Command Palette
Search for a command to run...
RecGPT-V2 기술 보고서
RecGPT-V2 기술 보고서
초록
대규모 언어 모델(Large Language Models, LLMs)은 추천 시스템을 암묵적인 행동 패턴 매칭에서 명시적인 의도 추론으로 전환하는 데 있어 놀라운 잠재력을 보여주고 있다. RecGPT-V1은 LLM 기반의 추론을 사용자 관심도 탐색 및 아이템 태그 예측에 통합함으로써 이 새로운 패러다임을 선도적으로 구현하였지만, 다음과 같은 네 가지 근본적인 한계를 지닌다: (1) 다수의 추론 경로에서 발생하는 계산적 비효율성과 인지적 중복; (2) 고정 템플릿 기반 생성으로 인한 설명 다양성 부족; (3) 감독 학습 패러다임 하에서의 제한된 일반화 능력; (4) 인간 기준에 부합하지 못하는 단순한 결과 중심 평가 방식. 이러한 문제를 해결하기 위해, 본 연구는 네 가지 핵심 혁신을 도입한 RecGPT-V2를 제안한다. 첫째, 계층적 다중 에이전트 시스템(Hierarchical Multi-Agent System)을 통해 의도 추론을 조율된 협업 구조로 재설계함으로써 인지적 중복을 제거하고 다양한 의도 커버리지를 가능하게 하였다. 이와 함께 사용자 행동 컨텍스트를 압축하는 하이브리드 표현 추론(Hybrid Representation Inference)을 도입함으로써 GPU 소비량을 60% 감소시키고, 독점적 리콜률(Exclusive Recall)을 9.39%에서 10.99%로 개선하였다. 둘째, 메타 프롬프팅(Meta-Prompting) 프레임워크를 도입하여 맥락에 적응하는 동적 프롬프트 생성을 실현함으로써 설명 다양성을 +7.3% 향상시켰다. 셋째, 제약 강화 학습(constrained reinforcement learning)을 적용하여 다중 보상 간의 갈등을 완화하였으며, 태그 예측 성능은 +24.1%, 설명 수용률은 +13.0% 개선되었다. 넷째, 에이전트를 심사관으로 활용하는 Agent-as-a-Judge 프레임워크를 도입하여 평가 과정을 다단계 추론 구조로 분해함으로써 인간 선호도와의 일치도를 향상시켰다. 타오바오에서 수행한 온라인 A/B 테스트 결과, 클릭률(CTR)은 +2.98%, 인구당 방문 횟수(IPV)는 +3.71%, 전환율(TV)은 +2.19%, 신뢰도 평가율(NER)은 +11.46% 개선되는 등 의미 있는 성능 향상을 확인하였다. RecGPT-V2는 LLM 기반 의도 추론을 대규모로 실현할 수 있는 기술적 타당성과 상업적 실현 가능성을 입증하며, 인지적 탐색과 산업적 활용 사이의 격차를 좁히는 데 기여한다.
One-sentence Summary
The authors propose RecGPT-V2, which addresses RecGPT-V1's four limitations through a Hierarchical Multi-Agent System eliminating cognitive redundancy while enabling diverse intent coverage, Meta-Prompting for dynamic explanation generation, constrained reinforcement learning for multi-reward optimization, and Agent-as-a-Judge for process-oriented evaluation, achieving 60% GPU reduction and significant Taobao performance gains including +11.46% novelty exposure rate.
Key Contributions
- RecGPT-V2 addresses critical deficiencies in prior systems including computational inefficiency from redundant intent reasoning and low-information-density explanations by introducing a Hierarchical Multi-Agent System with Hybrid Representation Inference, reducing GPU consumption by 60% and improving exclusive recall from 9.39% to 10.99%.
- The framework overcomes homogenized expression and weak temporal adaptation through Meta-Prompting for dynamic context-aware explanation generation, which increases explanation diversity by 7.3% while capturing real-time signals like seasonal trends.
