Source-linked AI summary
Latent Residual Reasoning for Ranking in Generative Recommendation
Haihan Shi, Ruidong Han, Wenjie Wang, Shangyu Chen, Sunhao Dai, Yongqi Li, Yangyang Li, Fei Jiang, Wei Lin, Fuli Feng, Xiangnan He
TL;DR
Generative ranking needs fine-grained scores for multiple candidates, yet existing models largely rely on one-shot prediction, and how latent reasoning should refine these scores remains less explored. Refiner uses cumulative residual score corrections, history queries, and parameter-efficient adaptation; experiments report consistent improvements over evaluated baselines across public and industrial settings. The evaluation remains limited by the number of directly comparable methods and the production setup available for comparison with OnePiece.
Problem
Generative ranking requires fine-grained scores for explicit candidates, but models mainly use one-shot prediction and the role of successive latent-reasoning steps in refining scores remains less explored.
Source: §1 · p. 2Method
Refiner accumulates additive score corrections across reasoning steps, uses history queries for compact global user context, and adapts a frozen base ranker with LoRA modules and a reasoning prediction layer.
Source: §1 · p. 2, §3.1 · p. 3, §3.2 · p. 4Results
Refiner shows consistent improvements over evaluated baselines across two public datasets, a large-scale offline industrial dataset, and an online A/B test, with gains across different user and item groups.
Source: §1 · p. 2, §6 · p. 10Takeaways & Limitations
The evaluated results support Refiner’s practical effectiveness in the production setting, including the offline industrial evaluation and online A/B test.
Source: §6 · p. 10Direct comparison is limited by task-interface and backbone differences, and the current production setup lacks the interaction and contextual inputs needed for a fair comparison with OnePiece.
Source: §6 · p. 10
Abstract
from arXiv · showhide
Generative recommendation (GR) models for ranking typically adopt a fast-thinking direct prediction paradigm with strong practical efficiency. However, ranking requires fine-grained scoring across multiple candidate items, while existing GR models still mainly rely on one-shot prediction. This raises the question whether additional inference-time computation can further improve ranking quality. Recently, inspired by Chain-of-Thought (CoT) reasoning in large language models (LLMs), some work has begun to introduce reasoning into recommendation systems. Some of these methods rely on LLM reasoning frameworks to perform explicit reasoning in textual space, but such designs are not suitable for more widely deployed non-textual recommendation models. A more relevant direction is to introduce latent reasoning into ID-based recommendation models. Although recent work has extended latent reasoning to industrial ranking, how successive reasoning steps should improve fine-grained candidate scores remains less explored. To address this problem, we propose Refiner, a latent residual reasoning framework with parameter-efficient fine-tuning for generative ranking. Our approach introduces three key components: (1) a residual reasoning mechanism that progressively refines ranking scores through cumulative additive corrections, (2) history queries that provide condensed user behavior representations to enhance candidate-specific reasoning, and (3) a parameter-efficient fine-tuning strategy that leverages LoRA modules to adapt pre-trained models for latent reasoning. Extensive experiments on two public benchmarks, an offline industrial dataset, and an online A/B test demonstrate that Refiner consistently improves over direct prediction, validating its effectiveness and practical utility.
1 Introduction
Generative ranking models score multiple candidates efficiently, but typically rely on one-shot prediction, leaving open whether added inference-time computation can refine scores. Refiner addresses challenges in adapting latent reasoning to candidate ranking, preserving global user context, and enabling parameter-efficient adaptation.
1 Introduction
- Ranking needs fine-grained scores for explicit candidates, but existing generative ranking methods mainly produce them in a single inference step.The open question is whether additional inference-time computation can refine those scores.Source: §1 · p. 2
- Textual LLM reasoning carries computational overhead and is less applicable to widely deployed non-textual recommenders, motivating latent reasoning in Transformer-based ID models.Prior latent-reasoning work targets semantic-ID generation or next-item recommendation; OnePiece demonstrates industrial ranking via block-wise candidate-representation refinement and progressive multi-task supervision.Source: §1 · p. 2
- Latent reasoning for ranking must progressively improve candidate scores under the ranking objective, rather than optimize objectives for semantic-ID generation or next-item recommendation.Candidate-specific reasoning also needs access to global user information, which is central to retrieval-oriented latent reasoning.Source: §1 · p. 2
- Refiner accumulates additive score corrections, uses history queries to retain global user context during candidate-specific reasoning, and adapts pretrained models with lightweight LoRA modules.Its fine-tuning framework freezes the pretrained base model and adds a dedicated reasoning prediction layer.Source: §1 · p. 2
- Experiments on public benchmarks, an offline industrial dataset, and an online A/B test show consistent improvements, including gains for cold-start items and long-sequence users.The work implements Refiner on two ranking models, including the deployed industrial model MTGR.Source: §1 · p. 2
2 Preliminary
Generative ranking organizes user and candidate features as a token sequence, then scores multiple candidates in one forward pass while masking interactions between candidates. The base model learns task-specific scores with binary cross-entropy supervision; latent reasoning adds continuous hidden-space computation to correct candidate scores.
