PaperlayerView exact revision on arXiv

Source-linked AI summary

TGR: Advancing Industrial Recommendation from Generative-Paradigm Ranking toward Unified Generation and Reasoning

TGR Team, Lei Cheng, Haonan Hu, Beibei Kong, Yudong Li, Zang Li, Yunsheng Pang, Hongyang Su, Jianchao Tu, Yunlong Wang, Bing Wen, Junzhang Zhu, Shaojie Zhu, Chengxiang Zhuo

arXiv:2609.00986v1cs.IR

TL;DR

Industrial recommendation uses separately optimized cascade stages, while fine-grained cross-field interaction is sacrificed to control sequence cost. TGR develops a three-direction generative recommendation stack; BARGE improves Hit@5 over OneRec by 10.2% and 16.9% in two scenarios.

  • Problem

    Industrial recommenders rely on separately optimized cascade stages, while fine-grained cross-field interaction is sacrificed to contain sequence cost.

    Source: §1 · pp. 3–4

  • Method

    TGR advances generative recommendation through coupled ranking, end-to-end generation, and reasoning-augmented recommendation directions.

    Source: §1 · p. 5, §6.1 · p. 43

  • Results

    BARGE leads all four Hit@K metrics in both scenarios, improving Hit@5 over OneRec by 10.2% in Scenario 1 and 16.9% in Scenario 2.

    Source: §4.2.3 · p. 21

  • Takeaways & Limitations

    BARGE retains a 12.6% relative improvement at Hit@50, showing its gains extend beyond the first few ranked positions.

    Source: §4.2.3 · p. 21

    TGR remains a coupled stack rather than one model: GenRank emits per-item multi-task scores for the downstream cascade, and GenRec uses two production models with two tokenizers.

    Source: §6.2 · p. 43

Abstract

from arXiv · show

Industrial recommender systems typically rely on cascaded retrieval, pre-ranking, ranking, and reranking stages, whose separately optimized models limit scaling, fragment decision making, and lack semantic knowledge and reasoning. We present TGR (Tencent Generative Recommendation), an industrial framework that advances recommendation toward the generative paradigm along three coupled directions. TGR-GenRank upgrades ranking through CCFormer, which combines unified feature tokenization, a scalable Transformer backbone, feature-field separated cross attention, subspace token mixing, and hierarchical sequence compression while retaining per-item multi-task outputs. TGR-GenRec explores end-to-end generation under two paradigms: BARGE bridges item-boundary loss and semantic drift in hierarchical semantic-ID generation through item context-aware attention, hierarchical path reranking, and orthogonal dual-path decoding; HiGR performs whole-slate generation with prefix-structured semantic IDs, coarse-to-fine decoding, and listwise multi-objective alignment. TGR-Reason injects offline-generated semantic-ID reason tokens into online decoding, providing reasoning without request-time rollout. TGR is deployed across Tencent production surfaces serving hundreds of millions of users. CCFormer delivers significant gains in five A/B-tested scenarios and is fully launched in two, including +3.57% CTR and +1.71% advertising revenue. BARGE improves Hit@5 by 10.2-16.9% and yields +0.60% CTR and +1.70% reading time after full rollout. HiGR improves offline slate quality by 15.9-21.3% with a 5x inference speedup and achieves up to +1.22% watch time and +1.73% video views. TGR-Reason raises cold-start new-user Hit@1 by 477.8% and delivers +1.75% effective consumption and +13.09% new-user exposure-to-conversion online.

1. Introduction

TGR advances industrial recommendation beyond the cascade through generative-paradigm ranking, end-to-end generation, and offline-precomputed reasoning. Its three directions target the scaling limits of ranking, misalignment between item-wise decisions and slate experience, and the serving cost of request-time reasoning.

