ManCAR: Manifold-Constrained Latent Reasoning with Adaptive Test-Time Computation for Sequential Recommendation

Kun Yang, Yuxuan Zhu, Yazhe Chen, Siyao Zheng, Bangyang Hong, Kangle Wu, Yabo Ni, Anxiang Zeng, Cong Fu, Hui LiarXiv 2026 · arXiv

TL;DR

ManCAR improves sequential recommendation by constraining multi-step latent reasoning to a "collaborative manifold" derived from user-item interaction graphs. This prevents the model's internal reasoning from wandering into unrealistic item spaces ("latent drift") and introduces an adaptive test-time stopping mechanism that dynamically scales computation based on recommendation complexity.

What Problem Does This Paper Address?

Sequential recommenders predict a user's next action based on their history. Recently, researchers have borrowed the concept of "recursive depth" or "latent reasoning" from Large Language Models (LLMs). Instead of generating text-based chains-of-thought, these models iteratively refine hidden vector representations over multiple steps before making a final prediction, effectively increasing computational depth without adding parameters.

However, existing methods suffer from latent drift. Because they are trained primarily using target-dominant objectives (i.e., forcing the final state to match the target item), the intermediate reasoning steps have too much freedom. They often wander into implausible regions of the item space that do not align with real user behavior patterns, leading to fragile generalizations at test time.

Figure 1: Illustration of constrained versus unconstrained latent reasoning. Graph-conditioned reasoning trajectories remain within a collaborative manifold defined by neighbor items, enabling stable and directed refinement toward the target. In contrast, unconstrained reasoning may drift out- side feasible regions, leading to inefficient or unstable paths.

Methods

The authors propose ManCAR (Manifold-Constrained Adaptive Reasoning) to anchor latent reasoning to real-world item relationships.

  1. Collaborative Manifold Constraint: The model constructs a global item interaction graph using a Swing-style co-interaction algorithm. For any given user history, ManCAR defines a local candidate set based on items reachable within $k$-hops from the user's recent actions. This candidate set acts as a low-dimensional "manifold" that bounds the reasoning process.
  2. Variational Training Objective (ELBO-like): The framework treats latent reasoning as approximate inference over an intent variable. It introduces a graph-conditioned "Teacher Prior" that assigns higher probability mass to strongly connected neighbors. The model minimizes a loss function that balances accurate target prediction with a KL-divergence regularization term, forcing the student distribution to stay on the manifold.
  3. Teacher Scheduling: During training, the teacher distribution transitions smoothly from a diffuse, graph-aware prior to a sharply peaked distribution centered on the target item.
  4. Adaptive Test-Time Halting: Instead of using a fixed number of reasoning steps for every recommendation, ManCAR monitors the KL-divergence between successive student distributions. When the distribution stabilizes ($D_{KL}(p_{t'-1} \parallel p_{t'}) < \epsilon$), the model halts early, saving compute on simple queries while spending more steps on complex ones.
Figure 2: Overview of ManCAR. ManCAR performs multi- step latent reasoning constrained by a graph-induced can- didate set. At each step, the reasoning state is regularized toward a scheduled teacher prior defined on collaboratively reachable items, ensuring manifold-consistent refinement. Adaptive test-time termination stops reasoning when the induced item distributions stabilize.

Datasets

ManCAR was evaluated on seven subcategory datasets from the standard Amazon 2023 Reviews corpus:

  • CDs & Vinyl (CDs)
  • Video & Games (Video)
  • Office Products (Office)
  • Arts, Crafts & Sewing (Arts)
  • Grocery & Gourmet Food (Grocery)
  • Musical Instruments (Music)
  • Toys & Games (Toys)

Key Insights and Ideas

  • Reasoning as Navigation, Not Free Refinement: Instead of treating latent refinement as an unconstrained walk through high-dimensional space, effective reasoning should be viewed as navigating a pre-existing collaborative manifold defined by collective user behavior.
  • KL Distillation as Gradient Flow: The authors mathematically prove that minimizing the KL loss between the student and the graph-conditioned teacher induces a gradient flow. This flow pulls the prediction barycenter toward the teacher barycenter within a graph-restricted convex hull, theoretically guaranteeing local graph smoothness.
  • Asymmetric and Data-Aware Compute: Unlike prior baselines that use fixed, shallow reasoning depths (e.g., 2–3 steps), ManCAR dynamically allocates compute. It naturally uses deeper reasoning for complex datasets (like CDs and Toys) and terminates early on simpler ones (like Arts), matching human-like intuition where harder problems require more thought.

Key Results

ManCAR consistently outperformed state-of-the-art sequential and reasoning-based baselines (such as SASRec, BERT4Rec, ReaRec, and LARES) across all seven benchmarks.

  • Substantial Accuracy Gains: On the CDs dataset, ManCAR achieved a 46.88% relative improvement over the second-best baseline in terms of NDCG@10.
  • Near-Ceiling Performance: Performance ceiling analysis shows that ManCAR's adaptive halting mechanism achieves results remarkably close to an "oracle ceiling" (the theoretical best step per sample), proving the effectiveness of its early-stopping criterion.
Figure 3: Performance ceiling analysis on Office and Toys.

Limitations and Future Work

The authors note that ManCAR's performance margins are tightly coupled with interaction density. On highly sparse datasets (such as Musical Instruments or Arts), the graph edges become noisier, which limits the effectiveness of the graph-conditioned manifold.

Future Direction: A key area for future work is improving the model's robustness under limited preference evidence, specifically addressing the cold-start problem where graph structures are weak.

Summary updated

Abstract

Sequential recommendation increasingly employs latent multi-step reasoning to enhance test-time computation. Despite empirical gains, existing approaches largely drive intermediate reasoning states via target-dominant objectives without imposing explicit feasibility constraints. This results in latent drift, where reasoning trajectories deviate into implausible regions. We argue that effective recommendation reasoning should instead be viewed as navigation on a collaborative manifold rather than free-form latent refinement. To this end, we propose ManCAR (Manifold-Constrained Adaptive Reasoning), a principled framework that grounds reasoning within the topology of a global interaction graph. ManCAR constructs a local intent prior from the collaborative neighborhood of a user's recent actions, represented as a distribution over the item simplex. During training, the model progressively aligns its latent predictive distribution with this prior, forcing the reasoning trajectory to remain within the valid manifold. At test time, reasoning proceeds adaptively until the predictive distribution stabilizes, avoiding over-refinement. We provide a variational interpretation of ManCAR to theoretically validate its drift-prevention and adaptive test-time stopping mechanisms. Experiments on seven benchmarks demonstrate that ManCAR consistently outperforms state-of-the-art baselines, achieving up to a 46.88% relative improvement w.r.t. NDCG@10. Our code is available at https://github.com/FuCongResearchSquad/ManCAR.
BibTeX
@misc{yang2026mancar,
  author = {Kun Yang and Yuxuan Zhu and Yazhe Chen and Siyao Zheng and Bangyang Hong and Kangle Wu and Yabo Ni and Anxiang Zeng and Cong Fu and Hui Li},
  title = {{ManCAR: Manifold-Constrained Latent Reasoning with Adaptive Test-Time Computation for Sequential Recommendation}},
  year = {2026},
  eprint = {2602.20093},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2602.20093},
}
Figure