acceptodds
Under review as a conference paper at ICLR 2027

LARCH: Latent Attribute Reasoning through Compositional Hierarchies for Generative Recommendation

Abstract

Generative recommenders predict items through multi-token identifiers. Beyond distinguishing individual items, these identifiers should expose shared relationships useful for prediction. This raises two connected questions: how to organize predictive information into shared components, and how to use those components to supervise intermediate computation and assign policy credit. We introduce LARCH, which derives latent supervision from target prefixes and policy feedback from sampled output words. Attribute-Compatible Set Supervision (ACS) trains continuous latent states through probability mass over entities compatible with the target prefix, without annotated reasoning traces. Structure-Aware Reinforcement Learning assigns token-specific credit according to output-component roles, with prediction-dependent selection of auxiliary latent supervision. For next-POI recommendation, where prediction depends on geographic, semantic, and behavioral relationships among locations, LARCH organizes these signals into shared Mobility Word components while preserving each POI’s identity. Across three location-based social network datasets, LARCH improves Acc@1 over our reproduced ROS baseline by an average of 2.52 percentage points. On Amazon Beauty, a Product Word instantiation surpasses the strongest reported baseline, demonstrating cross-domain use of structure-derived latent supervision.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Generative Recommendation, Latent Reasoning, Compositional Hierarchies, Structured Prediction, Reinforcement Learning, Large Language Models, Next-POI Recommendation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_QxPoOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes LARCH, a generative recommendation framework that enables latent attribute reasoning through compositional hierarchies of item identifiers. It targets two open challenges in generative recommendation: how to organize predictive information into shared components of tokenized item identifiers, and how to leverage these components to supervise intermediate latent computation and assign policy credit for reinforcement learning. The core contributions are threefold:

  1. Attribute-Compatible Set Supervision (ACS): A structure-derived latent supervision approach that trains continuous latent states by maximizing probability mass over all entities sharing the target prefix. It requires no annotated reasoning traces, and the same objective applies to both singleton and multi-entity compatible sets.
  2. Structure-Aware Reinforcement Learning: A token-specific credit assignment mechanism that aligns policy feedback with the structural role of each output component, paired with adaptive latent supervision that dynamically selects which prefix levels to supervise based on sampled prediction outcomes.
  3. Compositional entity representations: A 6-token Mobility Word for next-POI recommendation (encoding coarse/fine geography, function, local context, mobility behavior, and exact identity) and a 3-token Product Word for e-commerce recommendation. Both representations support the prefix-based supervision hierarchy while preserving unique entity identification. Experiments are conducted on three location-based social network datasets (NYC, CA, TKY) for next-POI recommendation and the Amazon Beauty dataset for product recommendation. LARCH improves Acc@1 over the reproduced SOTA baseline ROS by an average of 2.52 percentage points on mobility datasets, and outperforms the strongest reported baseline S²GR on Amazon Beauty, demonstrating cross-domain transferability of the framework.
Critical strengths
  1. Conceptually elegant and general latent supervision paradigm The ACS objective is a clean, general solution for intermediate supervision that directly exploits the prefix structure of compositional item identifiers. It eliminates the need for manually annotated reasoning traces or separate clustering steps, and naturally unifies singleton and multi-entity supervision under one mathematical formulation. This is a conceptually novel contribution that can be adapted to various generative recommendation tasks beyond the two domains tested in the paper.

  2. Nuanced structure-aware policy refinement design The combination of token-specific credit assignment and adaptive latent supervision is well-motivated and technically thoughtful. By matching reward signals to the structural decision encoded by each output token, and dynamically adjusting supervision granularity based on sampling outcomes, it addresses a common shortcoming of flat sequence-level RL in structured generation tasks. The ablation results clearly validate that each component contributes non-trivially to the final performance gain.

Critical weaknesses
  1. Limited manifestation of set-level supervision in the main mobility scenario In the core next-POI experiments, over 97% of prefix-compatible sets at the GH level are singletons across all three mobility datasets (Appendix D). This means ACS degenerates to exact-entity supervision on the vast majority of training samples, so the core advantage of "set-level probability aggregation" is not fully exercised in the main results. The supplementary coarser-prefix experiment shows only a small gain (~0.28 Acc@1 points) over full exact supervision, leaving the practical impact of ACS's set-level property underwhelming in the mobility domain.

  2. Incomplete coverage of recent competitive baselines The POI benchmark misses several recent generative recommendation methods (e.g., SPAR, DIGER) that represent strong competing approaches. On Amazon Beauty, the comparison does not include very recent works focused on latent reasoning (e.g., HiLaR, SIDReasoner). This gap makes it difficult to fully assess how LARCH positions itself against the full state of the art in generative recommendation.

Questions for authors

Clarification questions

  1. In the main mobility experiments, could you break down the accuracy gain of ACS over Full/Full/Full supervision into contributions from samples with multi-entity compatible sets versus samples with singleton sets? It would help clarify whether the gain comes primarily from set-level supervision on multi-entity cases, or from improved training stability on singleton cases.
  2. For adaptive latent supervision in Stage 3, the "all-incorrect" branch activates the earliest failing prefix level. Intuitively this applies coarser supervision when all samples fail, yet static all-level supervision performs worse. Could you explain the core mechanism why adaptive selection outperforms applying all supervision levels uniformly?
Reviewer_BKooOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper considers generative recommendations (GR). The authors proposes LARCH, which derives latent supervision from target prefixes and policy feedback from sampled output words. Experiments results on Amazon datasets are provided to support their claims.