Motivation

  • Industrial ranking struggles to convert added compute into quality, while long-sequence self-attention costs drive pre-compression or truncation that sacrifices fine-grained cross-field interaction.Generative-paradigm rankers tokenize user, item, context, and sequence features into a unified stream, but sequence-length costs remain a constraint.Source: §1 · pp. 3–4
  • Stage-wise cascade optimization and candidate truncation fragment decisions, while item-in-isolation scoring misses slate interdependencies and positional effects; autoregressive slate generation is too slow for real-time use.The user experiences an ordered slate, so list-level quality is not captured by pointwise scoring.Source: §1 · p. 4
  • Behavioral co-occurrence models lack multimodal semantics, world knowledge, and deliberate reasoning, limiting sparse-supervision generalization; per-request LLM reasoning exceeds industrial serving latency budgets.These limits particularly affect cold-start, long-tail, and ambiguous-intent regimes.Source: §1 · p. 4

Related Work

  • Prior work spans scalable generative rankers such as HSTU [6] and OneTrans [8], end-to-end systems including OneRec [16] [17] [18] and GPR [19], and reasoning systems OneRec-Think [10] and OneReason [20].These systems revisit parts of the cascade or replace it, and TGR benchmarks against or sits alongside them.Source: §1 · p. 4

TGR Stack

  • TGR spans Tencent’s news, video, music, and web-novel surfaces, serving hundreds of millions of daily active users under large-corpus and tight-latency constraints.The framework addresses these operating conditions through three coupled directions.Source: §1 · pp. 4–5

TGR-GenRank

  • CCFormer replaces production rankers with efficient Transformer ranking while retaining the cascade and per-item multi-task outputs.Feature-field separated cross attention, subspace token mixing, and hierarchical sequence compression support directed feature interaction and subquadratic long-sequence modeling without giving up access to the full history.Source: §1 · p. 5

TGR-GenRec

  • BARGE adapts single-item autoregressive generation to recommendation, while HiGR emits an entire slate as one structured object.BARGE uses item context-aware attention, hierarchical path reranking, and orthogonal dual-path decoding; HiGR combines PCRQ-VAE tokenization, a Hierarchical Slate Decoder, and ORPO-based listwise multi-objective alignment.Source: §1 · p. 5

TGR-Reason

  • TGR-Reason injects offline-exported reason tokens from a LatentRec-trained Think model into online generation without request-time reasoning rollout.It retains a single generator while adding deliberate reasoning, with Group Memory query capability.Source: §3 · p. 5

3. TGR-Reason — reasoning-augmented generative recommendation

TGR-Reason uses an offline-trained Think model to produce reasoning-enriched semantic-ID priors for a production generator, adding reasoning without request-time reasoning cost. It reports offline gains strongest for cold-start new users and statistically significant online A/B lifts.

TGR-Reason

  • All three TGR directions, including TGR-Reason, are deployed on production traffic and validated in online A/B tests; TGR serves hundreds of millions of users across Tencent surfaces.Source: §3 · p. 5
  • TGR-Reason exports reasoning-enriched semantic-ID priors from a LatentRec-trained offline Think model and injects them into the production generator at zero request-path reasoning cost.The design amortizes reasoning at training time.Source: §3 · p. 6
  • TGR-Reason delivers consistent offline gains, strongest for cold-start new users, and statistically significant online A/B lifts on a commercial content platform.Source: §3 · p. 6

2. The TGR Stack: Design Principles and Shared Foundations

TGR combines a production-grade ranking upgrade with end-to-end slate generation and reasoning, built on shared foundations. Its design centers on replacing cascade stages, sharing structure across models, and optimizing for whole-list outcomes.

2.1. Design Principles

  • TGR’s stack is governed by production-grade stage replacement, shared structure across models, and list-level optimization rather than isolated-item objectives.Each replacement must match or exceed incumbent offline metrics within the same online latency and GPU budgets. Shared tokenizer, encoding, and post-training infrastructure support the stack, while slate generation directly optimizes listwise preferences.Source: §2.1 · pp. 6–7

2.2. Problem Formulation

  • The task maps each user’s chronological history and request features to an ordered slate optimized for holistic list-level feedback, through ranking or end-to-end generation.In the cascade, CCFormer emits per-item multi-task scores for candidates; end to end, TGR-GenRec and TGR-Reason decode the slate as semantic-ID tokens, with reasoning-enriched priors conditioning generation.Source: §2.2 · p. 8

2.3. Shared Components

  • Shared components include paradigm-specific semantic-ID tokenizers, a common generative encoder–decoder interface, and training and serving infrastructure.HiGR and TGR-Reason use PCRQ-VAE, while BARGE uses OSQ-VAE; CCFormer instead consumes raw production feature fields. TGR-Reason reuses TGR-GenRec’s backbone, and the stack uses shared training and serving systems.Source: §2.3 · p. 8

3. TGR-GenRank: Generative-Paradigm Ranking

TGR-GenRank keeps the recommendation cascade but replaces its production ranker with a Transformer that borrows generative-paradigm inputs, backbone, and training infrastructure while retaining per-item multi-task outputs. CCFormer is designed to combine full-history interaction and scalable efficiency in industrial ranking.

