RecGPT-V2 Technical Report

1. Introduction

• RecGPT-V1’s multi-route reasoning redundantly encodes user histories and overlaps in candidate generation, while fixed prompts, static-data supervision, and one-shot judging limit adaptation and nuanced assessment.The cited limitations include a 13.46% inter-route duplication rate and average 32K-token behavior sequences.Source: §1 · pp. 3–4
• HMAS coordinates specialized agents to reduce duplicated reasoning; combined with compressed behavior representations, it raises exclusive recall from 9.39% to 10.99% and cuts GPU consumption by 60.0%.Hybrid Representation Inference compresses user-behavior tokens from 32K to 11K through atomized entity encoding.Source: §1 · p. 4
• +7.3% explanation diversity over RecGPT-V1’s static templates through prompts adapted to user interests, item attributes, and real-time contextual signals.The examples of contextual signals include weather, seasonal events, and trending topics.Source: §1 · p. 4
• +24.0% human-evaluated tag quality pass rate and +77.6% human-rated explanation acceptance over RecGPT-V1 with constrained reinforcement optimization.The approach shapes rewards rather than directly summing multiple signals.Source: §1 · pp. 4–5
• +0.46% human preference alignment on item tag prediction and +1.76% on recommendation explanation generation versus LLM-as-a-Judge baselines.Agent-as-a-Judge iteratively assesses dimensions such as relevance, diversity, and coherence; the pipeline also distills agent judgments into optimization reward signals.Source: §1 · p. 5
• +3.64% IPV, +3.01% CTR, +2.11% TV, and +11.46% NER over RecGPT-V1 in large-scale Taobao homepage A/B tests.These are the reported online performance improvements.Source: §1 · p. 5

2. Agentic Intent Reasoning

• RecGPT-V1’s routes each encoded an average 32K-token behavior sequence and had 13.46% inter-route redundancy; RecGPT-V2 jointly targets both inefficiencies.The proposed framework combines compact representations with coordinated reasoning.Source: §2 · p. 5

2.1. Hybrid Representation Inference

• 7× compression is achieved by atomizing behavioral entities while preserving task performance, addressing the fact that user lifelong behaviors account for approximately 95.89% of input tokens.The approach combines self-perception QA and production-task supervision, training adaptor parameters while keeping the LLM backbone frozen.Source: §2.1 · p. 6, ⇓Atomized Entity Compression · pp. 8, 10

2.1.1. Atomized Entity Compression

• A 21,349-token user profile is reduced to 5,158 tokens (76%) by replacing item descriptions and query texts with atomic representations while retaining attributes and temporal metadata.An embedding model and lightweight adaptor map entities into atomic units compatible with LLM input.Source: §2.1.1 · pp. 6–7

2.1.2. Infrastructure Engineering Optimization

• The serving design separates prefill and decode resources by phase-specific demands and replaces FlashInfer with an XQA kernel using FP8 inference on H20 GPUs.Prefill receives a larger GPU pool for long-context processing, while decode receives fewer resources for memory access patterns.Source: §2.1.2 · p. 10
• 53.11% higher MFU than RecGPT-V1, with MFU rising from 11.56% to 17.04%, prefill QPS improving ×69.30, and decode TPS improving ×7.35.These gains combine infrastructure optimizations, atomized entity compression, and the coordinated reasoning architecture.Source: Figure 4 · p. 10, §2.1.2 · pp. 10–11

2.2. Hierarchical Multi-Agent System

• The Hierarchical Multi-Agent System uses a Planner→Experts→Arbiter architecture to coordinate specialized tag prediction and synthesize candidates, reducing duplicated reasoning while preserving diverse intent coverage.The Global Planner analyzes hybrid compressed context and multi-source contextual signals; experts reason in parallel and the Arbiter synthesizes their predictions.Source: §2.2 · p. 11

2.2.1. Global Planner

• The Global Planner synthesizes rich contextual signals into a coherent strategic plan, decomposing complex user intent into specialized personas for downstream expert reasoning.Unlike isolated parallel routes that independently process the same contexts, it performs holistic intent analysis.Source: §2.2.1 · p. 11
• The Global Planner compresses behavioral context and decomposes hybrid user context into complementary personas, reducing redundant processing and overlapping expert reasoning.Behavioral entities use atomic compression, while user attributes and environmental signals retain natural-language encoding.Source: §2.2.1 · p. 12

