PaperlayerView exact revision on arXiv

Source-linked AI summary

Parallel Latent Reasoning for Sequential Recommendation

Jiakai Tang, Xu Chen, Wen Chen, Jian Wu, Yuning Jiang, Bo Zheng

arXiv:2601.03153v1cs.IR

TL;DR

Sparse behavioral sequences make it difficult to infer complex user preferences, while existing latent reasoning methods rely on single-trajectory depth scaling with diminishing returns. PLR introduces parallel latent reasoning streams using learnable triggers, diversity regularization, and adaptive aggregation, and experiments report substantial improvements over state-of-the-art baselines with real-time inference efficiency. The paper’s analysis and experiments support parallel reasoning as a way to improve recommendation reasoning beyond depth scaling.

  • Problem

    Sparse behavioral data challenges complex preference inference, and existing latent reasoning methods face diminishing or negative gains from deeper single-trajectory reasoning.

  • Method

    PLR uses learnable trigger tokens to create parallel latent reasoning streams, global reasoning regularization to preserve diversity, and mixture-of-reasoning-streams aggregation to combine outputs.

  • Results

    PLR substantially outperforms state-of-the-art baselines on three real-world datasets while maintaining real-time inference efficiency.

  • Takeaways & Limitations

    Parallel width-level reasoning extends recommendation computation beyond depth-only scaling and supports diverse reasoning trajectories for sparse behavioral sequences.

  • Takeaways & Limitations

    Existing latent reasoning methods can suffer overthinking after multiple reasoning steps, and shallow Transformer architectures typically rely on a small Lipschitz constant bound.

Abstract

from arXiv · show

Capturing complex user preferences from sparse behavioral sequences remains a fundamental challenge in sequential recommendation. Recent latent reasoning methods have shown promise by extending test-time computation through multi-step reasoning, yet they exclusively rely on depth-level scaling along a single trajectory, suffering from diminishing returns as reasoning depth increases. To address this limitation, we propose \textbf{Parallel Latent Reasoning (PLR)}, a novel framework that pioneers width-level computational scaling by exploring multiple diverse reasoning trajectories simultaneously. PLR constructs parallel reasoning streams through learnable trigger tokens in continuous latent space, preserves diversity across streams via global reasoning regularization, and adaptively synthesizes multi-stream outputs through mixture-of-reasoning-streams aggregation. Extensive experiments on three real-world datasets demonstrate that PLR substantially outperforms state-of-the-art baselines while maintaining real-time inference efficiency. Theoretical analysis further validates the effectiveness of parallel reasoning in improving generalization capability. Our work opens new avenues for enhancing reasoning capacity in sequential recommendation beyond existing depth scaling.

1 Introduction

Sequential recommendation must infer complex, evolving preferences from sparse behavior, while existing reasoning methods face latency and depth-scaling limitations. PLR addresses these issues with parallel latent reasoning streams, diversity preservation, adaptive aggregation, and contrastive robustness training.

  • Sparse behavioral data limits recommendation models’ ability to perform complex inference about users’ latent interests and preference dynamics.
  • Explicit reasoning methods incur substantial inference latency and face ambiguity in defining suitable reasoning chains for recommendation.
  • Latent reasoning methods avoid annotations but mainly deepen a single autoregressive reasoning trajectory.
  • Further increases in reasoning depth can produce diminishing or negative gains because initial directions may be suboptimal and errors accumulate.
  • PLR expands computation horizontally by exploring multiple latent reasoning trajectories to represent dynamic and multifaceted user preferences.
  • PLR uses learnable trigger tokens, global reasoning regularization, mixture-of-reasoning-streams aggregation, and reasoning contrastive learning to construct, diversify, combine, and robustify parallel streams.
  • Width-level scaling combines breadth and depth reasoning while maintaining real-time inference efficiency and improving sequential recommendation across three real-world datasets.

2 Preliminary

Sequential recommendation predicts a user’s next interaction from an ordered history. Latent reasoning methods iteratively refine representations, but stacking reasoning steps can cause overthinking and marginal or negative performance gains.

  • Sequential recommendation learns to rank candidate items and recommend the top-K items most likely to follow a user’s chronological interaction sequence.
  • Traditional recommenders use a fixed forward pass, whereas latent reasoning methods add iterative computation to refine representations for sparse-interaction scenarios.
  • A latent reasoning-enhanced model encodes an input sequence into h0, repeatedly applies a reasoning module, and predicts from the refined representation hT.
  • Iterative reasoning increases computational expressiveness and can improve accuracy on complex tasks.
  • Existing methods often show marginal or negative gains after stacking reasoning steps, a limitation termed overthinking.

