RecGPT-V3 Technical Report

• The RecGPT series faces stateless full-history processing, lossy natural-language tag-to-item grounding, and costly explicit reasoning that averages about 3,000 tokens per task.These challenges respectively cause redundant computation without carrying forward prior analysis, leave a gap between predicted intent and retrieved items, and impose prohibitive latency at billion-user scale and high QPS.Source: §1 · p. 3
• The Memory Hub distills long-term behavior into traceable, incrementally curated memory units, compressing tokens by 94.5% and reducing user-modeling computation by 55.8%.The system combines consolidated memory with a recent behavioral increment rather than reprocessing the full raw history.Source: §1 · p. 4
• The Hybrid-modal Foundation Model reasons over natural-language intent and Semantic IDs, combining open-ended world knowledge with concrete item representations enriched by collaborative signals.Generating representations compatible with the downstream model creates a higher-bandwidth channel from intent reasoning to product retrieval than the language-only interface.Source: §1 · p. 4
• Latent Intent Reasoning encodes lengthy chain-of-thought in learnable latent tokens, cutting token cost by 200× while allowing readable rationales to be decoded on demand.Dedicated latent <CoT> slots encode an entire reasoning trace.Source: §1 · p. 4
• In large-scale online A/B testing against live RecGPT-V2, RecGPT-V3 delivered +1.28% IPV, +1.00% CTR, +1.97% TC, and +3.97% GMV while reducing end-to-end serving compute by 52.4%.Source: §1 · p. 5

2. Memory Hub

• Memory Hub addresses repeated full-history analysis by making user memory condensed, traceable to source behaviors, and incrementally evolvable.The Planner previously reprocessed about 55K tokens for highly active users and accounted for about 95% of agentic intent-analysis computation; memory compression achieves an 80% token reduction.Source: §2 · p. 5

2.1. Structured Behavior Compression

• Structured Behavior Compression converts the full behavior sequence into fewer memory units, each pairing a taxonomy-based behavior pattern with a natural-language preference summary and provenance metadata.Units also record representative behavior indices, preferred brands, a temporal activity profile, and a timestamp; the initial compression runs once to establish persistent memory.Source: §2.1 · p. 5

2.2. Evolving Memory Curation

• Evolving Memory Curation selectively refreshes relevant existing units, retains unaffected ones, and extracts new units from coherent unmatched behaviors using the current memory and recent behavior delta.The update and new-pattern operations share one model forward pass and do not access the full behavior sequence.Source: §2.2 · pp. 7–8
• Selective updates re-synthesize affected units as coherent snapshots, preventing stale and emerging preferences from accumulating together while leaving unaffected units unchanged.The paper illustrates this with evolving age-range and cookware interests and a newly formed unit for a first-time interest.Source: §2.2 · p. 8
• 55.8% less user-modeling computation is achieved while preserving recommendation quality.In production, incremental curation runs every two months.Source: §2 · p. 5, §2.2 · p. 8

Problem addressed

• Text-only LLM experts predict free-form tags for downstream retrieval, but coarse tags form a lossy bottleneck that weakens the link between expressed intent and surfaced items.A broad tag can match heterogeneous items without conveying the fine-grained, item-level evidence retrieval depends on.Source: §3 · p. 8

Hybrid-modal model

• The model extends Qwen3-14B (Yang et al., 2025) with Semantic IDs (SIDs) as a second modality alongside natural language.Recommendation requires reasoning about user intent while grounding it in concrete items, which neither representation satisfies alone.Source: §3 · p. 8

Complementary modalities

• Natural language provides rich, open-ended world knowledge, while SIDs ground intent in item representations enriched by collaborative signals.Source: §3 · p. 8

Unified representation and pipeline

• A unified vocabulary lets the model emit retrieval-compatible identifiers instead of coarse tags, opening a high-bandwidth channel into the item space while preserving natural-language capability.The pipeline combines semantically grounded SID tokenization (§3.1) with two-stage training to ground the tokens in language and enable joint reasoning over SIDs and text (§3.2).Source: §3 · p. 8

3.1. Hybrid Tokenization

• 65,536 Semantic ID tokens encode items as two-level codes whose shared coarse prefixes expose semantic similarity for generalization.Multimodal representations are contrastively learned from behavior-derived co-occurrence pairs filtered by multimodal similarity, then quantized with RQ-VAE using two 32,768-entry codebooks.Source: §3.1 · pp. 9–10, Figure 4 · p. 9

3.2. Pre-training

• A two-stage pipeline appends learned SID embeddings to Qwen3-14B, first grounding them in item semantics and then teaching their use across recommendation tasks.General-domain data is mixed into both stages to balance domain knowledge with the backbone’s pre-existing general capabilities.Source: Figure 4 · p. 9, §3.2 · p. 10

3.2.1. Continual Pre-training

• Continual pre-training grounds SIDs in natural-language item descriptions and mixes approximately 10% general-domain text to guard against catastrophic forgetting.SID-grounding samples pair each identifier with item attributes, teaching the model to recover item content from the SID; this gives the tokens working semantics before instruction tuning.Source: §3.2.1 · pp. 10–11

