PERSONALIZATION · 2026-10-16

Personalization — Paper Daily, 2026-10-16

Community search over UGC posts, questions, answers and comments. Priorities: nDCG, CTR, semantic matching, query understanding, short-text retrieval, ranking and LLM-assisted search.

4.21
LEAD STORY · Beihang University / NTU / Chengdu University of Information Technology

1. Learning What to Keep: Transfer-Aware Personal Skill Learning from User Feedback

Beihang UniversityNTUChengdu University of Information Technology

Feedback mixes lasting preferences with one-off corrections. SkillGrain learns which atomic rules to retain and jointly trains skill induction and response generation. On one backbone, it improves the five-task average by 1.61 points over the strongest baseline.

Key insight: Reward rule quality and retention separately. A correct correction can still be unsuitable for future requests; preference stores benefit from explicit scope.

4.15

2. Bridging the Topological Divide: Dual-Branch Graph Distillation with Semantic Priors for Robust Cold-Start Recommendation

HKUST (Guangzhou)

DualGCN gives interaction-free items a representation through semantic-anchored graph distillation. A separate student learns directions, while inference restores vector magnitude. On Sports cold items, Recall@20 exceeds DropoutNet by 20.8%.

Key insight: Separate direction from magnitude: content predicts an item’s location more readily than its interaction-driven scale. The cold-only evaluation does not establish mixed-catalogue ranking quality.

📖 Read the deep dive →
4.15

3. Revisiting Generative vs. Dense Retrieval: A Static Value Prior Reverses the Gap

Shopify

A controlled study separates exact-target recall, best-item value and whole-list value. A static value prior reverses generative retrieval’s apparent value lead; on industrial data it improves dense retrieval NDCG@50 by 11.8%.

Key insight: Hold the popularity or value prior constant before attributing gains to personalization. Best-item value and whole-list relevance can favor different retrievers.

3.90

4. PersEvalBench: Can LLM Evaluators Judge Contextually Appropriate Use of Persona Information?

AmazonUIUCUCL

PersEvalBench tests whether evaluators can judge the appropriate use of user history through controlled response pairs. On the human-audited subset, the strongest reported judge detects only 63.9% of under-personalization errors.

Key insight: Distinguish omitted preferences from inappropriate reuse. These failures need opposite fixes, so an aggregate satisfaction score hides useful diagnostics.

3.85

5. Learning Individual Scientific Taste with Privileged Hindsight Distillation

Tsinghua University

The study separates shared, evidence-grounded proposal assessments from individual preference mappings. Hindsight is available only during teacher training. The complete method reaches 0.693 on the reported real-researcher RW-SP metric.

Key insight: Share factual assessment while personalizing trade-offs. Future evidence must remain outside inference inputs; the experiment focuses on research proposals rather than consumer recommendation.

3.80

6. xLightGCN: Explainable Graph Neural Network Based Recommendation Systems with Most Influential Peers

University of Waterloo

xLightGCN decomposes LightGCN-family scores into influential graph peers and quantifies unexplained residuals. Gowalla experiments average each setting over 100 user-item pairs; larger explanation subgraphs cover more of the score.

Key insight: Use the model’s actual computation as explanation evidence. Exact decomposition depends on linear propagation and does not extend directly to arbitrary recommenders.

3.60

7. Calibration and Risk-Budgeted Decisions with System One Models

The paper calibrates binary decisions with risk budgets and conservative fallback actions. Across six moderation, safety and reranking datasets, Learn Then Test exceeds the target risk in 7.8% of settings with defined test risk.

Key insight: Fit thresholds to the risk of the action actually taken. Calibration plots alone do not establish that an automated acceptance subset respects its error budget.

3.40

8. Similarity-aware generalization framework for zero-shot cross-domain sequential recommendation

SAGERec selectively aligns semantically compatible items and adapts regularization to domain discrepancy for zero-shot cross-domain sequential recommendation. The publisher abstract reports gains but gives no numerical effect sizes.

Key insight: Selective transfer preserves domain distinctions. Evidence is limited to the publisher abstract; the full page returned HTTP 403 and effect sizes could not be checked.

3.40

9. MgRec: Multi-granularity intent interaction based sequential recommendation method

MgRec combines Transformer-based sequential preferences with coarse-grained intents extracted through a knowledge graph and a neural topic model. The publisher abstract claims stronger recommendations and reduced popularity bias without numerical results.

Key insight: Keep recent behavior and broader intent as distinct signals before fusion. The accessible abstract does not establish the gain over simpler fusion baselines.

3.25

10. Intern-Decision: Efficient Multimodal Decision Making with Autoregressive Masked Language Modeling

Shanghai AI LabSJTUFudan UniversityNTU

Intern-Decision returns structured closed-set choices and candidate probabilities in one causal forward pass. Its 4B model averages 90.02% accuracy across seven evaluation suites. These are decision tasks, not recommendation CTR experiments.

Key insight: A compact choice model can support low-latency auxiliary judgments. Routing confidence still needs task-specific calibration.