acceptodds
Under review as a conference paper at ICLR 2027

LASAR: Latent Adaptive Semantic Aligned Reasoning for Generative Recommendation

16 views

Abstract

Large Language Models (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. Achieving this reveals three key challenges: (1) the gap between prior-less Semantic ID (SID) symbols and continuous latent reasoning, as 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 improving recommendation quality. Experiments on four real-world datasets show the best overall performance across the evaluated settings. It adds limited inference latency and is roughly 20x faster than generating explicit CoT text. Code: https://anonymous.4open.science/r/LASAR-E292

Primary area:
foundation or frontier models, including LLMs
Keywords:
Generative Recommendation, Recommender Systems, Latent Reasoning, Large Language Models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_kkKDOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

LASAR brings Coconut-style recurrent latent reasoning into SID-based generative recommendation. The last-layer hidden state is fed back as the next input embedding for N steps before the item SID is decoded (Eq. 3). Training runs SFT then RL, with three designs.

  1. Two-stage SFT: Stage 1 grounds SIDs with plain cross-entropy, and only Stage 2 adds the latent loop (Fig. 4).
  2. Step-wise bidirectional KL alignment: each latent step is aligned to a fixed anchor, namely the Stage-1 hidden state at a segment boundary of a GPT-5 CoT trace segmented by embedding similarity (Eq. 4).
  3. A Policy Head predicting per-sample depth N: it is CE-warm-started on the segment counts, then trained with REINFORCE and a step penalty alongside GRPO and a terminal-only KL (Eq. 5).

With Qwen3-0.6B on Amazon Beauty/Instruments/Sports and Steam, LASAR is best on 23 of 24 metric–dataset cells of Table 1. The baselines are SASRec, GRU4Rec, LC-Rec, MiniOneRec, ReaRec and a direct-answer GREAM variant. Inference overhead is 7–16% over MiniOneRec (Table 4). The paper also reports alignment and RL ablations (Tables 2, 3, 9), depth analyses (Fig. 5, 9) and a 1.7B LoRA scaling point (Table 5).

Critical strengths
  1. A well-designed adaptation of latent reasoning to SID-based generative recommendation. The paper identifies why Coconut-style latent reasoning fails out of the box in SID-based recommenders: SIDs carry no pretrained semantics, so jointly learning them with latent reasoning converges poorly (Fig. 4). Two-stage decoupling and alignment to CoT anchors address this directly. The implementation is also carefully engineered for beam-search recommendation:
    • the hidden-state feedback loop reuses the KV cache;
    • variable-depth samples are batched with masking;
    • predicting N before the loop lets all beams of a prompt share the same depth;
    • decoding is trie-constrained (§2.2, App. E). Unaligned latent reasoning can even fall below plain SFT (Instruments N@10 0.0576 vs 0.0634, Table 9), while KL alignment improves results on all four datasets (Tables 2, 9). The ablations expose these failure modes rather than hiding them, which makes the design choices convincing.
  2. Consistent gains at near-zero inference cost, with practical relevance. All generative baselines share the base model, prompt, SID tokenizer and training data, and the GREAM variant uses the same CoT traces (§3.1, App. D/E). Under that control LASAR is best on 23 of 24 cells of Table 1, including the non-Amazon Steam dataset. It adds only 7–16% GPU-seconds per sample over MiniOneRec (Table 4), and REINFORCE cuts mean depth from 4.20 to 2.47 while improving quality (Table 3). Inference cost is the main barrier to using LLM reasoning in production recommenders, so adding reasoning without generating text addresses a practically important problem.
Critical weaknesses
  1. Baseline consistency and statistical robustness. The MiniOneRec numbers in Table 1 are lower than the paper's own "Pure SFT" model (Table 9) on three datasets:

    • Sports HR@10: 0.0260 vs 0.0237.
    • Instruments HR@10: 0.0888 vs 0.0826, which is also above GREAM's 0.0850.
    • Steam N@10: 0.0372 vs 0.0357.

    The source of this gap must be clarified. Against the stronger numbers, LASAR's margins shrink substantially (e.g., Sports HR@10 from about +18% to about +8%). On Sports, KL-aligned SFT alone (HR@10 0.0281) already matches the full LASAR (0.0280), so the RL stage's contribution there is unclear. On Beauty, gains over the best baseline are small (N@10 0.0303 vs 0.0295) and only p < 0.1.

