Zhilin Wang

• Led research and training of SOTA reward models, defining evaluation protocols and interface with RLHF 1. Trained Llama-3.1-Nemotron-70B (Highest-ranked Open Model and Top 10 on LM Arena in Oct’24) 2. Trained Nemotron Reward Models that achieve SOTA on RewardBench, RM-Bench and JudgeBench 3. Developed new training methods for RM and RLHF including Binary Flexible Feedback and SteerLM 4. Designed benchmarks e.g. ProfBench: PhD/MBA-level Rubrics across Science and Business domains • Led curation of human-annotated helpfulness modeling dataset (HelpSteer 1/2/3) with 500k+ downloads 1. HelpSteer (original) contains 10k samples with Likert-5 ratings on 5 attributes (e.g. correctness) 2. HelpSteer2 incorporates multiple annotators per sample and pairwise preference on 10k samples 3. HelpSteer3 introduces natural language feedback and edits on 40k samples incl. Code/STEM/Multilingual • Designed and implemented NeMoLLM Service Python Client (400k+ PyPI downloads) • Mentored interns and junior researchers on Dialogue Modeling, Question Answering, Safety and RLHF

- Research assistant at Noah's Ark, UWNLP (Supervised by Prof Noah Smith) - Projects on knowledge-based approaches in personalizing dialogue and event representation in stories. - Improved NLU by 2-13% using constraint decoding, discriminative reranking, and commonsense integration

- Improved how well Alexa can understand what we refer to by over 80% using knowledge graph embeddings and NLP algorithms (Viterbi; Katz’s Backoff) - Building an internal package to facilitate rapid experimentation and iteration

• Interned at New Oriental Artificial Intelligence Research Institute • Researched on the use of active learning techniques to make the labelling of samples for supervised learning more effective for paragraph-level text classification. Code and Findings at https://github.com/Zhilin123/active_learning • Created knowledge map for a secondary school Physics curriculum to link classroom recordings to curricular content

• Helps teachers to make science fun to learn by recommending relevant popular science videos using machine-learning based text analysis (Latent Semantic Indexing & Latent Dirichlet Allocation) • Deployed on a Python Django-based website and accessible through third-party learning resource platform using a SaaS API • UCL Institute of Education Knowledge Lab EDUCATE Graduate (with EU grant) + Y Combinator StartupSchool Graduate + Runner-up, Cambridge University Entrepreneurs £2,000 competition (Social Enterprise) • I led a team of four with my role contributing to both software engineering and business development

• Interned at Computer Laboratory • Low-income students in India are unable to understand word problems in English, hindering their understanding of mathematical concepts. • Developed web-app that translates questions into their mother tongue and finds definitions of difficult words in a just-in-time manner. • Code at https://github.com/Zhilin123/urop/ • Demo at http://zw322.pythonanywhere.com/ • Presented poster at Cambridge Language Sciences Symposium https://github.com/Zhilin123/Publications/blob/master/CLS_2018.pdf

• Chemistry students can take a picture of a question they have trouble with to obtain videos and study notes as well as similar explained questions in order to solve the question • Helped >10,000 students + Winner, Cambridge University Engineering Society Ventures 2018 (Social Impact)+ AngelHack London 2018 Grand Prize • I led a team of five with my role contributing to both software engineering and business development • Uses natural language processing, information retrieval and full-stack web development

Assisted Assoc Prof Joseph Jay Williams from the School of Computing on several education technology products (ASSISTments and MOOClets). Drafted proposals, prepared informational materials for potential collaborators and reviewed research papers.

Researched on high capacity, low cost batteries based on the zinc-air technology. Found that a novel material (graphene aerogel) can improve the efficiency of these batteries, making them suitable for storing renewable energy locally.

LING 570-572 Core NLP CSE 573 Artificial Intelligence LING 575 Dialogue Systems LING 566 Computational Syntax