acceptodds
Under review as a conference paper at ICLR 2027

ClueRec: Learning User Preference Evidence for Recommendation

Abstract

Large language models (LLMs) have enhanced recommendation by bringing rich semantic knowledge into user and item modeling. However, existing LLM-based recommendation efforts mainly learn from positive interactions while overlooking the alternatives shown alongside each clicked item. Because a click is made within an exposure set, differences between the clicked item and its co-exposed alternatives provide valuable evidence about the user's relative preferences. Ignoring such evidence allows models to learn what users choose, but not why one item is preferred over another. To address this limitation, we propose **ClueRec**, an LLM-based recommendation model that explicitly learns preference evidence from exposure contexts. ClueRec comprises a Semantic Mapper, a Preference Evidence Miner and a Recommender. Its Semantic Mapper encodes multimodal item content into coarse-to-fine Semantic IDs (SIDs) and aligns the newly introduced SID tokens with the LLM's semantic space. The Preference Evidence Miner organizes clicked and co-exposed items into SID trees, extracts gated multi-level evidence from their divergent branches, and progressively injects this evidence into the LLM through attention. The Recommender utilizes this evidence-enhanced user representation to support both generative recommendation and CTR prediction. Extensive offline experiments on large-scale industrial data show that ClueRec consistently outperforms strong baselines on both generative recommendation and CTR prediction. Online A/B testing in a real-world advertising scenario further demonstrates stable improvements in user engagement and platform revenue. Our anonymous code is available at https://anonymous.4open.science/r/ClueRec-2E8B.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM-based Recommendation, Generative Recommendation, Click-Through Rate Prediction, Preference Evidence Modeling, Large Language Models

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_CHUWOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes ClueRec, an LLM-based recommendation framework that learns user preferences by comparing clicked items with co-exposed alternatives. The authors argue that existing LLM-based recommendation methods mainly focus on positive interactions while overlooking useful preference information from unclicked alternatives.

The authors evaluate ClueRec using industrial data from five million users for generative recommendation and CTR prediction. The model outperforms several baselines on offline prediction metrics. A one-week online A/B test also reports a 0.571% improvement in CTR and 1.024% improvement in revenue per mille (RPM).

The main claimed contribution is using hierarchical semantic differences between clicked and co-exposed alternatives to improve preference modeling in LLM-based recommendation.

Critical strengths

The paper addresses an important recommendation problem, and the use of hierarchical semantic representations to incorporate co-exposed alternatives is potentially useful. The framework supports both generative recommendation and CTR prediction.

I appreciate that the paper evaluates the model on large-scale industrial data and includes an online A/B test. The reported improvements across multiple metrics are encouraging. The ablation studies also provide some evidence about the contribution of different components.

Critical weaknesses
  1. Limited novelty. The main idea has been extensively studied in the economics and marketing literature for more than a decade.

My biggest concern is the limited contribution relative to existing research. The paper motivates its method by arguing that recommendation models focus on what users click but overlook the alternatives they choose not to click, and therefore fail to understand why users prefer one item over another. However, this is not a new insight.

The economics literature on consumer search and discrete choice has long studied how consumers' choices among available alternatives reveal their preferences. For example, De los Santos, Hortaçsu, and Wildenbeest (AER 2012) use browsing and purchasing behavior to model consumer search and preferences. Ursu (Marketing Science 2018) and Ghose, Ipeirotis, Li (Management Science 2019) examine consumer search and purchase decisions while accounting for endogenous product rankings and interactions with social media. Over the past decades, several related research has explicitly distinguished exposure, consideration, and choice. These studies have already recognized that understanding preferences requires examining not only selected products but also the alternatives consumers encounter, while accounting for the search and exposure process.

The authors do reference some prior exposure-aware recommendation methods, including ExpoMF and RACP, but they do not sufficiently engage with this established economics literature. More importantly, I do not see clearly what new conceptual understanding ClueRec adds beyond applying this established insight to LLMs.

The technical novelty mainly involves representing item differences through hierarchical SIDs and injecting them into an LLM using gating and attention. These are reasonable engineering choices, but I find the methodological advance incremental.

I am not convinced that these relatively small improvements establish a substantial new contribution, especially given the extensive prior literature on exposure-aware preference modeling.

  1. Insufficient identification of user preferences and limited validation of the proposed mechanism.

My second concern is that the paper assumes a clicked item is preferred over all co-exposed alternatives. However, non-clicking does not necessarily indicate lower preference. Users may not have noticed or considered all alternatives, and their decisions are influenced by ranking positions, presentation, and the platform's exposure policy. The exposure set itself is also endogenous, since the platform determines which products users see based on existing recommendation algorithms.