    The size of the gain also varies widely across datasets, and the paper's explanation (larger gains under higher sparsity, §3.2) is contradicted by its own data. Steam is the densest dataset (sparsity 99.468%) yet shows large gains (+9.2% N@10 over the strongest baseline). Beauty is sparser (99.935%) but shows the smallest. Notably, Beauty is the only dataset where the Table 1 MiniOneRec is stronger than the Table 9 "Pure SFT" (N@10 0.0295 vs 0.0277). Part of the cross-dataset variation may therefore reflect baseline strength rather than properties of the data. All results appear to be single training runs: the bootstrap in App. C resamples test users and does not capture training variance, and no validation split is described for selecting λ or checkpoints.

  2. The efficiency claims lack matched comparisons and full cost accounting.

    • The fixed-depth results (Fig. 5a, Fig. 9) force N on a checkpoint trained under adaptive N (App. G), so the fixed depths are out of distribution for that model. This does not show that adaptive allocation beats a model trained at a fixed depth. Fixed N = 1 on the same checkpoint is already close to adaptive.
    • The batch loop runs to max(N) with masked padding (App. E.1), so a lower mean N does not by itself imply less executed computation.
    • The "over 20× faster than explicit CoT" claim compares latency against GREAM's CoT-generation mode (Table 4), but accuracy against its direct-answer mode (Table 1). The accuracy of the CoT-generation mode is never reported.
    • The method needs a GPT-5 trace for every training sample (App. E.2), yet the cost of generating them (API spend or time) is not reported. This offline cost matters for deployment, especially when the model is retrained often.
Questions for authors
  1. Why is the Table 1 MiniOneRec weaker than the Table 9 "Pure SFT" on Sports, Instruments and Steam (e.g., RL stage, hyperparameters, epochs)? Would the conclusions change if the stronger of the two were used as the baseline?
  2. What explains the large variation in gains across datasets, given that sparsity does not account for it (Steam vs Beauty)? An analysis by user or item characteristics (e.g., history length, category diversity, item popularity) on existing outputs would help readers judge when LASAR is worth using.
  3. What was the cost of generating the GPT-5 CoT traces (number of traces, API spend or time)? How sensitive is LASAR to the teacher model, e.g., with a smaller open-source teacher?
2 comments

Dear Reviewer kkKD,

With less than one day remaining in the discussion period, we would like to kindly follow up, as we have not yet received your feedback on our responses. We would appreciate knowing whether our clarifications and additional results address your concerns, or whether any points require further explanation. We remain available to clarify them before the discussion closes.

Thank you again for your thoughtful review and time.

Best regards, The Authors

Thank you for the detailed response. It addresses many of my questions, including the Pure SFT comparison, the RL contribution on Sports, and the trace cost and teacher sensitivity. I will maintain my original assessment. Thank you for your efforts.

Reviewer_dmU4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper propose LASAR (Latent Adaptive Semantic Aligned Reasoning), an SFT-then-RL framework for adaptive recurrent latent reasoning in mainstream generative recommendation. There are three key challenges when Latent reasoning is applied in mainstream generative recommendation: the gap between prior-less Semantic ID (SID) symbols and continuous latent reasoning, representation drift and the suboptimality of applying a globally fixed reasoning depth. To address these challenges, LASAR uses two-stage decoupling to bridge the semantic grounding gap, explicit CoT semantic alignment to address representation drift and a Policy Head with REINFORCE to enable flexible step optimization. Experiments on four real-world datasets show LASAR achieves best performance on nearly all metric-dataset combinations, and ablations reveal the failure modes of latent reasoning and isolate each component's contribution.

Critical strengths
  1. This paper tries to address three challenges in introducing recurrent latent reasoning into generative recommendation and provides three corresponding solutions: two-stage decoupling, explicit CoT semantic alignment, and a Policy Head with REINFORCE. The motivation is clear and easy to follow.
  2. The paper evaluates LASAR on four benchmarks and shows strong performance. Ablation studies demonstrate the effectiveness of each module, while efficiency experiments show limited additional inference overhead.
Critical weaknesses
  1. This paper proposes flexible step optimization so that simple samples use fewer steps and complex samples can use more. And Section 3.4 Step Optimization Analysis shows the effectiveness of adaptive allocation, but the paper lacks an efficiency comparison with fixed-step reasoning. Therefore, it is unclear how much efficiency improvements adaptive allocation can provides. Section 3.5 Inference Efficiency and Model Scaling also shows that LASAR has slightly higher inference cost than MiniOneRec. Please explain this overhead, provide a detailed cost for each module, and discuss possible improvements.
  2. Please consider discussing and comparing several recent baselines, such as COBRA [1], CoderRec [2], and IDIOMoE [3]. These methods reflect recent advances in recommendation methods and performance. Comparing LASAR with one or two suitable baselines under the same evaluation settings would strengthen the robustness of this paper.
  3. Table 7 in the appendix reports the implementation details and hyperparameters. However, the paper should explain how important parameters, such as αalign\alpha_{align}, βpolicy\beta_{policy}, γRF\gamma_{RF} and γKL\gamma_{KL} were chosen and analyze how different values affect the final performance.