3.2. CCFormer Architecture in Brief

  • CCFormer replaces global self-attention with directed field-separated cross-attention, subspace token mixing, and hierarchical sequence compression.Target tokens act only as queries, enabling single-pass scoring of candidates; relative temporal and positional encoding adds recency-aware order information, and compression expands the receptive field with depth.Source: Figure 2 · p. 10, §3.2 · pp. 10–11

3.1. Motivation

  • CCFormer addresses the cost of full-history attention by separating required cross-field interactions from sequence-token interactions, making both subquadratic without retrieving or truncating history.Directed field-separated cross-attention handles user, sequence, and target interactions, while sequence-wide token mixing and hierarchical compression preserve whole-history coverage at decreasing cost.Source: §3.1 · p. 10

3.3. Training and Serving at Industrial Scale

  • Single-pass parallel candidate scoring raises online peak QPS by 30% at unchanged inference resources despite 20× higher per-sample computation.The deployment recipe also uses Numerous-Torch infrastructure, mixed-precision training that reduces memory overhead by more than 35%, and sparse-parameter compression through INT8 quantization and double hashing.Source: §3.3 · pp. 11–12

3.5. Scaling Behavior

  • CCFormer improves monotonically as industrial sequences grow from 0.5k to 2k tokens, reaching 78.17% AUC and 71.57% GAUC at 2k.Across sequence-length settings, average relative gains over HSTU were 1.08% AUC and 2.37% GAUC; with 0.5k tokens, CCFormer matched HSTU’s AUC and surpassed its GAUC with 2k tokens.Source: Figure 3 · p. 13, §3.5 · p. 13

3.4. Offline Experiments

  • CCFormer achieved the best score on every dataset and metric, beating the strongest sequential baselines on both public benchmarks and HSTU and STCA on the industrial dataset.It reached 93.67% AUC on Taobao, 0.86 points above STCA, and 83.35% on KuaiRec, 0.59 points above OneTrans. On the industrial dataset, it exceeded HSTU by +0.28 AUC and +0.50 GAUC points, and STCA by +0.21 AUC and +0.41 GAUC points.Source: §3.4 · pp. 12–13, Table 2 · p. 12, Table 3 · p. 12

3.5. Scaling Behavior

  • At matched feature dimensionality, CCFormer scales predictably while using roughly half HSTU’s GFLOPs, with 4× fewer GFLOPs at d = 256 and just 0.01 fewer AUC points.Across d = 128 to 512, AUC rose from 77.65% to 78.13% and GAUC from 70.76% to 71.52%; average relative gains over HSTU were 1.02% AUC and 1.93% GAUC.Source: §3.5 · pp. 13–14, Table 4 · p. 14
  • Ablations identify sequence compression as the efficiency lever and subspace token mixing as the main quality contributor.Removing compression leaves AUC at 77.96% but reduces training speedup from 2.21× to 1.29×. Removing mixing causes −0.21 AUC and −0.34 GAUC points; replacing it with self-attention yields only +0.02 AUC, −0.15 GAUC, and 1.41× speedup. Temporal-positional encoding adds +0.09 AUC and +0.23 GAUC points.Source: Table 5 · p. 14, §3.5 · p. 14

3.6. Online Deployment and A/B Results

  • +3.57% CTR in video recommendation and +1.71% advertising revenue accompanied CCFormer’s full launch in two of five Tencent scenarios.The five A/B-tested scenarios reported statistically significant lifts; after launch in Scenarios 1 and 2, gains remained stable across their production traffic.Source: §3.6 · pp. 14, 16, Table 6 · p. 15

4. TGR-GenRec: From Next-Token to Next-Slate Generation

TGR-GenRec advances end-to-end recommendation generation through two production paradigms: item-by-item generation with BARGE and whole-slate generation with HiGR, as intermediate steps toward replacing the cascade’s learned decision stages rather than a uniform full-pipeline replacement. HiGR uses structured semantic IDs, coarse-to-fine decoding, and listwise preference alignment.