2.2.2. Distributed Experts

• Persona-aligned supervised fine-tuning uses GPT-4 to identify categories in users’ subsequent interactions that match each persona’s intent focus.When relevant categories are insufficient, the training procedure augments the target set with GPT-4-generated tags; when there are more than 15, it samples 15.Source: §2.2.2 · p. 13
• Constrained Reward Shaping optimizes accuracy only after alignment, diversity, and length constraints are met, avoiding the conflicting gradients of weighted-sum reward aggregation.The composite reward is accuracy multiplied by indicators for the three secondary-reward thresholds, so any unmet constraint sets the total reward to zero.Source: §2.2.2 · p. 15, Figure 6 · p. 15
• CRS maintains lower gradient norms and KL divergence than SUM while sustaining positive optimization across objectives; SUM’s diversity gains later degrade accuracy.The report attributes CRS’s stability to decoupling constraint satisfaction from optimization of the primary accuracy objective.Source: §2.2.2 · pp. 15–16, Figure 7 · p. 16
• 32.60% HR@30 from GRPO (CRS) exceeded SFT by 3.40% and RecGPT-V1 by 6.31%.The evaluation uses Hit Rate at top-30 predictions to measure whether predicted tags mapped to categories match users’ actual interaction categories.Source: Table 2 · p. 16, §2.2.2 · p. 16

2.2.3. Decision Arbiter

• The Decision Arbiter jointly selects tags across the candidate pool for relevance, profile consistency, specificity, and validity, accounting for complementarity and avoiding redundancy.For online recommendations, multiple learned interest vectors support item matching, while quadratic-programming traffic allocation balances exploratory cognitive-channel items with the existing utility channel.Source: §2.2.3 · p. 17

3.1. Meta-Prompting

• Meta-prompting first synthesizes adaptive stylistic guidelines, then generates explanations conditioned on those guidelines.The two-stage design supports varied personas and contextually adaptive explanations.Source: §3.1 · p. 18
• The framework expands explanation evaluation from four dimensions to seven by adding timeliness, informativeness, and attractiveness.The expanded criteria guide both meta-prompt generation and subsequent evaluation.Source: §3.1 · p. 18

3.2. Preference-Aware Reinforcement Learning

• Constrained Reward Shaping combines rule-based diversity and model-based alignment rewards, with diversity acting as a constraint to mitigate reward conflicts.For explanation generation, human-preference alignment is the primary reward and diversity is secondary.Source: §3.2 · pp. 18–19
• +7.30% diversity and +13.04% quality acceptance rate for RecGPT-V2, validating gains from meta-prompting and preference-aware reinforcement learning.Table 3 compares explanation performance; CRS helps improve both outcomes by eliminating gradient interference.Source: Table 3 · p. 19, §3.2 · p. 19

4.1. Agent-as-a-Judge

• Dimension-specific sub-evaluators assess quality separately, and a Senior Reviewer aggregates their outputs into Superior, Average, or Bad judgments.The framework uses progressive, hierarchical review rather than one-step holistic scoring.Source: §4.1 · pp. 20–21
• Agent-as-a-Judge improves human-judge agreement: item-tag accuracy rises by +0.10pp, +0.20pp, and +0.38pp across three models.For explanation generation, accuracy improves by +0.67pp for GPT5-mini and +1.21pp, with +5.20pp F1, for Qwen3-SFT.Source: Table 4 · p. 21, §4.1 · pp. 21–22

4.2. Judge-as-a-Reward

• Judge-as-a-Reward distills discrete agent judgments into continuous, differentiable reward signals that guide policy optimization.The reward model learns preference structure through listwise contrastive training.Source: §4.2 · pp. 22–23
• The integrated evaluation-and-reward cycle can continue after initial human annotation, accumulating training signals for reward-model calibration and policy refinement.The paper describes this as a self-reinforcing flywheel intended to avoid recurring annotation costs.Source: §4.2 · pp. 22–23
• +24.1% HR@30 and +13.0% Quality over RecGPT-V1, while listwise reward modeling beats pointwise training by +4.4% and +8.2%, respectively.Table 5 compares reward-model training strategies; listwise learning models the S ≻ A ≻ B preference ordering.Source: Table 5 · p. 23, §4.2 · p. 23

