MERIT: Mitigating Exposure Bias in Generative XMC for User-Interest Propensity Modeling

Abhinav Mahajan; Arindam Sarkar; Prakash Mandayam Comar

cs.IRAug 28, 2026

MERIT: Mitigating Exposure Bias in Generative XMC for User-Interest Propensity Modeling

Authors: Abhinav Mahajan, Arindam Sarkar, Prakash Mandayam Comar

Organizations: Carnegie Mellon University Pittsburgh, PA, USA · Amazon Bangalore, India

Abstract

Matching users to interest categories at scale is central to personalized shopping, but the task is challenging in large e-commerce platforms, where label spaces continually evolve and user-interest signals are sparse and long-tailed. Autoregressive language models are appealing because their world knowledge and semantic priors over descriptors generalize across extreme label spaces and accommodate multiple valid label assignments. Yet under teacher-forced fine-tuning, inference-time predictions become part of the conditioning context: early errors steer later outputs toward co-occurring labels, over-generating near-correlates and missing unrelated true interests. We present MERIT, a framework for user-interest propensity modeling that mitigates this exposure bias through a self-correction objective. A permutation-invariant multi-target loss over shuffled mixtures of gold and mined hard-negative labels exposes the generator to erroneous prefixes while preserving the efficiency of teacher-forced training. This training objective concentrates supervision at classification positions, yielding propensity-aligned hidden states powering a lightweight scorer for bidirectional retrieval (interests for users and users for interests). On a proprietary e-commerce dataset with 250k+ interest categories, MERIT improves global recall by at least 11.9% and average Hit@k by 6.1%. In production A/B tests, it achieves +0.26% gain in user conversion.

Figures & tables

Figure 1: MERIT inference pipeline . User profile (aggregated interest-categories) and recent interactions (raw product data) are tokenized, embedded, and compressed via adaptive patching. The exposure-mitigated LLM autoregressively generates interest-categories. With mitigation (green), the model produces diverse predictions (e.g., Sports Enthusiast); without mitigation (red), it enters correlation loops (Home Décor → Home Appliances → Interior Design). Hidden states from the first token of the first few generated categories are extracted as enriched classification embeddings, which serve as input to the propensity model for scoring any interest category.

Baselines
Method	Global			Beauty			Apparel			Electronics			Sports			Toys
Method	R	F1	P	R	F1	P	R	F1	P	R	F1	P	R	F1	P	R	F1	P
XLGen	+2.1	+2.0	+1.9	+2.8	+2.2	+1.4	+3.5	+4.0	+4.7	+1.1	+0.9	+0.5	+1.7	+1.6	+1.4	+0.5	+0.4	+0.3
GROOV	+1.8	+2.3	+2.9	+2.3	+2.4	+2.4	+3.9	+5.3	+7.5	+0.2	+0.3	+0.6	+2.2	+2.6	+3.3	+2.0	+2.3	+2.7
PatchRec	+0.2	+1.2	+3.1	-2.3	-2.3	-2.1	-0.2	+1.3	+5.1	-1.0	-0.9	-0.6	-1.1	-0.6	+0.7	-1.2	-1.1	-0.7
DEXML-DS ( $\text{top\_k}{=}5$ )	+1.2	+2.2	+3.8	+2.6	+3.1	+3.7	+1.8	+2.6	+3.9	+1.6	+2.6	+4.5	-0.7	-0.5	-0.1	-0.3	-0.3	-0.2

Table 1: Generative classification performance ( P recision, R ecall, F1 ) and training-objective ablation. Absolute improvement over the trivial baseline across product verticals; The lower block ablates different training-objectives.

Method	Beauty			Apparel			Electronics			Sports			Toys
Method	@500	@5k	@20k	@500	@5k	@20k	@500	@5k	@20k	@500	@5k	@20k	@500	@5k	@20k
DEXML-DS	+7.8	+6.8	+5.2	+0.4	+4.4	+7.5	-2.4	+2.9	+4.7	-0.4	-0.9	+1.1	-1.6	-3.3	-3.0
MERIT (230M)	+20.8	+12.6	+21.3	+1.4	+9.3	+19.7	+3.4	+7.1	+14.0	+3.0	+7.8	+15.8	+4.2	+7.4	+15.8

Table 2: Propensity ranking performance (Hit@k). Results show absolute improvement over baseline: DEXML-Stk

