Yufan Zhao

Research profile

Research summary

Yufan Zhao is an emerging researcher in recommender systems and information retrieval, with a 2026 record centered on generative recommendation built on semantic IDs. Two arXiv preprints anchor the Evolve catalog: 'From Interests to Semantic IDs' (2026), which introduces retrieval-grounded credit assignment for generative recommenders, and 'Learning Better Reasoning for Generative Recommendation with Semantic IDs' (2026), which presents Evo-Rec, combining best-of-n reasoning supervised fine-tuning with NDCG-rewarded GRPO. Both papers situate the author as a middle-position contributor on team efforts rather than corresponding author. The methodological throughline is reinforcement-learning post-training — group-relative policy optimization, process supervision, credit assignment — applied to semantic-ID generative recommenders and evaluated on three benchmark datasets per paper. Evolve's reviews credit both works with solid ablations and measurable lifts, while flagging limits: partial significance with untested reward leakage in the first, and missing seed variance with second-hand baselines in the second; both carry medium reproducibility readiness. The OpenAlex snapshot is thin: 1 work indexed in the last three years, h-index and i10-index of 0, 7 distinct co-authors, and topic labels spanning recommender systems, advanced graph neural networks, and expert finding and Q&A systems. Radar axes for output, impact, focus, and reproducibility are unscored for lack of data, while collaboration (41) and activity (20) are modest. This is an early-stage profile; the record likely understates a just-starting publication stream.

Research keywords

Highlights

• Retrieval-grounded credit assignment for semantic-ID recommendation

  This 2026 arXiv preprint introduces retrieval-grounded credit assignment for generative recommendation. Evolve's review credits it with improving semantic-ID recommendation across three datasets and solid ablations, while noting the significance is partial and reward leakage was not tested.

• Evo-Rec: reasoning-enhanced generative recommendation with GRPO

  The second 2026 preprint, Evo-Rec, combines best-of-n reasoning supervised fine-tuning with NDCG-rewarded GRPO to lift semantic-ID generative recommendation on three Amazon datasets. Evolve notes solid ablations, but flags the absence of seed-variance reporting and reliance on second-hand baselines.

• A shared RL post-training agenda across both papers

  Both reviewed works apply group-relative policy optimization to semantic-ID generative recommenders — one via retrieval-grounded process supervision and credit assignment, the other via NDCG-rewarded reasoning training — marking a coherent early agenda in reinforcement-learning post-training for recommendation.

Timeline

1. 2026First Evolve-cataloged preprint: From Interests to Semantic IDs introduces retrieval-grounded credit assignment for generative recommendation (arXiv).
2. 2026Second cataloged preprint: Evo-Rec pairs best-of-n reasoning fine-tuning with NDCG-rewarded GRPO for semantic-ID generative recommendation (arXiv).
3. 2026OpenAlex snapshot for this profile shows 1 indexed work in counts by year and 7 distinct co-authors.

Papers in the Evolve catalog

• 2026-09-24

  From Interests to Semantic IDs: Retrieval-Grounded Credit Assignment for Generative Recommendation

• 2026-09-24

  Learning Better Reasoning for Generative Recommendation with Semantic IDs

Frequent co-authors