PaperlayerView exact revision on arXiv

Source-linked AI summary

OneLatent: Latent Reasoning for Efficient Foundation Recommendation Models

Hao Jiang, Peiru Du, Pengfei Yao, Mengting Li, Siyuan Lou, Kuo Cai, Sheng Yu, Qiang Luo, Jian Liang, Ruiming Tang, Fei Pan, Peng Jiang, Wenwu Ou

arXiv:2610.30675v1cs.IR

TL;DR

Explicit CoT reasoning can be costly and may miss diverse, changing user interests. OneLatent trains a compact latent-token reasoning interface from multi-view rationales, and it outperforms explicit-CoT and conventional baselines across the reported recommendation benchmarks. An online A/B test also indicates a revenue lift in Kuaishou local-services advertising.

  • Problem

    Explicit CoT can generate lengthy, redundant traces that increase inference latency and reduce throughput, while fixed single-path reasoning can overlook short-term, context-dependent interests.

    Source: §1 · p. 2

  • Method

    OneLatent uses multi-view adaptive rationales, progressive latent-token alignment, and curriculum post-training to support latent-reason-then-answer recommendation.

    Source: §1 · p. 2

  • Results

    Across the industrial Kuaishou dataset and public Kuaishou LLM-Rec benchmark, OneLatent outperforms explicit-CoT and conventional baselines; industrial SID@64 improves by 17.44% over matched Think variants.

    Source: §5 · p. 9

  • Takeaways & Limitations

    The online A/B test indicates an estimated 9.6% revenue lift over strong online baselines in Kuaishou local-services advertising.

    Source: §1 · p. 2

    The semantic reconstruction analysis measures association, not step-by-step faithfulness, and does not prove that latent IDs form a strict information bottleneck.

    Source: §4.4.3 · p. 9

Abstract

from arXiv · show

Large language models (LLMs) have demonstrated strong reasoning capabilities, motivating their use as the backbone of foundation recommendation models (FRMs). Existing methods enhance recommendations through explicit Chain-of-Thought (CoT) reasoning under a Think-then-Answer paradigm. However, explicit CoT incurs substantial inference overhead by generating lengthy reasoning traces and relies on manually designed templates that struggle to capture diverse, dynamic user interests. We propose OneLatent, an efficient latent reasoning framework that compresses explicit reasoning traces into several learnable latent tokens, enabling Latent-Reason-then-Answer inference without generating verbose traces. OneLatent first introduces Multi-View Adaptive CoT (MV-ACoT), which creates diverse, high-quality teacher-generated supervision by exploring user interests from multiple perspectives and automatically adapting reasoning complexity to each instance. Building on pretrained FRMs, it then uses a three-stage latent-token alignment paradigm to progressively internalize CoT traces into learnable latent tokens. Finally, a multistage curriculum-based post-training strategy activates latent-token reasoning for downstream recommendation tasks. Experiments on an industrial-scale Kuaishou dataset and the public Kuaishou LLM-Rec benchmark show that OneLatent consistently outperforms explicit CoT-based methods and traditional baselines. Compared with the Think and No-Think variants of FRMs, OneLatent improves SID@64 by 17.44% and 9.33%, respectively, while achieving over 17x higher online inference throughput. We further develop a production serving system for scalable, real-time FRM inference. An online A/B test in Kuaishou's local-services advertising scenario shows that deploying OneLatent with this system yields an estimated 9.6% revenue lift over strong online baselines, including OneRec and OneReason.

1 Introduction

OneLatent replaces verbose explicit reasoning with a compact latent-token interface for foundation recommendation models, addressing the latency and interest-drift challenges of explicit CoT. It reports stronger recommendation results and throughput, alongside an online revenue lift in Kuaishou local-services advertising.

1 Introduction

  • Explicit-CoT FRMs incur latency from redundant traces and can overlook short-term interests, while prior latent reasoning can lack interpretable, coherent task-relevant reasoning.The introduction also notes that generic CoT patterns do not directly fit diverse, evolving user interests.Source: §1 · p. 2
  • OneLatent inserts three learnable latent tokens between user context and target SID, using MV-ACoT supervision, staged rationale alignment, and curriculum post-training.Its production serving architecture prioritizes valuable asynchronous updates under limited GPU capacity and keeps foundation-model inference off the online request path.Source: §1 · p. 2
  • On the industrial dataset, OneLatent improves SID@64 by 17.44% over explicit Think variants and 9.33% over No-Think variants, with 17.49× throughput over explicit rationale decoding.The experiments also include the public Kuaishou LLM-Rec benchmark and a 9B backbone evaluation.Source: §1 · p. 2
  • The online A/B test indicates an estimated 9.6% revenue lift over strong online baselines, including OneRec and OneReason, in Kuaishou local-services advertising.The production serving system uses behavior-triggered admission, capacity-aware scheduling, and asynchronous result materialization.Source: §1 · p. 2

2 Related Work

Related work traces foundation recommendation models to recommendation-native backbones and pretrained-LLM adaptation, then contrasts latent and explicit reasoning approaches. It identifies limits in supervision for user needs and in the cost and interface sensitivity of explicit reasoning.