5.1. Online A/B Test

• The two-week online A/B test compared RecGPT-V2 with RecGPT-V1, allocating 1% of platform traffic to each group and evaluating item and feed scenarios.Table 6 reports relative percentage improvements across the two scenarios.Source: §5.1 · pp. 23–24, Table 6 · p. 24
• +3.26% IPV, +3.01% CTR, +2.11% TV, +3.39% GMV, and +3.47% ATC in the item scenario; feed CTR rose +1.50% and GMV +1.53%.NER increased +11.46% in the item scenario and +4.49% in the feed scenario.Source: Table 6 · p. 24, §5.1 · p. 24

5.2. Case Study

• In the case study, the Global Planner turns weather and upcoming holidays into three expert personas, whose complementary recommendations receive context-adaptive explanations.Examples include a cardigan for cooling weather and children’s Halloween clothing for the approaching holiday.Source: ③Decision Arbiter · pp. 25–26

6. Conclusion

• RecGPT-V2 combines agentic intent reasoning, meta-prompting, constrained reinforcement learning, and process-oriented Agent-as-a-Judge evaluation.The conclusion reports 60% lower GPU consumption and online gains of +3.40% IPV, +4.68% CTR, +4.05% TV, and +11.46% NER.Source: §6 · p. 26
• The authors identify end-to-end joint optimization of multi-agent collaboration with reinforcement learning as future work.The stated aim is to enhance recommendation performance and user experience.Source: §6 · p. 26
• The conclusion presents industrial deployment at scale as technically feasible and commercially viable.This conclusion is grounded in the reported large-scale Taobao deployment.Source: §6 · p. 26

• Core contributors include Chao Yi, Dian Chen, Gaoyang Guo, and Jiakai Tang†.Source: A. Contributors · p. 28
• Additional contributors listed are Jian Wu, Jing Yu, Mao Zhang, Wen Chen, Wenjun Yang, Yujie Luo, Yuning Jiang, and Zhujin Gao.Source: A. Contributors · p. 28
• The contributor list also includes Bo Zheng, Binbin Cao, Changfa Wu, Dixuan Wang, Han Wu, Haoyi Hu, Kewei Zhu, Lang Tian, Lin Yang, Qiqi Huang, Siqi Yang, Wenbo Su, Xiaoxiao He, and Xin Tong Xu Chen†.Source: A. Contributors · p. 28
• The dagger affiliation is Renmin University of China, and authors are listed alphabetically by first name.Source: A. Contributors · p. 28
• Other listed contributors are Xunke Xi, Xiaowei Huang, Yaxuan Wu, Yeqiu Yang, Yi Hu, Yujin Yuan, Yuliang Yan, and Zile Zhou.Source: A. Contributors · p. 28

• Table 7 lists the evaluation dimensions for item tag prediction.Source: Table 7 · p. 29, B. Evaluation Dimensions for… · p. 29
• Table 8 lists the evaluation dimensions for recommendation explanation generation, marking newly introduced dimensions with an asterisk.Source: Table 8 · p. 29, B. Evaluation Dimensions for… · p. 29

• Traffic allocation uses constrained quadratic programming to balance exploratory item discovery against conversion-driven optimization under limited exposure budgets.Source: C. Implementation Details · p. 29
• Each cognitive-channel candidate receives an exposure probability, with the objective maximizing click revenue while penalizing concentrated selection and constraints protecting conversion targets and exploration volume.Source: C. Implementation Details · pp. 29–30
• The exposure score combines click revenue, a weighted above-average conversion-revenue premium, and budget pressure that adjusts with constraint tightness.Source: C. Implementation Details · p. 31
• The analytical policy gives full exposure above the score threshold, fractional exposure proportional to score within the threshold range, and excludes items with negative scores.Source: C. Implementation Details · p. 31
• For production latency, intermediate-score items are excluded through a binarized hard-threshold policy, avoiding strict optimization at inference time.Source: C. Implementation Details · p. 31