3 Methodology

PLR extends latent sequential recommendation from depth-only reasoning to parallel streams, using learnable triggers, diversity-preserving objectives, robust contrastive learning, and adaptive aggregation.

  • 3.1 Architecture: Each stream performs iterative latent reasoning, pools its multi-step outputs, and produces a stream-specific representation for final multi-stream aggregation.The architecture retains depth-level refinement within each stream while adding parallel reasoning across streams.
  • 3.1 Architecture: PLR extends depth-level latent reasoning with width-level parallel streams guided by learnable trigger tokens.Each trigger perturbs the initial reasoning state to encourage distinct trajectories while streams share input context and remain causally isolated.
  • 3.2 Optimization: Mixture-of-Reasoning-Streams adaptively combines stream outputs instead of uniformly averaging them, avoiding contamination from inferior reasoning streams.The framework also identifies adaptive aggregation as a key mechanism for synthesizing multi-stream results.
  • 3.2 Optimization: Global reasoning regularization uses bidirectional KL divergence across reasoning states to prevent within-stream and across-stream homogeneity.The objective encourages different steps to explore distinct directions and different streams to discover diverse sequence patterns.
  • 3.2 Optimization: Reasoning contrastive learning applies independent dropout to representations and attention scores, treating same-user views as positives and other users as negatives.The resulting objective encourages task-relevant representations that remain invariant across augmented views.

4 Theoretical Analysis

The analysis explains PLR through ensemble diversity, depth-induced diversity decay, and adaptive gating: diverse streams can reduce error, while excessive depth can erase that benefit.

  • 4.1 Why Diversity Reduces Error: Representational diversity translates into prediction diversity, which reduces ensemble error when streams are not identical.The specialization benefit is nonnegative and becomes zero only when all stream predictions are identical.
  • 4.2 The Refinement-Diversity Trade-off: Depth creates a refinement-diversity trade-off because additional steps improve quality while progressively reducing diversity.The analysis formalizes this tension using contraction mapping theory.
  • 4.2 The Refinement-Diversity Trade-off: When the Lipschitz constant is below 1, diversity decays exponentially as D(T) = exp(-2γT)D(0).The result provides a theoretical explanation for why depth-only scaling can eventually collapse streams toward a common fixed point.
  • 4.2 The Refinement-Diversity Trade-off: Increasing reasoning depth initially improves refinement faster than diversity declines, but later diversity loss outweighs refinement gains.This behavior is identified as the empirical over-thinking phenomenon.
  • 4.3 Gating Benefits: Adaptive gating outperforms uniform averaging when streams specialize because conditional mutual information between stream selection and the target is positive.The benefit arises when different streams excel on different sequence perspectives.

5 Related Work

Sequential recommendation evolved from neural sequence models toward reasoning-enhanced methods, including explicit language-based reasoning and lower-latency latent reasoning in continuous representation space.

  • Sequential Recommendation: Sequential recommendation predicts users’ next interactions from chronologically ordered behavior using recurrent, convolutional, and Transformer-based architectures.SASRec introduced self-attention for modeling sequential dependencies.
  • Multimodal Recommendation: Multimodal methods enrich item representations beyond IDs, with UniSRec targeting universal text-based representations and MoRec examining multimodal fusion.These approaches address limitations associated with simple ID-based sequences.
  • Reasoning-Enhanced Recommendation: Chain-of-Thought prompting motivated reasoning-enhanced recommendation by extending computation at test time beyond shallow pattern matching.Explicit approaches generate interpretable reasoning chains for recommendation tasks.
  • Reasoning-Enhanced Recommendation: Latent reasoning methods perform multi-step computation in continuous representation space, avoiding explicit CoT annotations while maintaining low latency.ReaRec introduced multi-step latent reasoning, while LARES decoupled encoding and reasoning modules with self-supervised and reinforcement learning components.

6 Experiments

Experiments across three Amazon Review domains evaluate PLR against sequential-recommendation backbones and latent-reasoning baselines. PLR improves performance across most settings, with robustness, component, sensitivity, ceiling, and efficiency analyses supporting its parallel-reasoning design.