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Method	$\tau$ =0.70			$\tau$ =0.80			$\tau$ =0.90			$\tau$ =exact			$\tau$ -AUC
Method	$\Delta$ R	$\Delta$ F1	$\Delta$ P	$\Delta$ R	$\Delta$ F1	$\Delta$ P	$\Delta$ R	$\Delta$ F1	$\Delta$ P	$\Delta$ R	$\Delta$ F1	$\Delta$ P	$\Delta$ R	$\Delta$ F1	$\Delta$ P
Trivial	+0.0	+0.0	+0.0	+0.0	+0.0	+0.0	+0.0	+0.0	+0.0	+0.0	+0.0	+0.0	+0.0	+0.0	+0.0
XLGen	+2.0	+2.0	+1.9	+1.2	+1.0	+0.7	+0.9	+0.7	+0.3	+0.8	+0.7	+0.5	+1.3	+1.2	+0.9
GROOV	+1.8	+2.3	+3.1	+1.2	+1.5	+1.9	+1.0	+1.2	+1.5	+1.1	+1.2	+1.5	+1.3	+1.6	+2.1
PatchRec	-0.1	+0.8	+2.5	-0.7	-0.2	+0.8	-0.8	-0.4	+0.4	+0.1	+0.4	+1.1	-0.6	+0.0	+1.1
DEXML-DS	+1.1	+2.0	+3.7	+0.7	+1.5	+2.9	+0.7	+1.3	+2.6	-0.1	-0.1	-0.1	+0.8	+1.6	+3.0

Appendix

Table 3: Generative classification performance (Precision, Recall, F1) under different prediction vs. ground-truth matching criteria (

\tau

is semantic matching threshold). Absolute improvement over the trivial baseline aggregated over all product verticals is reported.

\tau

-AUC is the normalized trapezoidal area under the improvement-versus-threshold curve over

\tau\!\in\![0.70,0.90]

$k_{g}$	@0.5k	@1k	@5k	@10k	@20k	Overall
1	16/20	17/20	18/20	20/20	20/20	91%
3	18/20	18/20	19/20	19/20	20/20	94%
5 (paper)	17/20	17/20	19/20	19/20	19/20	91%
7	16/20	16/20	19/20	19/20	19/20	89%
10	17/20	17/20	18/20	19/20	19/20	90%

Appendix

Table 4: Sensitivity to

k_{g}

: fraction of the 20 product verticals, per Hit@

k

cutoff, where the given

k_{g}

beats both DEXML-DS and DEXML-STk simultaneously. Overall aggregates all 100 vertical–cutoff cells.

Method	@0.5k	@1k	@5k	@10k	@20k	Mean
DEXML-DS	+5.5	+5.3	+4.7	+5.3	+4.9	+5.2
MERIT, $k_{g}{=}3$	+14.3	+12.0	+12.5	+16.4	+19.8	+15.0
MERIT, $k_{g}{=}5$ (paper)	+11.6	+10.3	+11.1	+14.9	+18.2	+13.2
MERIT, $k_{g}{=}10$	+11.4	+10.0	+10.5	+13.9	+16.3	+12.4

Appendix

Table 5: Mean Hit@

k

improvement (absolute, averaged across all 20 verticals) over DEXML-STk, for representative

k_{g}

settings; DEXML-DS shown for reference.

Baselines
Method	Global			Beauty			Apparel			Wireless			Drugstore			Kitchen			Home
Method	R	F1	P	R	F1	P	R	F1	P	R	F1	P	R	F1	P	R	F1	P	R	F1	P
XLGen	+2.1	+2.0	+1.9	+2.8	+2.2	+1.4	+3.5	+4.0	+4.7	+6.9	+5.8	+4.4	+4.3	+5.0	+5.8	-3.3	-3.6	-3.9	+0.9	+0.7	+0.5
GROOV	+1.8	+2.3	+2.9	+2.3	+2.4	+2.4	+3.9	+5.3	+7.5	+3.3	+3.2	+2.9	+3.4	+4.6	+6.4	-3.5	-3.4	-3.2	+1.8	+2.0	+2.2
PatchRec	+0.2	+1.2	+3.1	-2.3	-2.3	-2.1	-0.2	+1.2	+5.1	+4.5	+5.1	+5.9	-2.3	-1.9	-1.0	-4.3	-3.4	-1.8	+8.3	+11.4	+17.0
DEXML-DS	+1.2	+2.2	+3.8	+2.6	+3.1	+3.7	+1.8	+2.6	+3.9	+11.5	+15.4	+22.6	+1.8	+2.7	+4.0	-3.8	-3.7	-3.6	-0.8	-0.6	-0.3

Appendix

Table 6: Extended generative-classification results, part A (absolute pp improvement over the trivial baseline). Companion to Table 1 ; baseline short-names are defined in § 4 .

