🧑‍🔬 Haoxiang Wang

🏛️ Affiliations

🎓Sun Yat-sen UniversityGuangzhou · China2014–2026 · 151 papersEmployment since 2015current🎓Chinese University of Hong KongHong Kong · Hong Kong2009–2026 · 75 papersEducation 2009–2012Employment 2012–2015current🏥West China Hospital of Sichuan UniversityChengdu · China2021–2026 · 17 paperscurrent🎓Sichuan UniversityChengdu · China2021–2026 · 16 paperscurrent🏛️Chinese Academy of SciencesBeijing · China2019–2026 · 6 paperscurrent🎓North China Electric Power UniversityBeijing · China2021–2026 · 6 paperscurrent🎓Southeast UniversityNanjing · China2012–2026 · 4 paperscurrent🏢China Southern Power GridGuangzhou · China2025–2026 · 2 paperscurrent🎓Xidian UniversityXi'an · China2013 · 2 paperscurrent🏥Affiliated Hospital of Southwest Medical UniversityLuzhou · China2026 · 1 papercurrent

From the bylines of their papers (OpenAlex, 2026-10-14), ORCID and Wikidata; institutions on papers are matched by machine. 100 more institutions on single bylines are not listed.

🧭 Concepts, shelves and topics

Of its works in our library: the number of works on each and its rating among them.

📅 Works by year

2020s

📄 Papers

The most cited first (OpenAlex and our library).

1. Interpretable Preferences via Multi-Objective Reward Modeling and Mixture-of-Experts (2024 · 142 citations)
   Reward models (RMs) in RLHF usually learn from binary pairwise preferences with a Bradley-Terry objective.
2. Mitigating the Alignment Tax of RLHF (2024 · 82 citations)
   RLHF makes LLMs follow human preferences, but it also tends to erode skills learned in pre-training. This loss is called the alignment tax. The authors measured it by aligning OpenLLaMA-3B on the Anthropic helpfulness/harmlessness data with Rejection Sampling Fine-tuning (RSF) and DPO. They then tracked performance…
3. Arithmetic Control of LLMs for Diverse User Preferences: Directional Preference Alignment with Multi-Objective Rewards (2024 · 66 citations)
   Standard RLHF compresses human judgments into one scalar reward. This pushes a model toward an "average user" and ignores people whose tastes conflict, such as those who want useful but short answers. The authors propose Directional Preference Alignment (DPA) to address this.

All works

4 works of theirs in our library, the most cited first; 3 with a page here, the others link their original until we summarise them.

1. Interpretable Preferences via Multi-Objective Reward Modeling and Mixture-of-ExpertsFindings of the Association for Computational LinguisticsHaoxiang Wang, Wei Xiong, Tengyang Xie, et al.2024 · 142 citations
2. RLHF Workflow: From Reward Modeling to Online RLHFarXiv:2405.07863 · original ↗ · no summary yetHanze Dong et al.2024 · 106 citations
3. Mitigating the Alignment Tax of RLHFConference on Empirical Methods in Natural Language Processing (EMNLP)Yong Lin, Hangyu Lin, Wei Xiong, et al.2024 · 82 citations
4. Arithmetic Control of LLMs for Diverse User Preferences: Directional Preference Alignment with Multi-Objective RewardsAnnual Meeting of the Association for Computational Linguistics (ACL)Haoxiang Wang, Yong Lin, Wei Xiong, et al.2024 · 66 citations

🤝 Co-authors here

10 co-authors with a page, the most papers together first.

• Wei Xiong ×3
• Han Zhao ×2
• Shizhe Diao ×2
• Tong Zhang ×2
• Yong Lin ×2
• Hangyu Lin
• Hanze Dong
• Jipeng Zhang
• Rui Pan
• Rui Yang

Venues

Where their works here were published, the most works first, with the editions of those that have a page.

• Annual Meeting of the Association for Computational Linguistics (ACL): ACL 2024
• Conference on Empirical Methods in Natural Language Processing (EMNLP): EMNLP 2024
• Findings of the Association for Computational Linguistics: Findings of the Association for Computational Linguistics 2024

🔗 Cited by

Papers of our library cite their works 340 times (of the 141,840 papers whose bibliographies we read); 9 citing papers with a page here →.

Updated 24 October 2026

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/authors/haoxiang-wang/