acceptodds
Under review as a conference paper at ICLR 2027

ControlRec: Editable Intent Modeling for Controllable Generative Recommendation

15 views

Abstract

Generative recommender systems directly generate items from behavioral history, but most leave user intent under-specified and difficult to revise. Sequential recommenders primarily infer preferences from past behavior, while conversational recommenders elicit immediate needs but often make limited use of long-term behavioral preferences. We propose ControlRec, a unified framework for generative recommendation with implicit and explicit intent guidance. Without dialogue, it infers personalized intents by softly assigning historical interactions to shared preference prototypes; with dialogue, it updates these intents from natural-language feedback while preserving behavioral personalization. ControlRec comprises three core designs: (1) A semantic alignment adapter bridges natural language and the semantic IDs (SIDs) space, enabling expressed preferences to guide item generation. (2) A multi-intent condition encoder constructs personalized intent conditions from historical interactions and shared preference prototypes, and updates an editable intent state from natural-language feedback when dialogue is available, preserving behavioral personalization. (3) An intent-factorized diffusion decoder guides generation via parallel intent-specific denoising, preserving distinct concurrent preferences while enforcing shared negative constraints. Experiments on three Amazon domains and evaluations derived from ReDial demonstrate competitive next-item recommendation performance and effective intent refinement. On a challenging target-recovery subset, successive feedback progressively improves target recovery up to 3.22 times, and on ReDial the model achieves strong positive, negative, and multi-intent control. Furthermore, ControlRec supports freely expressed natural-language requests, providing a promising interface for controllable generative recommendation. Code will be released soon, and an interactive online demo is available.

Primary area:
generative models
Keywords:
controlled recommendation, generative recommendations

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

The paper proposes ControlRec, a generative recommender which combines a semantic adapter aligning natural language to semantic IDs, a multi-intent condition encoder, and an intent-factorized discrete diffusion decoder. The main contribution of the paper is an intent-mediated recommender which effectively connects language feedback with behavior history for recommendation. Validation includes next-item recommendations on three Amazon review datasets, an Amazon recovery dataset built from HSTU top-100 misses, and the ReDial dataset. The paper reported target-consistent feedback raises performance metrics on the datasets, and conducted ablations on design components.

Critical strengths
  1. The paper provides a coherent mechanism to effectively combine user behavior history and conversational steering via persistent and editable intents.
  2. The empirical study demonstrated that overall the mechanism works, for both positive and negative steering.
Critical weaknesses
  1. The critical evidence for conversational recovery is based on target metadata, not natural user feedback, which largely limits the soundness of the result.
  2. The ablation study lacks detailed numerical results and the explanation of individual design choice's contribution is not convincing.
  3. Weak presentation: the main body of the paper lacks necessary details, e.g. discussions on inference efficiency which is an important aspect lives only in the appendix.
Questions for authors
  1. Can the authors provide stronger evidence based on real human-authored feedback, ideally ambiguous, diverse, and noisy enough?
  2. The effectiveness of intent control is solely validated based on the discrete diffusion model. Would it be possible to provide evidence on an autoregressive SID decoding framework?
  3. The paper should add the missing citation for HSTU. In addition, the linked online demo is unavailable.
Reviewer_3cczOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ControlRec extends the LLaDA-Rec masked-diffusion SID recommender with an editable intent state. It has three parts.

  1. Semantic Alignment Adapter. Sentence-T5 text embeddings and SID embeddings are fed to a shared Transformer trained on six tasks. At inference, the adapter maps a validated category's text to an embedding z(c) and depth-wise SID logits B(c).
  2. Intent state.
    • Without dialogue, history items are soft-assigned to learned prototypes, and the top-4 prototypes become intent slots.
    • With dialogue, a frozen Qwen2.5-7B parses each utterance into Enhance/Suppress/Refine/Undo operations over catalog categories. A deterministic updater keeps positive and negative category maps.
  3. Intent-factorized decoding. Each positive intent drives a separate denoising branch through cross-attention. Negative categories subtract SID logits in every branch. Branches are merged with max(score + log π).

Experiments:

  • dialogue-free next-item recommendation on three Amazon-2023 domains;
  • a hard-case subset of 43,577 HSTU misses@100, with LLM-simulated three-turn feedback derived from the target's category path;
  • positive, negative and multi-intent control on ReDial.
Critical strengths

1 Relevant problem. It combines long-term behavioral personalization with explicit, revisable user constraints in a generative recommender. The editable-state abstraction (with Undo and Refine semantics) is clear and inspectable.

2 Reasonable decoding design with direct support. App. B.3 compares Joint conditioning with Prob. Slate under paired-bootstrap CIs and breaks the results down by number of intents.

3 Useful controls. These include a shuffled-feedback control (B.2), a matched tag-reranking control (B.5), parser reliability (B.4) and a latency breakdown (B.1), plus a working demo.