4.2.2. Method Overview

  • BARGE restores item context with ICA, reranks paths with HPR, and adds orthogonal dual-path decoding (DPD) to address lost item boundaries and prefix-induced semantic drift without widening the beam.The modules adapt next-token prediction rather than replacing it.Source: §4.1 · p. 16, §4.2.1 · p. 19

4.2.1. Motivation: Two Structural Gaps at Industrial Scale

  • On TIGER, c3 target probability fell from 0.787 under teacher forcing to 0.015 autoregressively; c3 accuracy was 77.0% with correct prefixes versus 0.6% with erroneous ones.The results indicate that deep-layer errors arise largely from preceding path errors.Source: §4.2.1 · p. 18, Table 8 · p. 19

4.2.3. Offline Results

  • BARGE led all four Hit@K metrics in both industrial scenarios, improving Hit@5 over OneRec by 10.2% (Scenario 1) and 16.9% (Scenario 2).The gains extended to Hit@10: 6.2% in Scenario 1 and 14.7% in Scenario 2.Source: §4.2.3 · p. 21, Table 10 · p. 22

4.2.4. Efficiency and Engineering Considerations

  • BARGE was smaller than TIGER (19.91 M vs. 22.71 M parameters), while per-epoch train/infer cost grew by under 10%.A 2-layer encoder absorbs the parameters of ICA, HPR, and DPD’s dual towers, which share the encoder and run in parallel.Source: Table 11 · p. 22, §4.2.4 · p. 22

4.2.5. Deployment and Online A/B

  • In a two-week 6%-traffic A/B test, BARGE improved CTR by +0.60%, unique clicking users by +1.34%, and total reading time by +1.70% over the incumbent.BARGE was deployed as a generative retrieval channel on Tencent’s commercial media platform.Source: §4.2 · p. 17, §4.2.5 · p. 23
  • BARGE reached 9.93% CTR, versus 8.70% for Graph-based I2I, 6.94% for DNN-based retrieval, and 5.61% for content-based retrieval; its exposure adoption rate was 1.02% versus 1.03% for Graph-based I2I.This channel-level comparison measures CTR and the rate at which candidates pass through the downstream system and are adopted and exposed.Source: Table 12 · p. 24, §4.2.5 · p. 24

4.3.1. Motivation: Three Disconnects Between Item Generation and Slate Optimization

  • HiGR generates ordered slates as units, combining prefix-contrastive tokenization, hierarchical slate decoding, and listwise preference alignment; it achieves more than a 5× inference speedup over OneRec without KV caching.The design targets slate-level ordering, compatibility, and diversity, which item-wise objectives do not directly optimize.Source: §4.3 · p. 24, §4.3.1 · p. 24, Figure 7 · p. 25
  • HiGR addresses semantic-prefix, decoding-efficiency, and objective-alignment disconnects through structured SIDs, hierarchical slate decoding, and listwise ORPO.PCRQ-VAE aligns the first D−1 codebook layers for related items while preserving item identity in the leaf layer; HSD plans slate-position preferences before generating short item SID sequences, and ORPO learns from slate-level preference pairs covering ranking fidelity, genuine interest, and diversity.Source: §4.3.1 · p. 25, §4.3.2 · p. 26

4.3.2. Method Overview

  • HiGR combines a prefix-contrastive tokenizer, a hierarchical slate decoder, and listwise alignment to generate an M-item slate for holistic list quality.The planner autoregresses over one preference embedding per slate position, while a shared item generator grounds each preference in a short SID sequence; ORPO uses positive and negative slate pairs mined from implicit feedback without a reference model.Source: §4.3.2 · pp. 25–26