Critical strengths

The presentation of this paper, including the figures, formulas and tables are in general clear.

Critical weaknesses
  1. There are a lot of terms undefined or difficult to follow. For example "Attribute-Compatible Entity Set", "ACS (Attribute-Compatible Set Supervision)", "Mobility Word", "next-POI recommendation", etc. This raises a concern that the paper might use a lot of AI-generated words without verifying, understanding, and polishing them.

  2. The authors directly start to write their proposed methods without too much motivation and intuition. For example why we should define ACS in (2) and (3), and why we should use a 3-stage algorithm to tackle GR.

  3. There are 5 RQs in total but the numbers to answer those questions seem to be from nowhere. For example there is no reference to a

  4. No code provided.

Questions for authors

Can the authors provide more comments or insights on why many words are used unclearly without definition, and can the authors disclose the AI usage in writing this paper?

Reviewer_iUiJOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces LARCH, a framework that leverages compositional hierarchies of item identifiers to provide intermediate supervision and structured policy credit for generative recommendation. The authors propose Attribute-Compatible Set (ACS) Supervision, which trains continuous latent states by maximizing the probability mass over entities sharing a target prefix, eliminating the need for annotated reasoning traces. Furthermore, they introduce Structure-Aware Reinforcement Learning, which assigns token-specific credit based on output-component roles and adaptively selects latent supervision levels using sampled predictions. For next-POI recommendation, the authors design a six-token Mobility Word integrating geographic, functional, behavioral, and identity components. Experiments on three POI datasets and Amazon Beauty demonstrate that LARCH outperforms existing baselines, with detailed ablation studies validating its individual components.

Critical strengths

Strengths:

1) The formulation of Attribute-Compatible Set (ACS) supervision is theoretically elegant, providing a unified objective that naturally handles both singleton and multi-entity compatible sets without requiring expensive human-annotated reasoning traces.

2) The Token-Specific Credit Assignment mechanism effectively addresses the limitations of sequence-level reinforcement learning by aligning policy feedback with the distinct structural roles of different identifier components.

3) The proposed Mobility Word thoughtfully integrates heterogeneous signals—geography, function, local context, behavior, and identity—into a single ordered representation, offering a strong structural prior for next-POI recommendation.

4) The empirical evaluation is exceptionally rigorous, featuring highly controlled matched ablations across training stages, multiple random seeds, and cross-domain validation on both mobility and product recommendation tasks.

Critical weaknesses

Weaknesses:

1) The actual supervisory signal provided by ACS is severely undermined by the data statistics revealed in Appendix D. Table D.8 shows that for the primary POI datasets, the compatible sets are singletons in over 97% to 99% of training examples. Consequently, the proposed hierarchical set supervision mathematically collapses into repeated exact-entity supervision for almost the entire dataset, invalidating the core premise that ACS provides distinct intermediate structural guidance during canonical mobility training.

2) The overall framework suffers from extreme complexity and hyperparameter fragility. The three-stage pipeline relies on an intricate dynamic gradient-norm controller, specific temperature scaling, cross-state consistency regularization, and adaptive branch logic. This heavy engineering burden makes the method difficult to reproduce, tune, or deploy, especially when the marginal performance gains over simpler methods are relatively small.

3) The baseline comparisons for the POI task are potentially misleading. The authors reproduce ROS but report numbers significantly lower than the original ROS paper (e.g., 41.90% vs. 44.78% on NYC), attributing this to environment differences. By comparing LARCH against a degraded reproduction rather than the originally reported state-of-the-art, the claimed 2.52 percentage point improvement becomes unconvincing and potentially artificially inflated.

4) The Adaptive Latent Supervision mechanism in Stage 3 introduces problematic training dynamics. Conditioning the application of auxiliary latent losses on the discrete outcomes of sampled rollouts creates high-variance, non-stationary gradients. While the Train-calibrated static control performs worse, the paper fails to adequately analyze whether the adaptive version's superiority comes from genuine structural learning or simply acts as a noisy regularizer that prevents overfitting in the RL stage.

Questions for authors

Questions:

1) Given that over 97% of the canonical mobility training examples result in singleton compatible sets where ACS reduces exactly to standard cross-entropy, how can the authors justify the computational overhead of ACS for POI recommendation, and does the method offer any real benefit outside of the Amazon Beauty dataset?

2) If the reproduced ROS baseline were optimized to match its originally published performance, would LARCH still achieve statistically significant improvements, or would the gains fall within the margin of error given the small absolute differences?

3) How sensitive is the final performance to the dynamic gradient-probe loss-weight controller? Specifically, what is the performance degradation if fixed, manually tuned loss weights are used instead of the complex RMS-based adaptive calibration described in Appendix A.2?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.