2.1 Scaling Recommendation with LLMs

  • Foundation recommendation models span recommendation-native backbones and pretrained-LLM adaptation, but mainly model behavior and item generation, with limited supervision for latent needs and recommendation decisions.Examples include TIGER, HSTU, and OneRec among recommendation-native backbones, and LC-Rec and RecGPT among LLM-adapted models.Source: §2.1 · p. 2

2.2 Reasoning-Enhanced Recommendation

  • Latent methods reduce rationale costs but often lack structured recommendation supervision, while explicit reasoning methods remain costly and sensitive to natural-language and item-ID interfaces.The cited approaches include hidden-space refinement, industrial ranking modules, distilled reasoning states, and supervised or reinforcement-learning CoT methods.Source: §2.2 · pp. 2–3
  • OneLatent addresses these gaps with complementary multi-view rationales and staged alignment for SID prediction.The section frames this as training a latent-token-conditioned interface.Source: §2.2 · p. 3

3 Methodology

OneLatent uses multi-view, complexity-adaptive teacher supervision and staged latent-token training to predict recommendation SIDs without generating verbose rationales at inference. Its production serving design runs inference asynchronously and merges its candidates with the incumbent retrieval pipeline.

3.1 Problem Formulation and Framework

  • The framework adds three fixed latent-token IDs shared across samples, whose contextualized hidden representations support subsequent rationale or SID tokens through causal self-attention.The tokens are not sample-specific latent vectors.Source: §3.1 · p. 3

3.2.1 Multi-View CoT Task Design.

  • Exploration supports open-ended demand discovery and interest-shift analysis, while evaluation uses only the real positive target, adds no sampled negatives, and retains a rationale plus a discrete score where larger means stronger user–item consistency.Source: §3.2.1 · pp. 3–4

3.2.2 Adaptive CoT Generation.

  • Prompt-guided adaptation asks the teacher to reason quickly for simple cases and compare multiple dimensions for complex cases, selecting dimensions only when the instance supports them.The policy uses no separately trained difficulty classifier, difficulty label, or hand-set threshold.Source: §3.2.2 · p. 4

3.3 Latent Token Alignment

  • Three stages progress from reproducing single- and multi-view teacher outputs after latent tokens to predicting the next item SID from user profile, history, and request context.The target SID is excluded from the Stage III input.Source: §3.3 · p. 4, §3.3.3 · p. 4, §3.3.4 · p. 5

3.3.1 FRMs Pre-training.

  • Pretraining represents each item with a three-level Semantic ID and bidirectionally aligns that sequence with the item’s dense caption, alongside tasks covering recommendation and general-domain knowledge.The added tasks include spatiotemporal context and user-persona understanding.Source: §3.3.1 · p. 4

3.3.3 Stage II: Multi-View Latent Reasoning Alignment.

  • Stage II trains a shared latent-token-conditioned interface across available views, averaging token losses within each output so longer traces do not automatically receive greater weight.Student prefixes exclude candidate descriptions, positive-item descriptions, and target SIDs, even when the teacher used the positive item for supervision.Source: §3.3.3 · pp. 4–5

3.4.1 Multi-stage Curriculum-based Recommendation SFT.

  • Curriculum-based recommendation SFT moves from high-activity to low-activity users, then latent-reasoning SFT mixes latent and conventional recommendation samples at a 1:1 ratio.The latent mode predicts the target SID after three latent tokens while conventional recommendation performance is retained without them.Source: §3.4.1 · p. 5, §3.4.2 · p. 5

3.4.2 Latent Reasoning-based Recommendation SFT.

  • At inference, the FRM decodes only the target SID after three latent tokens, with fixed rather than rationale-length-dependent reasoning-token overhead.The teacher, view tokens, privileged item information, and textual rationale are absent.Source: §3.4.2 · p. 5

4 Experiments

The experiments assess OneLatent on public and industrial recommendation data, comparing its recommendation quality, reasoning supervision, inference efficiency, and online performance with established baselines. Across these evaluations, results favor multi-view adaptive supervision and latent reasoning, while the analysis also examines interpretability, token count, and general capabilities.