4.3.3. Offline Results

  • At 100M parameters, HiGR reaches 0.0831 NDCG@5 on the industrial dataset, a 41.1% gain over OneRec-25M, and leads all five metrics at 25M.At 25M, it improves Impression Hit@5 from 0.2438 to 0.2825 (+15.9%), Effective-View Hit@5 from 0.1603 to 0.1945 (+21.3%), and NDCG@5 from 0.0589 to 0.0714 (+21.2%). On KuaiRec, matched 25M HiGR improves NDCG@5 from 0.1496 to 0.1574 (+5.2%) and Effective-View Recall@5 from 0.0422 to 0.0491 (+16.4%) over OneRec; 100M HiGR raises them to 0.1651 and 0.0546, gains of 10.4% and 29.4% over OneRec-25M.Source: §4.3.3 · pp. 26–27, Table 13 · p. 27
  • Ablations show that preserving the leaf SID layer limits collision, HSD context embeddings support NDCG@5, and ORPO’s three objectives jointly deliver the best slate quality.First D−1-layer alignment yields 2.37% collision, 93% concentration, and 66.47% consistency, versus 8.73% collision when the leaf is included; removing context embedding lowers NDCG@5 from 0.0753 to 0.0664. ORPO reaches ILD 0.62 and 0.63 on the industrial dataset and KuaiRec, compared with DPO’s 0.57 and 0.60; joint objectives yield NDCG@5 0.0831 with ILD 0.62. From 0.05B to 2B parameters, loss and NDCG@5 follow power-law trends.Source: §4.3.3 · p. 27, Figure 8 · p. 28

4.3.4. Efficiency and Engineering Considerations

  • HiGR achieves more than a 5× inference speedup over OneRec-Beam while improving Recall@5 by over 5%, and keeps production P99 latency below 50 ms.The 14:2 planner-to-item-generator layer split lowers latency by 26% versus 12:4, with a 0.6% relative decrease in NDCG@5; production GPU demand falls approximately 60% at equivalent throughput. Beam search is confined to short item sequences, and batched beam decoding avoids redundant first-step calls and replicated encoder memory.Source: §4.3.4 · pp. 27–28, Figure 9 · p. 29

4.3.5. Deployment and Online A/B

  • In three Tencent A/B scenarios, HiGR increases engagement, with Average Watch Time up 1.22% and Average Video Views up 1.73% in Scenario 1.Scenario 1 also raises Average Stay Time by 1.03% and Average Request Count by 1.57%; Scenario 2 increases CTR by 0.68% and advertising revenue by 0.56%; Scenario 3 increases Average Watch Time by 1.14% and Average Video Views by 0.88%.Source: Table 14 · p. 29, §4.3.5 · p. 29

5. TGR-Reason: Reasoning-Augmented Generative Recommendation · 5.1. Motivation

TGR-Reason moves LLM-derived intent inference into offline-generated reasoning-enriched semantic IDs that guide generation and memory retrieval without request-time reasoning rollout. It targets cold-start, long-tail, and ambiguous-intent requests, where co-occurrence provides the least evidence.

5. TGR-Reason: Reasoning-Augmented Generative Recommendation

  • +477.8%: reason tokens raise cold-start new-user Hit@1 from 0.0451 to 0.2606 in the deployed direct-injection configuration.The paper reports consistent improvements across every user cohort.Source: §5 · p. 30
  • An offline Think model generates top-Kr candidate items as complete SIDs, which guide TGR-GenRec decoding and query Group Memory Retrieval.Retrieved group context is combined with the user’s Personal Memory before the generator consumes it; Figure 10 depicts these connections.Source: §5 · p. 30, Figure 10 · p. 31

5.1. Motivation

  • Co-occurrence-based generation works when history constrains the next item, but falls short for new users, sparse long-tail items, and ambiguous-intent sessions.These requests are underdetermined by co-occurrence and call for external world knowledge and intent inference.Source: §5.1 · p. 30
  • Reason tokens bridge incomplete Personal Memory and richer Group Memory by selecting intent-consistent behavioral evidence rather than retrieving popular but irrelevant behavior.They infer plausible interests beyond the observed sequence, allowing population behavior to supplement rather than overwrite individual history.Source: §5.1 · p. 30
  • OneRec-Think [10] emits textual reasoning before each SID, while ReaRec [11] and STREAM-Rec [41] perform multi-step reasoning at inference, tying capacity to per-request computation.OneRec-Think emits hundreds of extra tokens per request; the cited approaches incur inference-time reasoning computation that industrial latency budgets cannot absorb.Source: §5.1 · p. 30
  • TGR-Reason installs reasoning through training-time supervision and periodically generates reason tokens offline, amortizing its cost instead of running an LLM rollout on every impression.Its Think model builds on LatentRec; the paper describes it as among the first to bring training-time-amortized reasoning to an industrial generative recommender [25].Source: §5.1 · p. 30, §5 · p. 30

5.2. Method Overview

TGR-Reason exports offline-generated complete-SID reason tokens as a shared interface for recommendation decoding and memory retrieval. TGR-GenRec combines these tokens with personal and population behavior without extending autoregressive SID decoding.

