PERSONALIZATION · 2026-10-21

Personalization — Paper Daily, 2026-10-21

Community UGC search: relevance (nDCG), CTR, semantic matching, user interest modeling, and ranking efficiency.

4.45
LEAD STORY · NUS / USTC

1. FLMNet: Self-Ensemble Distillation from Masks to Seeds for Single-Tower CTR Prediction

NUSUSTC

CTR self-distillation can be replaced with a computable disagreement penalty. FLMNet keeps gates on clean inputs and distills independently trained towers into one serving model. The distilled model reaches 0.8169 AUC on Criteo.

Key insight: Separate mask disagreement from seed disagreement: regularize the former within a run and distill the latter across runs. The moment formulas require gates that do not read masked intermediate activations.

📖 Read the deep dive →
4.35

2. DuoLM: Fast and Effective Personalized Recommendation with Language Models

Google DeepMindOracleUC Irvine

DuoLM precomputes one vector per item and uses a personalized ranker with user history and feedback. Shared history computation and a specialized runtime reduce online work. Throughput reaches about 26,400 candidate scores per second on one H100.

Key insight: Train reusable content representations and personalized judgments jointly. Compression pays off in end-to-end serving only when repeated history prefill, input transmission, and scheduling are reduced too.

📖 Read the deep dive →
4.15

3. When Semantics Fall Short: Persona-Guided Cross-Domain Sequential Recommendation with Large Language Models

PerCDR uses inferred user personas to transfer behavior across domains, while local semantic relations strengthen within-domain modeling. On the Movie domain, NDCG@10 improves 15.40% over the strongest baseline.

Key insight: Transfer user behavioral tendencies rather than relying on global item-text similarity. Local semantic links and persona-guided cross-domain links serve different purposes.

4.15

4. Multimodal Decoupling for Diversified Recommendation through Semantic and Structural Shaping

HIT

MOSS decouples multimodal features, alternates collaborative and semantic graph propagation, and regularizes category exposure. Its diversity-oriented configuration reaches 0.6348 Coverage@10 on Game.

Key insight: Diversity depends on representation geometry as well as the final ranking objective. Compare accuracy-oriented and diversity-oriented configurations separately rather than combining their best numbers.

4.03

5. Don't Rewrite the SID Map: Topology Preservation for Generative Retrieval in Dynamic E-Commerce Search

AlibabaNankai UniversityUSTC

Updating semantic item codes can invalidate paths learned by a generative retriever. Preserving old codes, inserting new items, and warm-starting the model yields a 1.16% GMV lift in a three-week TmallAPP test.

Key insight: Treat code stability as part of the learned model. Small-beam hit rate and the fraction of remapped old items expose drift that a wide beam can conceal.

4.00

6. TaQFormer: A Treatment-as-Query Transformer with Data-Fusion for Multidimensional Treatment Modeling in Coupon Allocation

Didi

TaQFormer models interactions among category-specific coupon discounts. Observational pretraining is followed by randomized-data adaptation and request-specific fusion. At essentially unchanged subsidy cost rate, business value per user rises 1.26%.

Key insight: A frozen observational branch preserves broad patterns while a randomized-data branch corrects policy bias. Request-specific fusion makes their roles clearer than simply mixing the logs.

3.97

7. Structure-Aligned History Learning with Subspace-Structured Semantic IDs for Generative Recommendation

Chongqing UniversityChina TelecomCAS

SAHL reconstructs masked history at granularities defined by its structured item codes. This links item representation and persistent preference learning. NDCG@10 improves 18.6–30.0% over RPG across four Amazon categories.

Key insight: Align reconstruction units with the code hierarchy. Shuffle the preceding history while retaining current-item clues to test whether reconstruction truly uses the user trajectory.

3.85

8. OneChord: Distill What You Rank, Recommend in One Pass

UCAS

OneChord aligns recommendation supervision, refinement distillation, and serving through one normalized ranking fusion. Across six Amazon datasets, average relative NDCG@5 improves 7.12% over the strongest baseline.

Key insight: Distill the distribution actually used for ranking, including its candidate set and normalization. Matching an intermediate code objective alone can leave a training–serving mismatch.

3.85

9. SpecMRL: Redesign Matryoshka Representation Learning within Parametric Spectrum

SpecMRL derives an ordered spectral basis from model parameters so compact embedding prefixes retain more useful information. On MTEB, half-dimensional representations exceed full-dimensional standard MRL.

Key insight: Compact embeddings need a useful coordinate ordering as well as useful information. Changing the readout basis can improve short vectors while retaining one shared representation model.