These issues have been extensively recognized in economic search models, which distinguish consideration, search, and choice. Yet ClueRec directly treats differences between clicked and unclicked items as preference evidence without adequately addressing this distinction. As a result, the learned representations may capture the platform's historical exposure and ranking patterns rather than users' underlying preferences.

The empirical evidence mainly demonstrates improved prediction of observed clicks, not that the model learns why consumers prefer certain products. I think these are quite different claims.

Questions for authors
  1. Novelty relative to economic search models.

How is the central preference-learning idea different from the established economics literature that uses observed search, chosen alternatives, and non-chosen alternatives to estimate consumer preferences? In particular, what new understanding does ClueRec provide beyond implementing this existing idea within an LLM architecture?

  1. Identifying preference versus exposure bias.

How can the authors establish that the clicked–exposed differences represent actual user preferences, rather than the platform's ranking and exposure policy? Would the results remain under randomized exposure, position-controlled comparisons, or a change in recommendation policy?

  1. Incremental value of the architecture.

Can the authors compare ClueRec with simpler pairwise preference-learning or exposure-aware methods using the same LLM backbone and computational budget? Given the relatively small improvement over the w/o SPE ablation, what is the specific advantage of the hierarchical SID difference mechanism?

  1. Evidence of preference learning beyond prediction.

Is there any independent validation that the learned representations capture stable user preferences, rather than simply improving predictions under the existing recommendation policy? For example, can they predict choices under different consideration sets or exposure policies?

Reviewer_CoyMOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies LLM-based recommendation and argues that existing methods mainly learn from positive clicked items while overlooking the co-exposed alternatives available in the same impression. Because a click is made within an exposure set, differences between the clicked item and its co-exposed alternatives can provide evidence about why the user preferred one item over another. The authors propose ClueRec, which consists of three components: a Semantic Mapper that encodes multimodal item content into coarse-to-fine Semantic IDs (SIDs) and trains newly introduced SID token embeddings through the SID-to-text reconstruction task (referred to as SID Knowledge Injection (KI) in the ablation); a Preference Evidence Miner that organizes clicked and co-exposed items into SID trees, extracts gated multi-level evidence from their divergent branches, and injects this evidence into the LLM through attention; and a Recommender that uses the evidence-enhanced user representation for both generative recommendation (GR) and CTR prediction.

The main contributions are: (1) identifying the omission of preference evidence from co-exposed alternatives in LLM-based recommendation; (2) proposing ClueRec to extract and inject multi-level preference evidence from clicked-vs-exposed comparisons; and (3) reporting large-scale offline experiments and online A/B tests on an industrial advertising platform. The validation includes comparisons against sequential recommendation, SID-based generative recommendation, exposure-aware CTR, and LLM-enhanced CTR baselines, plus ablation studies, preference-evidence analysis, hyperparameter analysis, and a theoretical analysis in the appendix.

Critical strengths

1.The observation that LLM-based recommenders learn what users click but not why they prefer one exposed item over another is meaningful. Organizing clicked and co-exposed items into SID trees and extracting multi-level evidence from divergent branches is a clean and intuitive way to exploit exposure context. This is consequential because it addresses a real information loss in current LLM-based recommendation pipelines. 2.The combination of the Semantic Mapper’s SID-to-text reconstruction task (SID Knowledge Injection), the Preference Evidence Miner over SID trees, level-wise gating of multi-level preference evidence, and multi-stage evidence injection (MSI) through attention is technically coherent and supports both GR and CTR. The offline industrial experiments and online A/B claim suggest deployment potential. The ablation results, especially the drop when removing the Preference Evidence Miner, provide reasonable support for the core design.