Critical weaknesses

1. By the paper's own Table 6, a trivial category filter matches the main controllability result (RQ2, Table 2).

  • The feedback is derived from the target's category path. The target is therefore always inside the "T3: Fine + Suppression" compatible set.
  • Table 6 reports a median of 144 compatible items, so at least half of the cases have ≤ 144 candidates.
  • Returning 10 uniformly random compatible items thus gives R@10 ≥ 0.5 × 10/144 ≈ 0.035. That already matches ControlRec's T3 R@10 of 0.0344.
  • The true value of this random baseline is necessarily higher. Every case with fewer than 144 compatible items (the minimum is 2) contributes more than 10/144. It could plausibly exceed even ControlRec + Tag Reranking (0.0615, Table 11).
  • A filter that ranks the compatible items by a history-only model's full-catalog score (HSTU or LLaDA-Rec with a hard category constraint), or by popularity, should be stronger still.

The "LLaDA-Rec + Tag Reranking" reference only reorders candidates that were already generated, so it is not this baseline. Without "history-only scorer + hard category filter over the full catalog" and "random/popular within the compatible set", RQ2 does not show that the intent machinery uses the feedback well. As it stands, it suggests the feedback is used less effectively than a filter would use it. If Table 6 is computed on a different population than the 43,577 cases, please say so.

2. The ReDial control metrics are largely satisfied by construction, and the baselines lack the same structure. ControlRec receives a parsed, catalog-validated genre constraint and applies hard logit suppression. KBRD, KGSF, BARCOR, UniCRS and the LLMs receive raw dialogue. Sat.@10 ≈ 0.99 and Viol.@10 ≈ 0.01 therefore mostly reflect the deterministic constraint rather than learned control. Fair comparisons would be each baseline with the same parsed-genre filter, or ControlRec's parser combined with a simple filter on a non-diffusion recommender. The test sets are small (175/150/150 instances) and come without CIs. The 0.2914 vs. 0.2571 R@10 gap is about 6 instances, and All-Req 0.98 vs. 0.967 is 2 instances.

3. The Amazon hard-case baselines look non-functional or structurally disadvantaged.

  • KGSF, BARCOR and UniCRS have T0 R@10 ≈ 0.0001. These are KG- and ReDial-oriented CRSs. The paper does not describe how they were adapted to Amazon industrial products: which knowledge graph, and which item vocabulary.
  • The LLM-agent baselines get worse with feedback (Llama-3-8B: 0.0087 → 0.0037 at T1).
  • B.1 says the agents rank "prepared Top-100 candidates". If those candidates come from HSTU, the target is absent by construction, because the subset consists of HSTU misses@100.

Please specify each baseline's candidate pool and adaptation.

4. Where the dialogue-free gains come from is not ablated. The RQ1 improvements over LLaDA-Rec (e.g., R@1 0.0098 → 0.0115 on Industrial) could come from any of three sources:

  • (a) the extra multi-task alignment pretraining (MaskItem/NextItem/Match/Rank, with item text);
  • (b) four-branch decoding with merging, which is effectively a test-time ensemble;
  • (c) the prototype "intents" themselves.

The Fig. 3 ablation covers only the hard-case dialogue protocol. Please ablate RQ1 and add a compute-matched LLaDA-Rec (four decoding runs or beam groups, merged). Two further points:

  • The seed standard deviations (±0.00005 on R@1) suggest that only a small module varies across seeds. Is the backbone a single fixed checkpoint?
  • A.7 suggests the baseline numbers are copied from prior work, e.g. LLaDA-Rec (Shi et al., 2025). If so, please confirm identical splits and preprocessing.

5. The free-form-language claim is not what is evaluated. The abstract and contributions emphasize "freely expressed natural-language requests". The evaluated pipeline, however, collapses every utterance into a validated catalog category ID. The free-form mechanism shown in the demo (B.6: matching requests against titles and plot summaries, then aggregating into SID logits) is a different mechanism with no quantitative evaluation. The parser accuracy in B.4 is measured on synthetic, category-level trajectories. The "real case exploratory user study" in B.6 is three hand-picked examples, not a user study.

6. The novelty is mostly integration, and the related work misses directly relevant lines.

  • Per-interest retrieval with weighted aggregation is ComiRec's aggregation module (Cen et al., 2020). The paper cites ComiRec but does not discuss this overlap.
  • The related work omits user-controllable, editable and scrutable recommendation: UCRS (Wang et al., SIGIR 2022), LACE with editable user profiles (Mysore et al., SIGIR 2023), and scrutable user models (Balog et al., SIGIR 2019).
  • Parsing language into structured operations is common in tool-using recommender agents, e.g. InteRecAgent, which the paper times in B.1.
  • Suppressing negative attributes through logits is standard constrained decoding.

"We first introduce a unified intent-mediated formulation" and "we first achieve intent-guided parallel diffusion generation under conversational control" overclaim.