4.1.1 Datasets and Models.

  • Evaluation covers the public advertising benchmark and seven days of Kuaishou local-services interactions, using ID-based, SID-based, and reasoning-enhanced FRM baselines.Experiments primarily use Qwen3.5-0.8B and also evaluate a 9B checkpoint; metrics include SID and ID Hit@K.Source: §4.1.1 · p. 6, §4.1.2 · p. 6, §4.1.3 · p. 6

4.2 Overall Results (RQ1)

  • At 9B, OneLatent reaches SID@64 0.2008, 8.13% above the strongest reported non-latent 9B result of 0.1857.The matched-size comparison uses MV-ACoT–Merge supervision for No-Think, Think, and OneLatent.Source: §4.2 · pp. 6–7, Table 3 · p. 7
  • At K=64, OneLatent–Merge gains 12.21% and 19.10% on public SID and ID retrieval, and 1.58% and 0.45% on industrial SID and ID retrieval over the strongest non-OneLatent results.The separate diagnostic comparison reports industrial SID@64 gains of 17.44% over Think–OR and 9.33% over No-Think–OR.Source: §4.2 · p. 6, Table 2 · p. 7

4.3.1 LLM-as-a-Judge Evaluation.

  • The judge evaluation supports multi-view construction: Exploration, Evaluation, and Attribution respectively discover intent, match positive items to evidence, and explain known-positive interactions.Together, the views address distinct reasoning objectives rather than repeating one objective.Source: §4.3.1 · pp. 7–8, Table 4 · p. 7

4.3 CoT Analysis (RQ2)

  • Adaptive CoT reduces normalized average reasoning length by 35%, with approximately 0.92 semantic similarity on the longest-reasoning cases; its downstream results improve public metrics and remain comparable industrially.Single views have dataset- and metric-dependent results, while MV-ACoT-Merge combines complementary objectives and OneLatent outperforms explicit Think under the compared CoT strategies.Source: §4.3.2 · p. 8, §4.3.3 · p. 8, Figure 4 · p. 8

4.4.1 Recommendation Performance.

  • Relative to explicit CoT, OneLatent raises Hit@L1 from 0.354 to 0.401 and achieves the best Hit@L2 and Hit@L3.The full latent-token-conditioned training pipeline is evaluated; the result does not isolate whether gains reside only in latent-token positions or across the fully updated backbone.Source: §4.4.1 · p. 8, Table 6 · p. 8

4.4.2 Inference Efficiency.

  • OneLatent achieves a 17.49× QPS gain over explicit CoT while matching UnCoT latency and throughput by decoding the SID after three fixed latent token IDs.The efficiency comparison uses 872 test instances.Source: §4.4.2 · p. 8, Table 7 · p. 8

4.4.3 Interpretability.

  • View-conditioned reconstructions have cosine similarities of 0.788–0.804 with teacher CoT, but this association does not establish step-by-step faithfulness or a strict information bottleneck.Three latent tokens balance reasoning capacity and compactness; latent supervision also yields higher scores across evaluated MMLU categories while improving recommendation performance.Source: §4.4.3 · pp. 8–9, Table 8 · p. 9, §4.4.4 · p. 9

5 Conclusion

OneLatent distills CoT supervision into latent reasoning and outperforms explicit-CoT and conventional baselines across two Kuaishou benchmarks. Its industrial deployment combines higher inference throughput with an estimated 9.6% revenue lift in local-services advertising.

5 Conclusion

  • SID@64 rises 17.44% versus Think and 9.33% versus NoThink, with more than 10× higher inference throughput on the industrial dataset.Across the industrial Kuaishou dataset and public Kuaishou LLM-Rec benchmark, OneLatent outperforms explicit-CoT approaches and conventional baselines.Source: §5 · p. 9
  • The production system schedules FRM computation outside the critical online request path, then routes generated candidates through existing retrieval and fine-ranking.In Kuaishou local-services advertising, deployment yields an estimated 9.6% revenue lift over a strong baseline.Source: §5 · p. 9
  • Further analysis indicates that OneLatent strengthens recommendation reasoning without eroding general capabilities.Source: §5 · p. 9

A.1 Implementation Details

Implementation uses full-parameter supervised fine-tuning with standard causal attention and a distributed training setup. The configuration specifies optimization, precision, sequence length, and evaluation-budget details.

A.1 Implementation Details

  • All stages use the standard causal attention mask, without a special attention bottleneck, auxiliary encoder, projector, or rationale decoder.Source: §A.1 · p. 11
  • Training runs across seven nodes with eight GPUs each, using per-device batch size 4 and one accumulation step for an effective global batch size of 224.Source: §A.1 · p. 11
  • The model trains for two effective epochs in BF16 with AdamW, peak learning rate 1.0 × 10−4, and a maximum sequence length of 4,096 tokens.Evaluation provides enough Think-generation budget for the complete rationale and final SID; no SID truncation due to rationale length was observed.Source: §A.1 · p. 11

A.2 Supervised Dataset Construction

The supervised dataset turns temporally ordered user-item interactions into examples that predict each converted item's hierarchical SID from the user's preceding history and profile. Histories are bounded and behavior-aware to emphasize recent interests and distinguish feedback strength.