5.2. Method Overview

  • TGR-Reason generates reason tokens offline, while Personal Memory captures user behavior and Group Memory retrieves intent-consistent population evidence for TGR-GenRec.DRI injects reason-token representations; GMR retrieves and encodes Group Memory for adaptive fusion with Personal Memory. Both paths add no autoregressive reasoning step.Source: §5.2 · p. 31

5.2.1. Reason Tokens as a Shared Reasoning–Memory Interface

  • Complete-SID reason tokens provide a shared-vocabulary interface: they prompt TGR-GenRec decoding and query Group Memory, while latent reasoning remains internal to training.Unlike chain-of-thought text, they require no online reasoning rollout or natural-language rationale across module boundaries.Source: §5.2.1 · pp. 31–32

5.2.2. Offline Reason Token Generation with LatentRec

  • LatentRec trains the Think model with per-step latent supervision, then uses K=0 standard SID decoding to generate offline reason tokens without a production reasoning rollout.Soft Token Selection keeps feedback within the input-embedding manifold; raw hidden-state feedback with per-step head supervision diverges within the first training epoch. Per-step Reasoning Loss [11] applies the shared prediction head at every latent step.Source: §5.2.2 · pp. 32–34

5.2.3. Reasoning and Memory Injection into TGR-GenRec

  • DRI turns reason tokens into Personal-Memory-conditioned, hierarchy-aligned prompts; GMR uses each complete SID to retrieve frequent successor items and adaptively fuse population evidence with Personal Memory.The paths preserve the number of SID decisions and the beam-search space; Group Memory is encoded once per request and reused across decoding steps and beams.Source: §5.2.3 · pp. 34–36

5.3. Training Recipe

Section 5.3 trains the Think model in stages, uses its precomputed reason tokens to condition standard SID generation, and refreshes those tokens through an offline pipeline. A batched decoding schedule makes large-scale refresh practical while preserving the reference decoder’s Hit@5.

5.3.1. Think-Model Training

  • Think-model training expands the vocabulary, aligns content to the target surface, then trains only LoRA adapters and the Continuous Thought Module while freezing the backbone.Reasoning training uses Ktrain=2, λ=0.1, Tb=1, and Ts=5; chain-of-thought-distilled initialization is used only for deeper rollout.Source: §5.3.1 · p. 36

5.3.2. Gen-Model Training

  • The Gen-Model uses teacher-forced next-token prediction with precomputed reason tokens organized by SID level and conditioned on Personal Memory; gradients do not update the Think model.When GMR is enabled, retrieved context conditions generation and its Group Encoder and adaptive fusion gate are jointly optimized; no reasoning-specific objective is added.Source: §5.3.2 · p. 37

5.3.3. Offline Generation at Scale

  • A weekly pipeline retrains and versions the Think model, while a daily pipeline refreshes top-Kr reason tokens for users flagged by a staleness detector.Weekly retraining uses a fixed-size sample; separating retraining and prediction cadences lets freshness and compute cost be tuned independently.Source: §5.3.3 · p. 37
  • Batched single-token rounds reproduce standard beam enumeration, while continuous batching and prefix caching accelerate decoding without modifying the inference engine.Native sampling loses top-Kr recall and native beam search is slower than the eager baseline; the SID trie is used only as a final validity check.Source: §5.3.3 · pp. 37–38
  • The batched schedule matches reference-decoder Hit@5 bucket by bucket and cuts the end-to-end generation cycle 2.2×, from 174 to 80 minutes.On eight A100 GPUs, the ten-million-user refresh falls from about 2.9 days to 1.3, keeping reasoning offline-amortized rather than a per-request cost.Source: §5.3.3 · p. 38

5.4. Deployment

TGR-Reason keeps expensive user-level reasoning off the synchronous request path by materializing reusable Reason records, which online serving can consume while asynchronous jobs refresh and backfill them. Its deployment combines generation conditioning, direct SID retrieval, and ranking features with version-aware storage and failure isolation.