  • Experimental Setup: Experiments use CDs & Vinyl, Movies & TV, and Video & Games, with chronological user-sequence splits and positive interactions defined by ratings above 3.Models are evaluated using NDCG@K and Recall@K at K=10 and 20.
  • Overall Performance: Reasoning-enhanced methods outperform base models across most settings, including 5.15% higher Recall@10 and 2.23% higher Recall@20 for ReaRec-ERL on CDs & Vinyl with SASRec.On Movies & TV with BERT4Rec, ReaRec-ERL improves Recall@10 by 16.67% and NDCG@10 by 13.49%.
  • Overall Performance: PLR improves Recall@20 by 12.07% over ReaRec-ERL on CDs & Vinyl with SASRec and improves Recall@10 by 14.91% on UniSRec.PLR performs best across most metrics, particularly on the sparser CDs & Vinyl dataset.
  • Ablation Study: Removing MoRS lowers CDs & Vinyl Recall@20 from 0.0873 to 0.0785, while removing RCL lowers Video & Games Recall@20 from 0.1033 to 0.0970.Removing KL also reduces performance, but its effect is relatively modest.
  • Sensitivity Analysis: Performance peaks at M=2 reasoning streams, T=2 reasoning steps, λ=0.1 regularization, and dropout p∈[0.2, 0.5].More streams or deeper reasoning degrade performance, while extreme regularization or dropout harms optimization or information preservation.
  • Performance Ceiling Analysis: PLR and ReaRec-series methods have comparable oracle ceilings with NDCG@20∈[0.0402, 0.0416], but PLR has smaller current-to-oracle gaps.The result indicates that PLR more effectively uses reasoning capacity without relying as heavily on oracle step selection.
  • Efficiency Analysis: PLR adds 5.22% FLOPs and 5.80% inference latency over base SASRec while delivering substantial performance gains.The reported overhead supports real-time deployment efficiency.

7 Conclusion

PLR introduces width-level computational scaling for sequential recommendation by exploring diverse reasoning trajectories in parallel. The framework outperforms state-of-the-art baselines while maintaining real-time inference efficiency, and motivates future joint scaling of parallel and depth reasoning.

  • PLR pioneers width-level computational scaling for sequential recommendation by exploring multiple diverse reasoning trajectories simultaneously.
  • Learnable trigger tokens, global reasoning regularization, and mixture-of-reasoning-streams aggregation construct, diversify, and combine parallel reasoning streams.
  • PLR substantially outperforms state-of-the-art baselines while maintaining real-time inference efficiency.
  • Theoretical analysis validates parallel reasoning for reducing prediction error through ensemble diversity and mitigating over-thinking issues.
  • Future work will explore jointly optimizing parallel and depth reasoning scaling for multidimensional reasoning architectures.

A.1 Proof of Theorem 4.1 (Ensemble Error Decomposition)

The proof uses Jensen’s inequality to compare independent stream losses with an ensemble loss. The resulting gap is nonnegative and is strictly positive when stream predictions differ.

  • Jensen’s inequality relates the expected negative log-likelihood under the ensemble distribution to the individual stream losses.
  • The ensemble gap is strictly positive when stream predictions differ.
  • The diversity benefit I(S_u) = L̄_ind − L_ens(S_u) is nonnegative, with equality only when all stream predictions are identical.

A.2 Proof of Proposition 4.2 (Diversity-Specialization Connection)

The proposition connects separated stream representations to specialization benefits. Under the stated embedding and scoring conditions, the benefit is lower bounded by pairwise divergence between stream predictions.

  • Under linear scoring, stream predictions are obtained by applying softmax to the inner products between stream states and item embeddings.
  • Total variation distance between stream predictions is related to embedding separation through bounds involving item-embedding radius and minimum separation.
  • The specialization benefit is lower bounded by the average pairwise KL divergence between stream predictions.

A.3 Proof of Theorem 4.4 (Diversity Decay)

The proof analyzes how stream diversity evolves through recurrent updates and how adaptive stream weighting compares with uniform aggregation. It also links specialization to conditional mutual information.

  • The recurrent update contracts pairwise stream-state differences by at most a factor L at each reasoning step.
  • After iterating across reasoning steps, the diversity bound takes an exponential form when L < 1.
  • The gated ensemble improves over uniform aggregation by weighting streams according to their sequence-dependent losses.
  • When streams specialize, the selected stream conveys positive information about the target, expressed as I(Z;V|S_u) > 0.
Loading 2601.03153v1…