Baselines
Method	PC			Grocery			Shoes			Book			Wireless Acc.			Electronics			Home Improv.
Method	R	F1	P	R	F1	P	R	F1	P	R	F1	P	R	F1	P	R	F1	P	R	F1	P
XLGen	+0.8	+0.6	+0.2	+2.7	+1.8	+0.3	+1.3	+1.3	+1.1	+0.3	+0.2	+0.1	+1.4	+1.1	+0.8	+1.1	+0.9	+0.5	+1.4	+1.3	+1.3
GROOV	+0.3	+0.5	+0.8	+1.8	+1.7	+1.3	+2.2	+3.1	+4.9	+3.5	+4.1	+4.8	+0.7	+0.8	+1.0	+0.2	+0.3	+0.6	+0.9	+1.3	+1.7
PatchRec	-2.1	-2.1	-1.8	-3.8	-4.3	-4.6	-1.4	-0.5	+3.2	-1.8	-1.1	+0.0	-1.5	-1.4	-0.9	-1.0	-0.9	-0.6	+12.5	+15.3	+19.2
DEXML-DS	+1.9	+3.7	+8.5	-0.0	-0.8	-2.2	-0.5	-0.5	-0.6	-3.1	-2.8	-2.4	+1.1	+2.3	+5.0	+1.6	+2.6	+4.5	+0.2	+0.7	+1.2

Appendix

Table 7: Extended generative-classification results, part B (absolute pp improvement over the trivial baseline). Companion to Table 1 ; baseline short-names are defined in § 4 .

Baselines
Method	Automotive			Baby Product			Sports			Toys			BISS			Lawn & Garden			Personal Care
Method	R	F1	P	R	F1	P	R	F1	P	R	F1	P	R	F1	P	R	F1	P	R	F1	P
XLGen	+1.1	+1.0	+0.9	+5.3	+5.9	+6.6	+1.7	+1.6	+1.4	+0.5	+0.4	+0.3	+0.9	+0.8	+0.7	+2.2	+2.3	+2.5	+2.1	+1.8	+1.5
GROOV	+0.8	+1.1	+1.3	-0.4	+0.6	+1.9	+2.2	+2.6	+3.3	+2.0	+2.3	+2.7	+0.6	+0.9	+1.3	+0.5	+1.0	+1.7	+3.1	+3.5	+4.0
PatchRec	+11.3	+14.3	+18.7	-4.4	-3.9	-3.2	-1.1	-0.6	+0.7	-1.2	-1.1	-0.7	+3.0	+4.4	+6.1	+4.3	+6.0	+8.7	-2.3	-2.4	-2.3
DEXML-DS	+3.3	+4.7	+6.6	+3.4	+6.1	+10.6	-0.7	-0.5	-0.1	-0.3	-0.2	-0.2	-0.3	+0.1	+0.5	+1.6	+2.2	+3.2	+3.2	+3.6	+4.0

Appendix

Table 8: Extended generative-classification results, part C (absolute pp improvement over the trivial baseline). Companion to Table 1 ; baseline short-names are defined in § 4 .

Vertical	DEXML-DS					MERIT (230M)
Vertical	@0.5k	@1k	@5k	@10k	@20k	@0.5k	@1k	@5k	@10k	@20k
Beauty	+3.8	+4.5	+6.1	+6.5	+5.0	+20.8	+18.1	+12.6	+16.8	+21.3
Apparel	-4.8	-0.5	+2.6	+4.8	+5.4	+1.4	+2.9	+9.3	+15.5	+19.7
Wireless	+64.6	+53.6	+22.1	+25.3	+22.5	+58.4	+51.5	+21.5	+25.9	+27.7
Drugstore	+3.2	+2.4	-0.5	-1.0	-1.5	+12.4	+13.0	+9.2	+12.6	+15.2
Kitchen	-2.4	-2.4	-2.5	-3.1	-2.0	+0.0	+0.0	+0.6	+0.7	+3.1

Appendix

Table 9: Extended propensity-ranking results (Hit@

k

, absolute pp improvement over the DEXML-STk baseline) across all 20 verticals and all

k

. Companion to Table 2 ; short-names are defined in § 4 . Both method blocks are expressed relative to DEXML-STk; MERIT uses the 230M propensity model on frozen generator embeddings.