5.4. Deployment

  • Nearline refreshes and offline backfills conditionally update reusable Reason records, so newly materialized outputs serve subsequent requests rather than the request that triggered them.Request-triggered jobs are admitted under gain, freshness, and budget constraints; scheduled jobs recover missing, stale, failed, or version-incompatible records.Source: §5.4 · p. 38, §5.4.2 · pp. 40–41

5.4.1. Serving Efficiency

  • Pre-materialized Reason records provide DRI soft prompts and GMR memory, while stored top-k SIDs enable decoder-free retrieval without adding SID decoding steps.These bounded operations leave the SID decoding horizon, beam width, and search space unchanged; direct-retrieval and generated candidates receive separate quotas before ranking.Source: §5.4.1 · pp. 38–39

5.4.2. Online Serving

  • A valid Reason record conditions TGR-GenRec, supplies features to CCFormer, and enables direct retrieval; a miss or invalid record falls back to baseline serving without blocking the request.Nearline triggers run fire-and-forget, and the current request neither waits for nor consumes the refreshed record.Source: §5.4.2 · p. 40

5.4.3. Store Consistency and Failure Isolation

  • Version-aware conditional writes accept only compatible records and, within a version, replace stored data only when its behavior cutoff is newer, preventing stale jobs from overwriting fresher results.Failures in trigger, nearline, or offline tiers reduce freshness or delay coverage recovery, but do not interrupt online requests, which can use a compatible record or fall back to no-Reason generation.Source: §5.4.3 · p. 41

5.5. Evaluation Results

TGR-Reason improves Hit@K across traffic cohorts, with its largest offline gains among cold-start users, and delivers significant online lifts in consumption and new-user conversion.

Offline results.

  • Hit@1 rose 28.2% across all recommendation traffic and 58.6% in primary recommendation traffic, with the benefit strongest near the list head.Across all traffic, Hit@1 increased from 0.3385 to 0.4339; in primary traffic, it increased from 0.2247 to 0.3563. TGR-Reason improved every reported cutoff in each cohort, while gains narrowed as K grew.Source: Offline results · p. 42
  • +477.8% relative Hit@1 gain among cold-start users in primary recommendation traffic was the largest reported cohort improvement.Hit@1 rose from 0.0451 to 0.2606; Hit@5, Hit@10, Hit@20, and Hit@50 improved by 39.2%, 24.9%, 20.1%, and 11.6%, respectively. Reason-token semantic priors help narrow the target space when behavioral history is sparse.Source: Offline results · p. 42
  • +1.75% Effective Consumption Rate and +13.09% Exposure-to-Conversion Rate for new users were statistically significant online lifts over the incumbent production generator.These gains show that offline improvements carried through the complete production cascade to measurable user-value gains.Source: Offline results · pp. 42–43, Table 17 · p. 43

6. Conclusion, Limitations, and Future Directions

TGR combines generative-paradigm ranking, end-to-end generation, and reasoning-augmented generation in an industrial recommendation stack. The paper reports production gains while identifying current limits in model unification, cold-start items, global objective weighting, and request-time reasoning, alongside three future directions.

6.1. Conclusion

  • TGR deploys CCFormer, BARGE, HiGR, and TGR-Reason across Tencent surfaces; CCFormer lifts CTR by +3.57% and advertising revenue by +1.71%, while TGR-Reason raises new-user Hit@1 by +477.8%.TGR-Reason also delivers statistically significant online gains of +1.75% Effective Consumption Rate and +13.09% new-user Exposure-to-Conversion Rate; the stack serves hundreds of millions of users.Source: §6.1 · p. 43