2 Preliminary
- The input comprises user profile features, a historical interaction sequence, and K candidate items, whose categorical and numerical features are projected into dense embeddings.The embedded features are unified into a single token sequence of length L and embedding dimension d.Source: §2 · p. 3
- A candidate mask lets each candidate attend to user-side information while preventing attention between different candidates, enabling parallel scoring without candidate-to-candidate leakage.Transformer blocks process the masked sequence, and candidate-position representations pass through a prediction module.Source: §2 · p. 3
- The base model produces task-specific scores for K candidates across T tasks and is optimized with binary cross-entropy against the task labels.The labels form a K × T binary ground-truth matrix.Source: §2 · p. 3
- Latent reasoning adds computation in continuous hidden space without textual traces, using the resulting states to produce successive corrections to candidate scores.This connects the ranking formulation to Refiner’s score-refinement process.Source: §2 · p. 3
3 Method
Refiner adds latent reasoning to generative ranking through candidate-specific intermediate states, condensed history queries, parameter-efficient adaptation, and cumulative score corrections. Its three components are input construction, fine-tuning, and residual reasoning.
3 Method
- The method comprises input construction with reasoning tokens and history queries, parameter-efficient fine-tuning, and residual reasoning for progressive score refinement.Figure 2 illustrates the three-stage framework, including autoregressive steps for score correction.Source: §3 · p. 3, Figure 2 · p. 5, §3.3 · p. 5
3.1 Input Construction
- Candidate-specific reasoning tokens carry intermediate states, while history queries provide a compact representation of global user behavior.Reasoning tokens are generated autoregressively from prior-step hidden states, and the fixed history queries are reused across reasoning steps.Source: §3.1 · p. 3
- The extended mask lets reasoning tokens access their corresponding candidate, earlier reasoning tokens, and user-side information while preventing information leakage.History queries attend to user history and retain a causal mask within the query segment.Source: §3.1 · p. 3, Figure 1 · p. 4
3.2 Parameter-efficient Fine-tuning Framework
- Refiner freezes the pre-trained base model and adds LoRA-enhanced Transformer blocks and a reasoning prediction layer for residual scores.This lightweight adaptation preserves the deployed base ranker while minimizing disruption to existing recommender systems.Source: §3.2 · p. 4
3.3 Residual Reasoning
- Each reasoning step predicts an additive score correction, and cumulative scores are trained with MSE to incrementally reduce remaining prediction error.The cumulative score sums the initial score and all refinements through the current step; the total loss sums the per-step MSE losses.Source: §3.3 · p. 4
4 Experiments
Experiments compare Refiner with ranking and latent-reasoning baselines on public and industrial data, then test its components, robustness, inference trade-offs, and online deployment. Refiner improves over direct prediction and ReaRec in the evaluated settings, with additional online gains over MTGR.