[1] Sparse Meets Dense: Unified Generative Recommendations with Cascaded Sparse-Dense Representations. NeurIPS 2025.

[2] Cross-Scale Collaboration between LLMs and Lightweight Sequential Recommenders with Domain-Specific Latent Reasoning. AAAI 2026.

[3] Catalog-Native LLM: Speaking Item-ID Dialect with Less Entanglement for Recommendation. ICLR 2026.

Questions for authors
  1. Teacher-generated CoT text provides the basis for the explicit CoT anchors and the reasoning-depth labels used in latent SFT. Its quality may therefore affect both latent-state alignment and step allocation. The authors should evaluate different teacher CoT models to assess how teacher choice affects recommendation performance and the robustness of LASAR.
  2. In Appendix E.2 Prompt Format and CoT Reasoning Details, the authors state that "LASAR aligns its latent reasoning steps only to Steps 1–3 of the teacher trace (analyze history, build preferences, uncover intent) and excludes Step 4 before segmentation." However, since the teacher model has access to the target item when generating the CoT, Steps 1–3 may still contain target-specific information or cues. Removing Step 4 may not fully remove this influence. Please compare the current setting with CoT generated from the interaction history without access to the target item.
2 comments

Dear Reviewer dmU4,

With less than one day remaining in the discussion period, we would like to kindly follow up, as we have not yet received your feedback on our responses. We would appreciate your thoughts on the additional comparisons, teacher-model experiments, and clarifications provided in our replies. Please let us know if any concerns remain unresolved so that we can respond before the discussion closes.

Thank you again for your constructive review and time.

Best regards, The Authors

Dear Reviewer dmU4,

Thank you again for your thoughtful suggestions. We hope that our responses and additional experiments provide a clearer picture of LASAR’s contribution and effectiveness. We would be grateful for your consideration and reassessment of our work.

Best regards,

The Authors

Reviewer_aoMkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

LASAR adapts recurrent latent reasoning to SID-based generative recommendation. It first grounds item IDs through plain SFT, then aligns latent states with segment-boundary representations of teacher CoT traces. A policy head predicts reasoning depth and is subsequently optimized with REINFORCE alongside GRPO and terminal alignment. The paper reports improved recommendation accuracy on three Amazon domains and Steam, component ablations, and fewer latent steps after RL. Reported inference overhead over MiniOneRec is 7–16%, while explicit CoT generation is substantially slower.

Critical strengths
  1. The combination addresses a useful practical problem: obtaining the benefits of additional computation in generative recommendation without decoding a long explanation. The staged grounding, intermediate supervision, and learned depth allocation form a coherent approach, and the discussion of unsuccessful unaligned variants makes the contribution more informative than a simple transfer of Coconut.
  2. The empirical analysis is reasonably broad. Table 1 includes a non-Amazon domain, Tables 2, 3, and 9 examine alignment and RL components, and Figure 9 tests forced depths within the same user groups. The matched backbone and input protocol, together with shared CoT data in the GREAM comparison, are useful controls. These results support the potential value of the method, subject to resolving the implementation concerns below.
Critical weaknesses
  1. The released training paths do not match the central alignment protocol. sft_two_stage.py:1092–1105 omits use_kl_alignment when constructing the Stage 2 simplified wrapper, whose false default leads to MSE rather than Eq. 4. In RL, rl_laser.py:73–84 loads anchors keyed by training-row csv_index, but laser_grpo_trainer.py:824–829 retrieves them using the target item's catalog index, which does not identify that instance's CoT trace. These discrepancies affect attribution of the alignment gains in Tables 2 and 3. The result-producing code and commands are needed; static inspection alone cannot establish which paths the reported experiments used.
  2. When selecting the simplified wrapper identified in the README as the paper's core model, the released evaluator does not carry recurrent states into answer decoding. evaluate.py:488–529 passes the wrapper's returned inputs_embeds to base-model generation, but laser_model_simplified.py:123,150,294 returns the original embeddings without replacing thought embeddings by recurrent hidden states. The evaluator also offers a full-wrapper branch that does perform replacement. Clarifying which branch produced Tables 1 and 4 is therefore essential to establish that their accuracy and latency measure the proposed inference mechanism.
Questions for authors
  1. Could you provide the result-producing training commit and commands, showing that Stage 2 uses Eq. 4 and that each RL instance retrieves its own terminal CoT anchor? If the released paths were used, do corrected Beauty and Sports ablations preserve the reported gains?
  2. Which evaluator and wrapper produced Tables 1 and 4? Please explain how the recurrent hidden states condition SID decoding in that configuration. Resolving this together with the training correspondence above would materially improve my recommendation.
1 comment