3.2.2. Instruction Tuning

• Instruction tuning covers SID–text translation, sequential recommendation in SID space, and general-domain tasks to retain the model’s existing capabilities.The sequential task predicts next-click SIDs from click histories across multiple temporal windows, without textual item descriptions; roughly 20% of tuning data is open-domain supervised tasks.Source: §3.2.2 · p. 11

4. Latent Intent Reasoning

• Latent Intent Reasoning compresses thousands of autoregressively decoded reasoning steps into a handful of tokens, cutting reasoning-token cost 200× while retaining decodable explanations.The tokens shift computation toward parallelizable prefill and can be decoded into faithful, human-readable rationales.Source: §4 · p. 12

4.1. Reasoning Internalization

• Latent slots encode contiguous trace segments, and warm-up plus single-, multi-, and full-trace reconstruction align each token with its reasoning content.Full-trace reconstruction trains the latent sequence to recover the reasoning chain, preserving the ability to decode a readable rationale.Source: §4.1 · pp. 13–14

4.2. Post-training

• Post-training has two stages: distilling and internalizing teacher reasoning, then optimizing the latent model against online business objectives.The first stage uses the §4.1 alignment curriculum; the second applies reinforcement learning from ranking feedback.Source: §4.2 · pp. 14–15

4.2.1. Stage 1: Explicit-to-Implicit CoT Alignment

• Stage 1 compresses teacher traces averaging roughly 2,300 tokens into at most 10 latent tokens, using C = 20 and a mixture of alignment, general-reasoning, and tag-prediction data.Tag prediction supervises output tokens only, helping prevent latent-position collapse.Source: §4.2.1 · p. 15

4.2.2. Stage 2: Reinforcement Learning from Ranking Feedback

• RLRF replaces sparse offline HitRate with dense CTRScore feedback from the production ranker, aligning reinforcement-learning rewards with the serving pipeline.It averages the top 100 ranking scores of items retrieved by generated tags; retained alignment, diversity, and length constraints gate the primary reward.Source: §4.2.2 · pp. 15–16

5.1. Online A/B Test Results

• In the online A/B test’s feed scenario, RecGPT-V3 improved relative to RecGPT-V2 by +1.28% IPV, +1.00% CTR, +0.83% PV, +0.56% DAU, +3.97% GMV, and +1.97% TC.Source: Table 6 · p. 17, §5.1 · p. 17

5.2. Memory Hub Evaluation

• Memory-unit accuracy was 82.89% across 2,514 behavior patterns and 95.27% across 21,268 behavior-index annotations; the Memory Hub cut Global Planner compute cost by 55.80% versus RecGPT-V2.Each inference pass conditions on compressed memory plus recent behavioral delta, reducing per-pass cost to 33.43%; incremental curation contributes 10.77%.Source: §5.2.1 · pp. 17–18, §5.2.2 · p. 18

5.3. Foundation Model Evaluation

• General-domain data mixing preserved most backbone capabilities and kept SID–text alignment essentially unchanged, while lifting category-level HR@30 by 26.2% (0.2250 →0.3050).Without general-domain data, scores fell to 4.70% on GSM8K, 0.12% on MMLU, 0.01% on CMMLU, and 23.29% on IFEval; with mixing, GSM8K dropped 1.66%, MMLU 2.65%, and CMMLU 4.49%, while IFEval fell from 81.52% to 75.60%.Source: §5.3.1 · pp. 18–19, Figure 6 · p. 19, Figure 7 · p. 19

5.4.1. Post-training Effectiveness

• RL on latent reasoning reached HR@30 0.3693 and CTR 0.0679, exceeding explicit CoT on both metrics; latent reasoning compressed ∼2,700 reasoning tokens into 10 latent tokens.Latent reasoning achieved HR@30 0.3462 and CTR 0.0649 before RL, comparable in quality to explicit CoT’s HR@30 of 0.3508.Source: §5.4.1 · p. 20, Table 10 · p. 21

5.4.2. Inference Efficiency

• Latent reasoning cut output length from 2,840 to 122 tokens (95.7%) and processing time from 1,020s to 295s (3.46× speedup); overall resource savings reached 52.4%.On identical hardware, input throughput rose from 166K to 498K tokens per minute; despite 15% expert-model overhead, Global Planner savings outweighed it because planner cost is approximately 20 times expert-model cost.Source: Table 11 · p. 21, §5.4.2 · p. 21

5.5.1. SID Modality Analysis

• Text tags covered more categories (1.40 vs. 0.84), while SIDs had lower cross-user overlap (4.61% vs. 11.36%); hybrid retrieval reached HR@500 0.1571, above SID-only 0.1539 and text-tag-only 0.1503.The embedding visualization showed tag-matched items more spatially dispersed and SID-matched items in tighter clusters, consistent with broader tag coverage and greater SID specificity.Source: §5.5.1 · pp. 22–23, Table 12 · p. 22, Figure 8 · p. 22

