Mengdan Zhu

Research profile

Research summary

Mengdan Zhu is an early-stage researcher working at the intersection of recommender systems, information retrieval, and reinforcement learning. The current profile is anchored on two 2026 first-author arXiv papers in the Evolve catalog, both attacking one problem: how to make semantic-ID-based generative recommendation learn better. The first, 'From Interests to Semantic IDs', reframes credit assignment as a retrieval-grounded task, applying group-relative policy optimization and process supervision to lift recommendation on three datasets; Evolve credits it with good ablations while flagging partial significance and untested reward leakage. The second, Evo-Rec, combines best-of-n reasoning supervised fine-tuning with an NDCG-reward GRPO stage, improving SID-based generative recommendation on three Amazon datasets, again with solid ablations but no seed-variance reporting and second-hand baselines. Together the papers define a coherent agenda: importing modern RL post-training machinery — GRPO, process supervision, rejection sampling, chain-of-thought reasoning — into generative recommenders. Collaboration data shows a compact circle of 7 distinct coauthors (including Yufan Zhao, Sridhar Iyer, and Liang Zhao) and a collaboration score of 41, consistent with focused small-team work; both papers are first-authored but not marked corresponding. Five of six radar axes — output, impact, focus, activity, reproducibility — remain unscored for lack of data. As a name-ambiguous profile built on thin coverage, this should be read as an early-stage snapshot rather than a settled record.

Research keywords

Highlights

• Retrieval-grounded credit assignment for semantic IDs

  2026 first-author arXiv work that reframes credit assignment in generative recommendation as a retrieval-grounded task, applying group-relative policy optimization and process supervision; Evolve reports gains across three datasets with good ablations, while noting partial significance and untested reward leakage.

• Evo-Rec: reasoning-enhanced generative recommendation

  Combines best-of-n reasoning supervised fine-tuning with an NDCG-reward GRPO stage; Evolve finds consistent gains for SID-based generative recommendation on three Amazon datasets, with solid ablations but no seed-variance reporting and second-hand baselines.

• A coherent RL post-training agenda for generative recommenders

  Across the two 2026 first-author papers, the same toolkit recurs — GRPO, process supervision, rejection sampling, chain-of-thought reasoning — defining a focused agenda of importing RL post-training into semantic-ID generative recommendation, with medium repro-readiness on both.

Timeline

1. 2026First-author arXiv paper 'Learning Better Reasoning for Generative Recommendation with Semantic IDs' (Evo-Rec): best-of-n reasoning SFT plus NDCG-reward GRPO, evaluated on three Amazon datasets.
2. 2026First-author arXiv paper 'From Interests to Semantic IDs': retrieval-grounded credit assignment with group-relative policy optimization and process supervision, improving SID recommendation on three datasets.
3. 2026Both catalog papers enter Evolve review with medium repro-readiness tags, establishing a focused 2026 research line on semantic-ID generative recommendation within a 7-person collaboration circle.

Papers in the Evolve catalog

• 2026-09-24first author

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

• 2026-09-24first author

  Learning Better Reasoning for Generative Recommendation with Semantic IDs

Frequent co-authors