LASAR: Latent Adaptive Semantic Aligned Reasoning for Generative Recommendation

Abstract: LLMs have demonstrated powerful reasoning capabilities through Chain-of-Thought (CoT) in various tasks, yet the inefficiency of token-by-token generation hinders real-world deployment in latency-sensitive recommender systems. Latent reasoning has emerged as an effective paradigm in LLMs, performing multi-step inference in a continuous hidden-state space to achieve stronger reasoning at lower cost. However, this paradigm remains underexplored in mainstream generative recommendation. Adapting it reveals three unique challenges: (1) the gap between prior-less Semantic ID (SID) symbols and continuous latent reasoning - SIDs lack pre-trained semantics, hindering joint optimization; (2) representation drift due to a lack of reasoning chain supervision; and (3) the suboptimality of applying a globally fixed reasoning depth. To address these, we propose LASAR (Latent Adaptive Semantic Aligned Reasoning), an SFT-then-RL framework. First, we bridge this gap via two-stage training: Stage 1 grounds SID semantics before Stage 2 introduces latent reasoning, ensuring efficient convergence. Second, we mitigate representation drift through explicit CoT semantic alignment. Step-wise bidirectional KL divergence constrains the latent reasoning trajectory using hidden-state anchors extracted from CoT text, while a Policy Head predicts per-sample reasoning depth. Third, during the GRPO-based RL phase, terminal-only KL alignment accommodates variable-length reasoning, and REINFORCE optimizes the Policy Head to dynamically allocate steps. This nearly halves the average latent step count while simultaneously improving recommendation quality. Experiments on three real-world datasets demonstrate that LASAR outperforms all baselines. It adds marginal inference latency and is roughly 20 times faster than generating explicit CoT text.

• The paper introduces LASAR, a two-stage pipeline that decouples semantic grounding and latent reasoning to accelerate and improve generative recommendation quality.
• It anchors latent reasoning via bidirectional KL divergence alignment with teacher-generated chain-of-thought semantics and adaptively allocates computation steps using an RL-based policy head.
• Empirical results on Amazon review datasets show LASAR achieving superior performance and a 20× reduction in inference latency compared to explicit chain-of-thought models.

LASAR: Latent Adaptive Semantic Aligned Reasoning for Generative Recommendation

Motivation and Challenges

Integrating advanced reasoning into LLM-based generative recommender systems is thwarted by the inefficiency of explicit textual chain-of-thought (CoT) generation, resulting in prohibitive inference latency. While latent reasoning paradigms, such as Coconut, enable multi-step inference through recurrent computations in continuous hidden-state space, their application to generative recommendation is nontrivial and previously underexplored. The paper identifies three fundamental obstacles: (1) the semantic mismatch between ungrounded vector-quantized Semantic IDs (SIDs) and latent reasoning spaces; (2) representation drift due to the lack of explicit reasoning chain supervision; and (3) the inefficiency and suboptimality of globally fixed reasoning depth across samples.

The LASAR Framework

LASAR is introduced as a structured SFT-then-RL pipeline designed to address the above challenges for decoder-only LLM-based recommendation with full latent reasoning and adaptive computation depth. The approach builds on a Coconut-style hidden-state feedback loop but introduces several architectural and training innovations:

Two-Stage Decoupling for Semantic Grounding

Recommendation-specific SIDs, unlike NLP tokens, lack pretrained semantics, impeding the optimization of latent reasoning in generative recommendation. LASAR decouples the learning process into:

• Stage 1: The backbone model is trained to align SIDs to items, establishing a stable and meaningful item-symbol space via cross-entropy before any latent reasoning is introduced.
• Stage 2: The latent reasoning mechanism is introduced, allowing the model to perform reasoning atop the now-grounded SID representations. This ordering ensures rapid and stable convergence, empirically yielding a threefold acceleration.

Latent Reasoning with CoT Semantic Alignment

To prevent representation collapse in the continuous latent loop, LASAR extracts hidden states from semantically segmented, teacher-generated CoT text (provided by an external LLM only at training time). Each latent step is aligned with its corresponding CoT segment via bidirectional KL divergence, anchoring the internal reasoning trajectory. This distinguishes LASAR from previous approaches using cosine or MSE, which empirically degrade performance or fail to yield improvements.

Adaptive Step Allocation via Policy Head + REINFORCE

LASAR employs a Policy Head—a two-layer MLP—to predict the number of latent reasoning steps, $\mathcal{N}$, on a per-sample basis. The head is pre-trained in SFT using the number of CoT segments as supervision, then optimized via REINFORCE in the RL phase to maximize group-based recommendation rewards and minimize unnecessary computation. A batch-efficient masking and padding scheme allows all samples in a minibatch to share the maximum $\mathcal{N}$, enabling full parallelism.

RL Phase: Joint Quality and Efficiency Optimization

The RL stage integrates three objectives:

• GRPO for maximizing generation and ranking quality;
• Terminal-only KL Alignment to maintain semantic consistency at the end-point of variable-length latent chains;
• REINFORCE-based compression of step counts, guided by reward signals and explicit penalties.

This synergy achieves simultaneous improvement in recommendation performance and drastic reduction in computational steps for most samples.

Empirical Results

Experiments are conducted on three Amazon product review datasets of varying sparsity. LASAR outperforms traditional discriminative (e.g., SASRec, GRU4Rec), generative (LC-Rec, MiniOneRec), latent reasoning (ReaRec), and explicit-CoT (GREAM) baselines across almost all NDCG@K and HitRate@K metrics. Notably:

• On highly sparse settings, performance gains are maximized, indicating latent iterative reasoning compensates for weak collaborative signals.
• RL-driven step adaptation halves the mean latent reasoning depth relative to the SFT initialization, with no drop—and frequent increase—in quality.
• Inference latency is essentially unchanged relative to direct generation baselines, and is approximately 20$\times$ faster than explicit CoT text generation.

Ablation studies demonstrate the necessity of each core LASAR innovation: latent reasoning without semantic alignment fails, alignment via non-KL divergences yields no gains, and both terminal KL and step-adaptive RL are essential for compression without sacrificing accuracy.

Theoretical and Practical Implications

LASAR is the first framework to realize full multi-step latent reasoning with recurrent state feedback and per-sample adaptive depth in mainstream LLM-based generative recommendation. It systematically resolves the lack of semantic grounding in SID-based recommendation, prevents drift inherent to unconstrained latent loops, and disambiguates resource allocation through direct step optimization. The methodology demonstrates the feasibility and advantages of latent (i.e., unobserved) reasoning trajectories in recommendation, avoiding the mode competition and latency bottlenecks of explicit textual reasoning.

From a practical perspective, LASAR positions latent reasoning as viable in production-scale recommenders due to its minimal inference overhead and ability to harvest semantic benefit under sparse data conditions. The adaptive step mechanism can serve as a foundation for test-time compute allocation in resource-constrained deployment scenarios.

Limitations and Future Directions

A major limitation is the inherent computational serialization within each latent reasoning chain: teacher-forcing is infeasible for latent steps, which must be computed sequentially. Efficient latent loop execution remains an open engineering and algorithmic challenge. Future research may extend LASAR's paradigm to conversational and cross-domain recommendation scenarios, or investigate more granular adaptive compute mechanisms and parallelizable architectures.

Conclusion

LASAR establishes a new standard for integrating latent recurrent reasoning with adaptive computation in generative recommendation. Its decoupled semantic alignment, KL-based trajectory anchoring, and sample-wise step optimization result in robust gains in both quality and efficiency, with broad implications for the future design of reasoning-augmented, LLM-driven recommender systems (2605.10207).