Critical weaknesses

1.The abstract and introduction claim “stable improvements in user engagement and platform revenue” from online A/B testing, but the main paper provides no quantitative online results, confidence intervals, p-values, traffic split, experiment duration, or effect sizes. Offline results are only from a single proprietary industrial platform and cannot be reproduced. Table 1 reports only means without variance despite five repetitions. The improvements over the strongest baselines are also relatively small (e.g., Recall@20 0.4218 vs. 0.4126, AUC 0.7545 vs.0.7492. This is consequential because the paper’s main claim of real-world effectiveness cannot be independently verified from the presented evidence. 2.The core signal is the difference between a clicked item and co-exposed unclicked items. In real advertising logs, co-exposed items are not random alternatives; they are chosen by an upstream policy and affected by position, popularity, budget, and user attention. A click may therefore reflect display bias rather than stable preference. The theoretical analysis assumes a conditional logit model with orthogonal preference and noise, which is unlikely to hold exactly. The paper does not include position-bias correction, propensity weighting, randomization, or a debiasing discussion. This is consequential because it threatens the validity of the central “preference evidence” claim. Additional concerns: the training protocol for the recommendation stage is unclear—the SID Knowledge Injection stage freezes the original LLM and updates only SID embeddings, but it is not explicit which parameters are updated during GR/CTR training. The fairness of baselines is also not fully established, since ClueRec uses a Qwen3.5 backbone, multimodal encoder, SID Knowledge Injection, and QA mixture data, while baselines may not receive equivalent capacity or pretraining.

Questions for authors

1.Online A/B details: Please provide the full online A/B results, including each metric, relative and absolute improvements, confidence intervals or p-values, experiment duration, traffic ratio, and whether the improvement in platform revenue is statistically significant. This is necessary to support the central real-world claim.

2.History-length inconsistency: Table 3 states “Max History N=128”, while Appendix C.1 states “at most 512 impression events.” Which is correct? If both are used in different stages, please clarify exactly where each limit applies and how the evidence memory Pu=512 is truncated or sampled.

3.Training protocol: During GR and CTR training, which parameters are trainable? Are the original Qwen parameters frozen throughout, or only during the SID Knowledge Injection stage? How are the QA and SID-to-text data mixed with recommendation data for each task, and what is the computational cost of the full training pipeline?

4.Exposure bias: How do you ensure that the mined evidence reflects stable user preference rather than position bias, popularity bias, or upstream display policy? Do you include any propensity/position features, randomization, or debiasing objective? If not, how do you justify the conditional-logit and orthogonal-noise assumptions empirically?

5.Baseline fairness and variance: Were all baselines tuned with comparable model capacity, pretraining data, and SID tokenizer? Please report standard deviations and significance tests in Table 1. The current table format is unclear and does not allow a rigorous comparison.

6.SID-preference alignment: The theoretical analysis assumes additive preference features per SID level, but RQ-VAE codebooks are not trained to be preference-aligned. How do you verify that SID levels capture preference-relevant semantics rather than content compression artifacts? The SID-to-text reconstruction objective only trains input embeddings; is this sufficient to align SID tokens with preference-relevant semantics?

Reviewer_puG8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposed ClueRec, a new method of building LLM recommenders based on semantic IDs. In addition to use clicked items, ClueRec also learns from unengaged items in the same impressions.

ClueRec learns a five level SIDs from the item text and image embeddings. The SID is aligned to Qwen3.5-0.8B via knowledge injection. To leverage the impressions in learning, ClueRec finds the first level of SID difference for each click and unclick pairs, takes the diff and store them in memory and later used in attention layers.

To validate the effectiveness, ClueRec was trained on recommendation (predicting next item's SID) and CTR prediction. The model uses 5MM user data sampled over 7 consecutive days. The model beats the best baseline on recommender recall@20 (0.4128 vs 0.4126) and CTR AUC (0.7545 vs 0.7492). The model was used as user-embedding in an online A/B test which runs on 20% of Ads traffic. ClueRec improves CTR by 0.571% and RPM by 1.024%.

Critical strengths
  1. The model leverages the impression that are not clicked on SID learning, which is common and meaningful data to improve the representation.
  2. The learning method using SID difference is also cheap that adds almost no additional parameters.
  3. The proposed method provides a single user representation across recommendation tasks (recommender and CTR).
  4. The method is tested in real production systems with business metric lift
Critical weaknesses
  1. The core method does not show substantial lift. In ablation table, replacing SPE with MLP impacts the model quality minimally (-0.003 recall@20)
  2. The method is not tested in public benchmark. The only data points are from private dataset.
  3. The baseline did not include: 1) treating unclicked impressions as hard negatives or 2) list-wise loss.
  4. The online testing is a different system. It tested the embedding in online system, while offline numbers are from the recommendation tasks.
Questions for authors
  1. What are the standard deviations over the 5 runs and what are the dataset statistics? (catalog size, CTR, history length ,etc.)
  2. Do the same results hold on a public dataset/benchmark?
  3. How does this model compares with backbone trained with unblock as hard negative or list-wise loss?
  4. How are position bias, multi-click on the same impressions, and SID collisions handled?
  5. What's the online metrics confidence intervals and user embedding cost to serve?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.