4.1 Experimental Setup
- Experiments use Steam and MovieLens10M plus an industrial exposure-log dataset, evaluate AUC and GAUC, and compare traditional rankers, MTGR, SASRec-ranking, ReaRec, and Refiner.Public-data candidates use K=2 with one sampled negative; industrial candidate sets follow the logged production pipeline.Source: §4.1.1 · pp. 5–6, §4.1.2 · p. 6, §4.1.3 · p. 6
4.2 Overall Performance (RQ1)
- Refiner consistently improves over direct-prediction backbones and outperforms ReaRec on public and industrial datasets; the reported CTR GAUC gain has a paired t-test p-value of 0.008.The comparison covers public-dataset CTR results and industrial CTR and CTCVR results.Source: §4.2 · p. 6, Table 2 · p. 7, Table 3 · p. 7
4.3.1 Ablation Study (RQ2)
- Removing any evaluated Refiner component degrades Steam performance; the End-to-End variant performs worst among the tested variants.Replacing residual refinement with direct BCE supervision also lowers performance, while removing history queries degrades results.Source: §4.3.1 · pp. 7–8, Table 4 · p. 7
4.3.2 Group Analysis (RQ3)
- Refiner improves performance across all user-sequence and item-popularity groups, with the largest gains for the least popular items.MTGR performance declines as user sequences lengthen and item popularity decreases.Source: §4.3.2 · p. 8, Figure 4 · p. 8
4.3.3 Impact of History Queries Number (RQ4)
- On Steam, two or four history queries achieve the best performance, while a small number keeps relative inference latency below 1.1 versus the baseline.Too few queries provide limited condensed user context; too many may introduce redundant history signals.Source: §4.3.3 · pp. 8–9, Figure 3 · p. 8
4.3.4 Impact of Total Reasoning Steps (RQ4)
- On Steam, one or two reasoning steps provide a favorable performance–latency trade-off; later residuals shrink as latency rises, and reported RIL is not a production-efficiency comparison.The additional reasoning passes have not undergone dedicated serving-level optimization.Source: §4.3.4 · p. 9, Figure 5 · p. 9
4.3.5 Online Experiments (RQ5)
- In a seven-day online A/B test, Refiner increased PV-CTR by 0.75%, UV-CTCVR by 0.21%, and GMV by 0.46% over MTGR.The incremental update framework freezes the latest MTGR checkpoint and optimizes Refiner’s additional parameters on current-day data.Source: §4.3.5 · p. 9, Figure 6 · p. 9, Table 5 · p. 10
5 Related Work
Generative recommendation spans semantic-ID generation and discriminative ranking, while reasoning-enhanced recommendation includes explicit textual and latent approaches. Refiner differs from prior latent methods by refining candidate scores additively while preserving isolated candidate processing and training only lightweight modules.
5.1 Generative Recommendation
- GR ranking encodes user context and explicit candidates jointly to produce parallel candidate scores, building on discriminative objectives and dense item-feature embeddings.Semantic-ID approaches instead tokenize items into discrete codes for autoregressive prediction.Source: §5.1 · p. 10
5.2 Reasoning for Recommendation
- Reasoning-enhanced recommendation includes explicit textual reasoning with LLMs and latent multi-step computation in continuous hidden spaces.LatentR3 optimizes compact latent reasoning in LLM-based recommenders through reinforcement learning.Source: §5.2 · p. 10
- Unlike methods refining semantic-ID generation or user representations, Refiner adds score corrections while preserving isolated candidate processing and training lightweight modules.OnePiece instead uses grouped candidate interaction and, in its full ranking configuration, trains a set-wise backbone end to end.Source: §5.2 · p. 10
6 Conclusion and Future Work
Refiner combines residual score refinement, history queries, and parameter-efficient fine-tuning for generative ranking. Its evaluations report consistent improvements, while the paper identifies limits in method comparability and fixed reasoning depth.
6 Conclusion and Future Work
- Refiner combines cumulative score corrections, compact user context for candidate-specific reasoning, and parameter-efficient adaptation without full-model retraining.Experiments across two public datasets, an industrial offline dataset, and an online A/B test show consistent improvements over evaluated baselines.Source: §6 · p. 10
- Refiner applies fixed reasoning depth to every instance, although simpler cases may need fewer steps and harder cases may benefit from more computation.Future work may explore adaptive depth allocation or stopping to balance refinement quality and inference cost.Source: §6 · p. 10
- Comparability is limited by differing task interfaces and backbone requirements; the current production setup lacks the backbone and contextual inputs needed for a fair OnePiece comparison.The industrial offline results and online A/B test provide evidence of practical effectiveness in the evaluated production setting.Source: §6 · p. 10
A.1 Daily Online Performance Trend
Across the seven-day online observation window, relative gains in PV-CTR and UV-CTCVR remain positive despite mild fluctuations, supporting stability under incremental updates.
A.1 Daily Online Performance Trend
- PV-CTR and UV-CTCVR relative gains remain positive throughout the 7-day window, despite mild fluctuations under evolving traffic and model checkpoints.The overall trend supports Refiner's stability under incremental updates.Source: Figure 7 · p. 10, §A.1 · p. 10
GenAI Usage Disclosure
The authors report using ChatGPT and OpenAI Codex for language refinement, LaTeX checks, and code development and debugging, while retaining responsibility for the work.
GenAI Usage Disclosure
- ChatGPT and OpenAI Codex assisted with language refinement, LaTeX checks, and code development and debugging; the authors reviewed and validated outputs and retain responsibility.Source: GenAI Usage Disclosure · p. 11