5.5.2. Case Study

• In the case study, the Memory Hub compressed behavior into preference units and selectively updated them as new interactions arrived.Updates included refreshing technology and badminton interests, evolving baby-care interests, retaining stable home-improvement preferences, and creating a new running preference.Source: §5.5.2 · p. 23
• 10 latent tokens replace long explicit chain-of-thought during inference and can be decoded on demand into a readable rationale.The case uses curated memory as the basis for Latent Intent Reasoning.Source: §5.5.2 · pp. 23–24
• The reconstructed rationale identifies a serious badminton player with an offensive style and infers needs for rackets, strings, maintenance tools, and related equipment.Source: §5.5.2 · p. 24
• The inferred intents become concrete memory-driven recommendations grounded with SIDs, demonstrating evolving user understanding and explainable reasoning.Source: §5.5.2 · p. 24

6. Conclusion

• +1.28% IPV, +1.00% CTR, +1.97% TC, and +3.97% GMV in large-scale online A/B tests, with end-to-end serving compute reduced by 52.4% in Taobao’s “Guess What You Like” feed.Source: §6 · p. 24
• RecGPT-V3 combines a continually evolving structured Memory Hub, joint natural-language and Semantic-ID item grounding, and decodable latent tokens that compress explicit chain-of-thought.Source: §6 · p. 24
• It addresses stateless behavior modeling, the tag-to-item information bottleneck, and inefficient explicit reasoning, with gains showing prediction accuracy and serving efficiency can improve together.Source: §6 · p. 24

B. Intent-to-Item Retrieval

• Tags convey open-ended, generalizable world knowledge, while SIDs ground intent in concrete item representations enriched by collaborative signals.Earlier RecGPT versions used a text-only retrieval pathway and did not exploit the collaborative signal carried by SIDs.Source: B. Intent-to-Item Retrieval · p. 27
• The hybrid retrieval model takes both text tags and SIDs as input and learns a preference ordering over candidate items.Its design addresses the need to consume both channels jointly for downstream item retrieval.Source: B. Intent-to-Item Retrieval · p. 27
• The model favors items users are likely to click while preserving semantic alignment with upstream intent predictions.The section proceeds to describe the hybrid input architecture (§B.1) and joint training objective (§B.2).Source: B. Intent-to-Item Retrieval · p. 27

B.1. Architecture

• Text tags express what the user wants, while paired SIDs encode which items satisfy that intent in the discrete item space.The retriever takes both channels as input and assigns relevance scores to candidate items.Source: §B.1 · p. 27
• The model separately embeds each tag and SID, concatenates their embeddings, and projects them into unified intent vectors for target-attention retrieval scoring.The shared attention computation enables cross-channel interaction while preserving channel-specific information in early layers.Source: §B.1 · pp. 27–28
• The SID channel adds collaborative signals unavailable to text alone, pairing language’s semantic breadth with discrete representations’ collaborative precision.Contrastive pre-training encodes item-level signals including co-purchase patterns, popularity dynamics, and item-to-item affinity.Source: §B.1 · p. 28

• The preference hierarchy makes click utility primary and uses semantic relevance to distinguish candidates of comparable utility, combining both signals in one training objective.Click-only supervision can favor globally popular items that do not match upstream tags and SIDs.Source: §B.2 · p. 28
• Utility uses a candidate-set softmax over clicked items; relevance uses progressively coarser contrastive positives and restricts negatives by interaction depth (clicked > exposed > non-exposed).Negatives are limited to candidates whose priority is no higher than the positive item's.Source: §B.2 · pp. 28–29
• The final objective is a weighted sum of utility and relevance losses, with α = 1 and β = 0.5 in production experiments.The combined optimization is intended to favor items that are both click-worthy and semantically aligned with upstream intent.Source: §B.2 · p. 29

Reconstruction prompts

• Reconstruction prompts ask the model to explain each <cot> token for single- or multi-segment tasks, or recover the entire chain of thought for full-trace tasks.The prompt is placed in the system turn; the user turn supplies the recommendation context, a <think> block with masked reasoning replaced by latent tokens, and the final output.Source: C. Latent Reasoning Reconstruction · p. 29

Single- and multi-segment reconstruction

• Segment-level examples reconstruct bag preferences and apparel recommendations from user profiles, purchase histories, click histories, and seasonal needs.The reconstructed reasoning identifies preferences such as bag styles and practical features, and apparel styles, materials, and comfort needs.Source: C. Latent Reasoning Reconstruction · pp. 30–31

Full-trace reconstruction

• The full-trace automotive example recommends specific car-related products, including a cute cartoon car ornament and a sun-protective cartoon sunshade.The reconstructed trace connects the user’s car-related purchase history and preference for cute cartoon decorations with recommendations across practical and decorative dimensions.Source: C. Latent Reasoning Reconstruction · p. 32