledge boost” without waiting for interaction data.
• Cross‑domain recommendation – Because the item knowledge graph captures universal attributes (e.g., “outdoor sport”), the same embeddings can be reused across different product categories or even different platforms.
• Reduced reliance on massive logging – Companies with strict privacy constraints can still benefit from LLM‑derived world knowledge, lowering the volume of user‑level data needed for high‑quality rankings.
• Interpretability for product teams – The explicit CoT steps produce human‑readable attribute lists, making it easier to audit why a recommendation surfaced (useful for compliance and trust).
• Scalable architecture – The multi‑agent design runs inference in parallel and fits within typical latency SLAs (≈30 ms on commodity GPUs), meaning it can be rolled out to any high‑traffic e‑commerce site.

Limitations & Future Work

• LLM hallucination risk – While diffusion reasoning is constrained by semantic vectors, occasional generation of implausible items was observed; tighter grounding mechanisms are needed.
• Domain‑specific jargon – The current prompts are tuned for consumer goods; adapting to highly technical domains (e.g., B2B software) may require custom knowledge extraction pipelines.
• Compute cost – Adding two LLM inference stages increases GPU usage; future work will explore distillation or quantization to keep operational expenses low.
• User privacy – Although ReaSeq reduces raw log dependence, it still consumes session data; integrating differential privacy guarantees is an open research direction.

Overall, ReaSeq demonstrates that marrying world knowledge with reasoning can break the “log‑only” ceiling that many recommender systems face today, opening a path toward more intelligent, context‑aware, and universally applicable recommendation engines.

Authors

• Chuan Wang
• Gaoming Yang
• Han Wu
• Jiakai Tang
• Jiahao Yu
• Jian Wu
• Jianwu Hu
• Junjun Zheng
• Shuwen Xiao
• Yeqiu Yang
• Yuning Jiang
• Ahjol Nurlanbek
• Binbin Cao
• Bo Zheng
• Fangmei Zhu
• Gaoming Zhou
• Huimin Yi
• Huiping Chu
• Jin Huang
• Jinzhe Shan
• Kenan Cui
• Longbin Li
• Silu Zhou
• Wen Chen
• Xia Ming
• Xiang Gao
• Xin Yao
• Xingyu Wen
• Yan Zhang
• Yiwen Hu
• Yulin Wang
• Ziheng Bao
• Zongyuan Wu

Paper Information

• arXiv ID: 2512.21257v1
• Categories: cs.IR, cs.CL
• Published: December 24, 2025
• PDF: Download PDF