A.2 Supervised Dataset Construction

  • Interactions are sorted chronologically, and duplicate items retain only their strongest associated signal, prioritized as conversion, click, long view, then impression.Source: §A.2 · p. 11
  • Each conversion becomes a target, with only preceding interactions used as context, so users with multiple conversions can contribute multiple examples without target leakage.Source: §A.2 · p. 11
  • Each input combines a compact user profile, local-life demand, explicit intent, and at most the latest 50 interactions represented by hierarchical SIDs and behavior types.The target is the hierarchical SID of the next converted item.Source: §A.2 · p. 11
  • Examples with unavailable user profiles or target SIDs are discarded, and historical items without valid SID mappings are filtered from the input.Valid samples are stored in sharded JSONL files for memory-efficient streaming and distributed training.Source: §A.2 · pp. 11–12

A.3 CoT Supervision Construction

CoT supervision is generated by a teacher through three complementary perspectives with distinct inputs and outputs. The construction configuration is documented in Table 12.

A.3 CoT Supervision Construction

  • Exploration uses the user profile and historical interactions to generate a rationale and multiple potential intents without observing a candidate or target.Source: §A.3 · p. 12
  • Evaluation uses the observed positive item as teacher-side guidance to generate a rationale and a 1–5 matching score, without sampling negative candidates.Source: §A.3 · p. 12
  • Attribution receives the observed positive item and generates a rationale explaining supporting evidence, while its description and SID stay out of the Stage I/II student prefix.Table 12 is identified as the configuration for constructing CoT supervision.Source: §A.3 · p. 12, Table 12 · p. 12

B Latent Reconstruction Case Study

The reconstruction shares key commercial semantics with explicit CoT, while the prompt design grounds teacher reasoning in behavioral evidence and adapts its depth to user-signal complexity.

B Latent Reconstruction Case Study

  • The reconstruction aligns with explicit CoT on commercial semantics, with approximately 0.80 embedding similarity, despite differing on geographic accessibility.Shared semantics include men’s-health interest, lead-generation intent, price sensitivity, offline fulfillment, and dense L3 behavioral signals.Source: B Latent Reconstruction Case Study · p. 12
  • The case study does not prove that latent token IDs alone store the rationale or form a strict information bottleneck.The user profile remains visible and the whole checkpoint is fine-tuned.Source: B Latent Reconstruction Case Study · p. 12

C MV-ACoT Prompt Design

  • Business-aware prompt constraints ground teacher reasoning in local-service lead-ad needs and prioritize strong behavioral evidence over weak exposures.The prompts translate content interests into actionable commercial demands rather than relying on generic preferences.Source: C MV-ACoT Prompt Design · p. 12
  • MV-ACoT adapts reasoning paths to signal strength, consistency, and uncertainty, selecting relevant dimensions and comparing plausible explanations for complex cases.This is in-prompt self-assessment, without an external difficulty classifier, binary label, threshold, or reported classification accuracy.Source: C MV-ACoT Prompt Design · p. 12

D CoT Quality Evaluation Prompt

The evaluation prompt uses an LLM judge to assess teacher-generated traces against user evidence and task objectives, with multidimensional scoring and explicit factuality caps.

D CoT Quality Evaluation Prompt

  • The LLM judge receives the task type, history, profile, generated trace, and— for Evaluation or Attribution—the positive item.It judges whether each trace derives a calibrated judgment from available user evidence while respecting its view’s objective.Source: D CoT Quality Evaluation Prompt · p. 12
  • The rubric evaluates factuality, evidence selection, intent translation, causal logic, advertising fit, spatiotemporal reasoning, confidence calibration, and task fulfillment.Tables 13 and 14 summarize task-specific requirements and scoring dimensions, each rated on a five-point scale with quality anchors.Source: D CoT Quality Evaluation Prompt · p. 12, Table 13 · p. 13, Table 14 · p. 14
  • Each instance supplies task type, history, profile, and reasoning trace; Evaluation and Attribution also include the positive item, with one 1–5 score and reason per dimension.The output format requires one JSON object and permits no other fields or markdown.Source: D CoT Quality Evaluation Prompt · p. 12
  • Fabricated behaviors absent from history cap factuality at 2, while fabricated locations, times, stages, or needs cap it at 3.Backfilling candidate or target information into history also caps factuality at 2; the constraints prioritize evidence-grounded reasoning.Source: D CoT Quality Evaluation Prompt · pp. 12–13

Related papers

Loading 2610.30675v1…