6.2. Limitations

  • TGR remains a coupled stack: CCFormer emits per-item multi-task scores for downstream cascade stages, while BARGE and HiGR ship with separate OSQ-VAE and PCRQ-VAE tokenizers.The generative models share a user/context encoding interface and train and serve on Numerous-Torch, but the stack is not yet one model.Source: §6.2 · p. 43, §6.1 · p. 43
  • Cold-start items remain limited by tokenizer quality because new items enter hierarchical SID space from semantic evidence before collaborative relations shaping SID prefixes can be estimated.The reported TGR-Reason cold-start gains concern new users, not cold-start items.Source: §6.2 · p. 43
  • HiGR’s ORPO alignment uses one global weighting for ranking fidelity, genuine user interest, and diversity; personalized or context-conditioned weighting remains open.The current weighting is uniform across users and contexts.Source: §6.2 · p. 44
  • TGR-Reason amortizes reasoning offline: Think is retrained weekly, reason tokens are regenerated daily or within seconds for trigger-gated users, and serving reads precomputed tokens at K=0.Intra-session intent shifts reach the generator through input context, while explicit in-text reasoning remains too expensive on the request path.Source: §6.2 · p. 44

6.3. Future Directions

  • TGR will extend HiGR’s observed 0.05B–2B scaling range, explore a unified lazy decoder, and co-design sparse attention, sequences, and topology within a < 50 ms P99 fixed-GPU envelope.The passage cites ULTRA-HSTU’s over 5× faster training scaling and 21× faster inference scaling through model–system co-design [43].Source: §6.3 · p. 44
  • TGR plans selective request-time deliberation with K>0 decoding for high-value or cold-start requests, latency-tolerant reasoning variants, and memory beyond the current Personal/Group split.The passage situates this direction alongside OneRec-Think [10], ReaRec [11], and RecGPT-V3 [44].Source: §6.3 · p. 44
  • TGR will pursue agentic self-evolution, building on reported recommender agents that propose, screen, and A/B-validate improvements across optimizers, architectures, and reward functions [45].The passage says these agents have surpassed conventional engineering workflows in development velocity and model performance; surveys chart the broader trajectory [46].Source: §6.3 · p. 44

Appendix · B. Notation

The notation appendix distinguishes report-wide stack symbols from chapter-specific notation, which must be interpreted within its stated scope.

B. Notation

  • Table 18 summarizes stack-level notation and the principal model-specific symbols used in each technical chapter.Source: B. Notation · p. 45
  • The production models are presented as independent, parallel works in §4 and follow the notation of their source papers.Source: B. Notation · p. 45
  • L denotes per-item SID depth in §4.2 and §5, but interaction-block count in §3; elsewhere, per-item SID depth is denoted D.Source: B. Notation · p. 45
  • Symbols not listed in the appendix are local to the subsection where they appear.Source: B. Notation · p. 45

C. Default Configurations and Hyperparameters

The section provides a consolidated reference for default settings across three production models and TGR-Reason, drawing on source papers [22] [23] [24] and §5 while omitting unpublished values.

C. Default Configurations and Hyperparameters

  • Table 19 consolidates defaults for three production models and TGR-Reason from [22] [23] [24] and §5, omitting unpublished values such as HiGR’s ILD regularization weight β.The body sections provide context for each setting.Source: C. Default Configurations and… · p. 45

D. Complexity Analysis

The complexity analysis attributes efficiency to CCFormer’s structured sequence processing, BARGE’s fixed-width correction of semantic drift, and HiGR’s decomposed slate decoding.

CCFormer (§3)

  • At matched capacity, CCFormer uses roughly half the per-sample GFLOPs of HSTU by replacing global attention with local grouping, linear-cost token mixing, and progressively compressed sequences.Relative temporal-position encoding costs O(L_s m), and strided convolution shortens the sequence between blocks while preserving access to the full history.Source: D. Complexity Analysis · p. 45

BARGE (§4.2)

  • BARGE corrects semantic drift at fixed beam width because latency scales approximately linearly with beam width and decoding memory grows with B × L.HPR adds lightweight per-layer dual-tower scoring with a once-per-request context projection; DPD runs two decoder towers in parallel without changing per-channel beam width or final Top-K length, and combined parameters remain below TIGER.Source: D. Complexity Analysis · p. 45

HiGR vs. OneRec-style decoding (§4.3)

  • HiGR reduces decoding complexity from B·M^3D^3·l_slate·d to M^3·l_slate·d + B·M D^3·l_item·d through greedy slate planning and local per-item beam search.The slate-planner depth matches the OneRec-style baseline’s lazy-decoder depth.Source: D. Complexity Analysis · p. 45

Related papers

Loading 2609.00986v1…