7. The necessity of the Semantic Alignment Adapter is unclear. At evaluation time the parser outputs a catalog category c, and the adapter only maps c's text to (z(c), B(c)). A training-free alternative would derive B(c) from the empirical SID-code distribution of items in c. This alternative is not compared, and the paper does not state what replaces the adapter in "w/o Align". SID codes are also shared across categories, so suppressing a category's codes can suppress unrelated items. The collateral recall loss on non-suppressed items should be measured.

Questions for authors
  1. What do "HSTU/LLaDA-Rec full-catalog scores + hard category filter" and "random/popular within the compatible set" achieve at T1–T3? Is Table 6 computed on the 43,577 hard cases?
  2. What candidate pool does each CRS and LLM baseline rank on Amazon, and how were the KG-based CRSs adapted?
  3. Can you provide the RQ1 ablation and a compute-matched LLaDA-Rec? Are the Table 1 baselines copied from prior work?
  4. On ReDial, what is a user's "behavioral history"? Please give CIs for Table 3.
  5. Is there a quantitative evaluation of the free-form, description-matching mechanism used in the demo?
  6. Does it go to production? Any AB test, what is the gain?
Reviewer_NBGyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ControlRec combines behavioral history with an editable intent state for controllable generative recommendation. It uses language–SID alignment, structured preference updates, and intent-specific diffusion branches to incorporate positive and negative feedback. Experiments demonstrate competitive history-only recommendation and improvements in feedback-guided target recovery and intent coverage.

Critical strengths

1.The paper addresses the practical need to revise recommendations while retaining behavioral context. 2. Useful supporting experiments. Shuffled-feedback experiments show that improvements depend on relevant feedback. Applying identical tag reranking to different candidate sources helps distinguish generation gains from postprocessing effects.

Critical weaknesses
  1. The approach seems similar to disentangled collaborative filtering and intent-based recommendation, without introducing improvements specifically tailored to SID-based generative recommendation; see [1, 2].
  2. How can the authors establish that the performance gains do not simply result from the additional parameters introduced by the Intent Cross-Attention component?
  3. The main text lacks a systematic hyperparameter sensitivity analysis. No systematic experimental comparison is provided for different values of the additional hyperparameters introduced by the method. Such important analysis should be included in the main text.
  4. Comparisons with more recent baselines are missing. The most recent baseline, LLaDA-Rec, dates from 2025.
  5. The case study refers to “Qwen 2.5-VL 7B/72B,” whereas Table 2 lists Qwen2.5 text models. These model names are inconsistent and should be clarified. [1] Disentangled Contrastive Collaborative Filtering. [2] Exploring the Individuality and Collectivity of Intents behind Interactions for Graph Collaborative Filtering.
Questions for authors

Refer to weakness.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of additional results and request for feedback

Dear Reviewers 3ccz, NBGy, and pKhr,

As the author–reviewer discussion closes on November 18 and we have not yet received any follow-up feedback, we briefly summarize the clarifications and additional results addressing your concerns.

  • Full-catalog filtering comparisons (3ccz). We added random, popularity-based, and HSTU-based filtering baselines using the same natural-language feedback. At T3, ControlRec with tag reranking achieves R@10/N@10 of 0.06150/0.04156, compared with 0.03658/0.01946 for popularity filtering. We also clarified that Table 6 describes feedback-construction metadata, which should not be treated as target-category information available to the recommender.
  • Matched conditions and collateral effects (3ccz). We provided BARCOR and UniCRS with the same parsed feedback conditions. ControlRec retains stronger performance in negative and multi-intent control. On the 150 negative-control instances, the complete pipeline, including reranking, preserves all 51 previously retrieved non-forbidden targets and retrieves 14 additional targets, while reducing Viol.@10 from 0.4393 to 0.0100.
  • Component contributions and capacity controls (NBGy, pKhr, 3ccz). We supplied numerical component ablations, dialogue-free ablations, parameter-matched continued-training controls, and sensitivity analyses. Expanding LLaDA-Rec alone does not reproduce ControlRec’s gains. We clarified the distinction between SID–text alignment pretraining and recommendation training.
  • Autoregressive SID validation (pKhr). We adapted intent control to TIGER and evaluated it on the same control tasks. It achieves 0.8667 All-Req@10, providing preliminary evidence beyond diffusion, although the current diffusion implementation performs better.
  • Real user feedback (pKhr, 3ccz). We added a Tencent News evaluation involving 5,000 users in total. Completion rate increases from 77.21% for the history baseline to 78.88% with inferred-intent guidance; the subset providing real feedback achieves 81.40%. User requests include “I hate snakes—they’re scary” and “Stop recommending celebrities,” without supplied target-category information.
  • Additional baselines and presentation (NBGy, pKhr). We added DiffGRM and FLASH comparisons, clarified model names and baseline implementations, and provided efficiency details and demo restart instructions. A central clarification is that ControlRec uses natural-language preferences and behavioral history to guide full-catalog generation; predefined target categories are not required to restrict its generation space.

We would appreciate your feedback on whether these responses address your concerns. If any issues remain, please let us know so that we can clarify them before the discussion closes. If the additional evidence changes your assessment, we would be grateful if you could update your review accordingly.

Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.