Context Strategy	Global			Beauty			Apparel			Electronics			Kitchen			Toys
Context Strategy	R	F1	P	R	F1	P	R	F1	P	R	F1	P	R	F1	P	R	F1	P
4k recent only	+13.6	+3.5	-0.1	+14.9	+3.3	-1.3	+13.2	+4.5	+1.0	+11.3	+2.9	-0.6	+11.0	-0.7	-4.1	+7.0	+2.0	+0.2
8k recent only	+13.9	+3.5	-0.2	+15.4	+3.4	-1.3	+13.6	+4.5	+0.9	+11.2	+2.7	-0.7	+11.1	-0.8	-4.2	+7.2	+1.9	+0.2
8k full history (patched)	+14.1	+3.4	-0.2	+15.5	+3.2	-1.5	+13.8	+4.5	+0.9	+11.5	+2.8	-0.7	+11.4	-0.8	-4.3	+7.3	+1.9	+0.1

Appendix

Table 10: Sequence compression ablation. Results show absolute improvement over trivial baseline. Adaptive patching enables fitting full user histories within fixed context budgets.

Campaign	Lift	95% CI	$P(\mathrm{lift}{>}0)$
Personal Care	$+0.23\%$	$(-0.12\%,\,+0.57\%)$	$0.90$
Toys	$+0.19\%$	$(-0.06\%,\,+0.43\%)$	$0.93$
Sports	$+0.37\%$	$(+0.02\%,\,+0.72\%)$	$0.98$
Average	$+0.26\%$	—	—

Appendix

Table 11: Per-campaign A/B-test results. Lift is the relative gain in purchase conversions of treatment over control. CI denotes the 95% credible interval;

P(\mathrm{lift}{>}0)

is the posterior probability of positive lift under our Bayesian covariate-adjusted framework.

User Preference Modeling

User preference modeling infers what people value from choices, behavior, and feedback to personalize recommendations, predictions, and AI actions.

See all 272 papers

Generative Long-term User Interest Modeling for Click-Through Rate Prediction
May 15, 2026Jiangli Shao, Kaifu Zheng, Hao Fang +5CTR Prediction User Modeling
MeiTuan Beijing, China
GISTBench: Evaluating LLM User Understanding via Evidence-Based Interest Verification
Mar 31, 2026Iordanis Fostiropoulos, Muhammad Rafay Azhar, Abdalaziz Sawwan +8LLM Evaluation User Modeling
Meta Recommendation Systems (MRS)
Self-supervised User Profile Generation for Personalization
Jun 3, 2026Clark Mingxuan Ju, Yuwei Qiu, Tong Zhao +1LLM Personalization Self-Supervised Learning
Snap Inc. · Bellevue, WA USA

May 15, 2026J/K move · Enter open · S save

Generative Long-term User Interest Modeling for Click-Through Rate Prediction

Jiangli Shao, Kaifu Zheng, Hao Fang, Huimu Ye+4

MeiTuan Beijing, China

CTR Prediction User Modeling Recommender Systems User Preference Modeling

Open paper AskPDF arXiv

Modeling long-term user interests with massive historical user behaviors enhances click-through rate (CTR) prediction performance in advertising and recommendation systems. Typically, a two-stage framework is widely adopted, where a general search unit (GSU) first retrieves top-

k

relevant behaviors towards the target item, and an exact search unit (ESU) generates interest features via tailored attention. However, current target-centered GSU would ignore other latent user interests, leading to incomplete and biased interest features. Additionally, the matching-based retrieval process in GSUs depends on the pairwise similarity score between target item and each historical behavior, which not only becomes time-consuming for online services as user behaviors continue to grow, but also overlooks the interaction information among user behaviors. To combat these problems, we propose a Generative Long-term user Interest model named GenLI for CTR prediction. GenLI consists of an interest generation module (IGM), a behavior retrieval module (BRM), and an interest fusion module (IFM). The IGM generates multiple interest distributions to indicate different aspects of real-time user interests, which is target-independent and incorporates interaction information among behaviors, ensuring complete and diverse interest features. The BRM selects related behaviors via a simple lookup operation, reducing the time complexity for weighting each behavior to

O(1)

. Finally, the IFM uses delicate gating mechanisms to generate interest features. Based on the generation process, GenLI improves the diversity of user interests and avoids complex matching-based behavioral retrieval, achieving a better balance between accuracy and efficiency for CTR prediction.

Abstract

User Preference Modeling

Explore similar work

Generative Long-term User Interest Modeling for Click-Through Rate Prediction

GISTBench: Evaluating LLM User Understanding via Evidence-Based Interest Verification

Self-supervised User Profile Generation for Personalization