- Preference-aware reinforcement learning resolves multi-reward conflicts in complex generation tasks, achieving 24.1% higher tag prediction accuracy and 13.0% greater explanation acceptance, with online A/B tests on Taobao confirming significant gains in CTR (2.98%), IPV (3.71%), and NER (11.46%).
Introduction
Recommendation systems increasingly rely on personalized explanations to boost user engagement with suggested items, but static explanation approaches suffer from repetitive, context-ignorant outputs that reduce effectiveness. Prior work like RecGPT-V1 faced critical limitations: rigid prompt templates produced low-information-density explanations with poor adaptation to temporal trends or user context, while evaluation frameworks failed to capture key quality dimensions like stylistic diversity. The authors address this by introducing Meta-Prompting to dynamically synthesize context-aware prompt templates and preference-aware reinforcement learning that optimizes explanations via human-aligned multi-reward modeling. Together, these innovations shift explanation generation from inflexible templating to adaptive reasoning, directly tackling engagement shortcomings observed in live deployments.
Dataset
The authors use a bilingual e-commerce product title dataset for translation tasks. Key details include:
- Composition and sources: The dataset consists of paired Chinese product titles and their English translations, sourced from online retail platforms.
- Subset details:
- Contains user-generated product listings with verified translations.
- Specific split sizes are unspecified, but examples follow a strict {Chinese: English} format.
- No explicit filtering rules are stated; entries appear curated for direct translation relevance.
- Usage in training:
- The data forms the core training split for sequence-to-sequence translation models.
- Used as a single-task mixture without additional data sources or ratio adjustments.
- Processing:
- Raw titles undergo minimal preprocessing; the example shows direct character-level pairing.
- No cropping strategy is applied—full titles are preserved as input-output pairs.
- Metadata is derived implicitly from the paired structure, with no auxiliary annotations added.
Method
The authors leverage a comprehensive, multi-component architecture in RecGPT-V2 to overcome the computational inefficiency, cognitive redundancy, and evaluation limitations of its predecessor. The system is structured around three core innovations: Agentic Intent Reasoning, Dynamic Explanation Generation, and an Agentic Judge Framework, each addressing a specific bottleneck in the recommendation pipeline.
The Agentic Intent Reasoning module forms the backbone of the system, replacing RecGPT-V1’s parallel, redundant LLM routes with a coordinated Hierarchical Multi-Agent System (HMAS). This system operates in three stages: a Global Planner decomposes user intent into specialized personas by synthesizing user behavioral history, static/dynamic profile attributes, and real-time environmental signals (e.g., weather, trending events). These personas are then distributed to Distributed Expert agents, each responsible for generating item tags aligned with a specific intent facet. Finally, a Decision Arbiter consolidates the expert outputs, performing joint reasoning over the entire candidate tag pool to select the most behaviorally relevant, profile-consistent, and non-redundant tags for downstream retrieval. This coordinated architecture eliminates redundant full-context encoding and cognitive overlap, as illustrated in the comparison between RecGPT-V1’s isolated routes and RecGPT-V2’s HMAS.
To enable this architecture at scale, the authors introduce Hybrid Representation Inference. This technique compresses the user’s lifelong behavioral sequence—often exceeding 32K tokens—into a compact hybrid context. The core of this compression is Atomized Entity Compression, which encodes item and query descriptions into dense vector representations using pretrained embedding models (e.g., BGE, Qwen3-Embedding). These vectors are then projected into the LLM’s input space via a lightweight, trainable adaptor network, replacing multi-token textual descriptions with single atomic tokens denoted as [entity]. This process achieves a 7x compression ratio while preserving semantic integrity, as shown in the example where a 21,349-token user profile is reduced to 5,158 tokens.
The authors further enhance efficiency through Infrastructure Engineering Optimizations, including a Disaggregated Prefill-Decode Serving Architecture. This design allocates separate GPU pools to the compute-intensive prefill phase (processing long contexts) and the memory-intensive decode phase (generating outputs), significantly improving Model FLOPs Utilization (MFU). Combined with the use of XQA kernels for FP8 precision inference, these optimizations reduce GPU consumption by 60% and improve MFU by 53.7% compared to RecGPT-V1.