Dear Reviewer aoMk,

With less than one day remaining in the discussion period, we would like to kindly follow up, as we have not yet received your feedback on our response. We would appreciate your thoughts. We remain available to provide further clarification.

Thank you again for your careful review and time.

Best regards, The Authors

Comments on the paper

2 comments

Dear Reviewers,

Thank you again for your time and constructive feedback. As the discussion period is nearing its end with less than 2 days remaining, we would like to follow up on our rebuttals and ask whether any concerns remain unresolved. We would be happy to provide further clarification. If there are any feedback you'd like us to tell us, please let us know. Your insights are invaluable to us.

Best regards,

The Authors

AuthorsOpenReview ↗

General Response and Summary of Additional Evidence

Dear Reviewers and Area Chair,

Thank you for your careful reviews and constructive feedback. We summarize LASAR’s contributions, the strengths recognized in the reviews, and the additional evidence provided during the discussion.

Contributions and recognized strengths

  • Practical motivation and contribution. To our knowledge, LASAR is the first framework to combine recurrent hidden-state reasoning with instance-specific depth in SID-based generative recommendation, avoiding lengthy CoT generation at inference. Reviewer aoMk recognized the practical value and described the contribution as more informative than a simple transfer of Coconut; Reviewer dmU4 found the motivation clear.

  • Coherent methodology and informative analysis. LASAR addresses SID semantic grounding, representation drift, and fixed reasoning budgets through two-stage training, CoT-guided alignment, and adaptive depth optimization. Reviewers aoMk and dmU4 recognized the coherent design, while aoMk highlighted the informative failure-mode analysis.

  • Strong and controlled empirical evaluation. Experiments across four datasets show the best overall recommendation performance, with limited overhead over MiniOneRec and over 20× faster inference than explicit CoT generation. The reviews highlighted the strong results, component ablations, matched evaluation protocol, and shared CoT supervision.

Additional evidence addressing the reviews

  • [kkKD, dmU4] Comparative performance. LASAR consistently outperforms Pure SFT in the reported comparisons and exceeds the adapted CoderRec baseline on all six Beauty metrics under a shared backbone and evaluation protocol.
  • [aoMk] Component effectiveness. Our additional experiments further support the benefits of terminal KL alignment and adaptive depth optimization.
  • [dmU4, kkKD] Teacher robustness. The smaller open-source Qwen3.6-35B-A3B teacher yields comparable results to GPT-5 at both the SFT and SFT+RL stages.
  • [dmU4] Supervision analysis. We clarify the shared teacher-supervision protocol. A supplementary content-selection experiment on Instruments shows improvements across all six metrics when retaining only the history-analysis portion of the existing traces.
  • [kkKD, dmU4] Adaptive reasoning and sensitivity. Adaptive allocation outperforms fixed one-step reasoning on Sports by 6.1% in HR@10 and 12.6% in NDCG@10 on the same checkpoint. Additional step-penalty sensitivity results support performance stability across the evaluated values.
  • [kkKD, dmU4] Cost transparency. We provide a module-level inference cost breakdown and dataset-level offline CoT synthesis statistics, clarifying both inference computation and the one-time supervision cost shared with the Explicit CoT baseline.

With less than half a day remaining in the discussion period, we would like to kindly draw attention to our responses and additional results. Reviewer kkKD confirmed that our responses address many of their questions and maintained their positive assessment. Although we have not yet received follow-up feedback from Reviewers aoMk and dmU4, we sincerely hope that our detailed responses and additional evidence help clarify their questions and demonstrate the value of LASAR. We would be grateful if could take the time to review these additions and reconsider their assessments in light of them. We greatly appreciate all reviewers’ time, thoughtful feedback, and consideration of our work.

Thank you again for your time and consideration.

The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.