Liang Zhao

Research profile

Research summary

This is an early-stage profile for Liang Zhao, name-resolved from public bibliographic data: the Evolve catalog holds two 2026 arXiv preprints, both on generative recommendation with semantic IDs, with Zhao in the last author position on each. The first, 'From Interests to Semantic IDs: Retrieval-Grounded Credit Assignment for Generative Recommendation,' proposes retrieval-grounded credit assignment; Evolve's review reports improved semantic-ID recommendation across three datasets, with good ablations but partial significance testing and untested reward leakage. The second, 'Learning Better Reasoning for Generative Recommendation with Semantic IDs,' presents Evo-Rec, pairing best-of-n reasoning supervised fine-tuning with NDCG-reward GRPO; the review finds gains on three Amazon datasets with solid ablations, while flagging missing seed-variance reporting and second-hand baselines. Both papers carry a medium reproducibility readiness rating. Together they indicate a coherent research program: applying reinforcement learning — group-relative policy optimization, process supervision, chain-of-thought reasoning, and rejection sampling — to semantic-ID-based generative recommenders. Collaboration metrics record 7 distinct coauthors and a collaboration score of 41 out of 100, pointing to a small, stable working group. No institutional affiliation, country, ORCID, citation counts, or output-volume statistics are present in the input data, so output, impact, focus, activity, and reproducibility axes remain unscored pending more coverage.

Research keywords

Highlights

• Retrieval-Grounded Credit Assignment for Semantic-ID Recommendation

  Evolve's review of this 2026 arXiv paper reports that retrieval-grounded credit assignment improves semantic-ID generative recommendation across three datasets, with good ablations, while noting partial significance testing and untested reward leakage.

• Evo-Rec: Reasoning Fine-Tuning with NDCG-Reward GRPO

  The review finds that Evo-Rec's combination of best-of-n reasoning supervised fine-tuning and NDCG-reward GRPO lifts semantic-ID generative recommendation on three Amazon datasets, with solid ablations, though seed variance is unreported and baselines are second-hand.

• A Consistent Reinforcement-Learning Research Line

  Both 2026 papers apply reinforcement learning — group-relative policy optimization, process supervision, chain-of-thought reasoning, and rejection sampling — to semantic-ID generative recommenders, with Zhao as last author on each, indicating a sustained, supervised research program.

Timeline

1. 2026Publishes 'From Interests to Semantic IDs: Retrieval-Grounded Credit Assignment for Generative Recommendation' on arXiv as last author; Evolve's review finds gains on three datasets with good ablations but partial significance.
2. 2026Publishes 'Learning Better Reasoning for Generative Recommendation with Semantic IDs' (Evo-Rec) on arXiv, again as last author; reviewed as lifting SID generative recommendation on three Amazon datasets with solid ablations.
3. 2026Accumulates 7 distinct frequent coauthors across the cataloged work — including Mengdan Zhu, Yufan Zhao, Sophie Di, Yao Zhao, Tao Di, Yulan Yan, and Sridhar Iyer — yielding a collaboration score of 41/100.

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