For dynamic explanation generation, the authors move beyond RecGPT-V1’s fixed templates by introducing a Meta-Prompting framework. This two-stage process first synthesizes a contextually adaptive stylistic guideline based on user interests, item attributes, and situational signals (e.g., seasonal trends). The second stage then generates the final explanation conditioned on this guideline, enabling diverse, emotionally resonant, and timely outputs. This approach improves explanation diversity by +7.3% and better aligns with user expectations across expanded evaluation dimensions, including Timeliness, Informativeness, and Attractiveness.
Finally, to address the limitations of outcome-focused evaluation, the authors propose an Agent-as-a-Judge framework. This system decomposes holistic quality assessment into Multi-Dimension Sub-Evaluators, each specializing in a specific criterion (e.g., Relevance, Factuality, Timeliness). A Senior Reviewer Agent then aggregates these sub-evaluations into a three-tier judgment (Superior, Average, Bad) through a structured, two-stage process: first detecting critical defects, then elevating high-quality outputs. This process-oriented evaluation improves human preference alignment by +0.46% on tag prediction and +1.76% on explanation generation. To enable continuous improvement, the authors introduce Judge-as-a-Reward, a distillation framework that converts these discrete judgments into dense, differentiable reward signals for reinforcement learning, establishing a self-reinforcing flywheel for policy optimization.
Experiment
- Conducted two-week online A/B test on Taobao comparing RecGPT-V2 against RecGPT-V1 across item and feed recommendation scenarios with 1% traffic allocation per group
- Achieved significant improvements in short-term metrics: +3.40% IPV, +4.68% CTR, +4.05% TV, and +11.46% NER (Novelty Exposure Rate) versus baseline
- Demonstrated enhanced long-term retention with +0.04% LT-14 and +0.05% LT-30 user retention rates
- Validated dynamic intent adaptation through real-world case analysis showing context-aware recommendations integrating environmental signals like weather and holidays
- Reduced GPU consumption by 60% while improving generation quality for item tag prediction and explanation tasks
The authors use RecGPT-V2 in a two-week A/B test on Taobao, comparing it against RecGPT-V1 across item and feed recommendation scenarios. Results show consistent improvements in both short-term engagement and long-term retention, with the item scenario achieving the highest gains in IPV (+3.64%) and NER (+11.46%), while the feed scenario shows modest but meaningful retention gains (LT-14 +0.04%, LT-30 +0.05%).

The authors evaluate RecGPT-V2 using two reward modeling approaches against RecGPT-V1, showing that List-wise RM achieves the highest HR@30 (Tag) at 32.60% and Quality (Explanation) at 40.73%, indicating improved tag prediction accuracy and explanation quality over prior versions. Results confirm that List-wise reward modeling enhances both retrieval effectiveness and explanatory output compared to Point-wise RM and the baseline.

The authors evaluate RecGPT-V2 against baseline models on item tag prediction and explanation generation tasks, showing consistent improvements in both accuracy and F1 score across all tested models. For item tag prediction, Qwen3-SFT achieves the highest performance with 0.8248 accuracy and 0.8228 F1 in V2, while for explanation generation, Qwen3-SFT also leads with 0.7006 accuracy and 0.7307 F1 in V2, indicating enhanced generation quality. These results support the system’s effectiveness in producing more precise and contextually relevant outputs.

The authors evaluate RecGPT-V2 against RecGPT-V1 using diversity and quality metrics, showing that RecGPT-V2 achieves higher diversity (0.677 vs. 0.631) and quality (40.73% vs. 36.03%). These results indicate improved recommendation variety and output accuracy in the updated system.

The authors evaluate RecGPT-V2 using HR@30 across multiple configurations, showing that the GRPO (CRS) variant achieves the highest score at 32.60%, outperforming both the baseline RecGPT-V1 (26.29%) and other variants including SFT (29.20%) and GRPO (SUM) (27.38%). Results indicate that the CRS optimization strategy contributes most significantly to recall performance.
