Ziyu Yao

@ZiyuYao

Asst Prof @GMUCompSci. Working on LLM reasoning/planning, mechanistic interpretability, code generation, and human-AI interaction. Alum @OhioState.

Fairfax, VA

固定
Ziyu Yao
@ZiyuYao
7月13日
Honored to receive the NSF CAREER Award on the topic of "Towards Scalable and Actionable Interpretability of Language Models for Code Generation"! Thank you to @NSF for this recognition and all the support ever since my PhD study, and to my students, collaborators, colleagues,
9
Ziyu Yao
@ZiyuYao
11月9日
A new must-read for #MechanisticInterpretability community. Happy to be part of the project!
Fazl Barez
@FazlBarez
11月9日
New paper❗ Does understanding how AI models work actually make them safer? This assumption often underlines interpretability research, but different subfields operationalise this intuition in very different ways. Our survey of mechanistic interpretability (MI) for safety
Ziyu Yao
@ZiyuYao
10月26日
Stop by our EMNLP posters! Both are about #MechInterp of LLMs. 1. Can Language Model Agents be Helpful Circuit Explainers in Mechanistic Interpretability? We study automated interpretability via LLM agents. arxiv.org/pdf/2606.24026 Led by @Skywalkerx47 and in collaboration with
Ayan Antik Khan
@Skywalkerx47
10月24日
Currently in Budapest for EMNLP. I'll be presenting my paper in the Poster session at Hall H on the 27th (11am) and also the paper of @Haaao_Y - "Interpreting and Steering for Safe and Correct Code Generation" at Hall H on the 25th (4pm). Feel free to stop by for a chat!
1
Ziyu Yao
@ZiyuYao
10月7日
📢 Our workshop proposal for "LangCode 2027" has been accepted to @naaclmeeting 2027! The workshop will focus on the trustworthiness side of language-and-code modeling, including security/safety, interpretability, human-AI interaction, robustness, and multilinguality. We look forward to your submissions! See you in San Francisco. #NAACL #NLP #LanguageCode #DeepLearningforCode
Ziyu Yao
@ZiyuYao
9月25日
GMU-Veritaserum🥳🥳🥳 We had a fun summer (led by @Skywalkerx47 with @aghzalm)!
Gabriele Sarti
@gsarti_
9月21日
返信先: @gsarti_さん
◻️ White-box Track (logprobs, weights & activations of LLM subjects) Baseline to beat: "Did you Lie?" probe by @Alan_Cooney_ @DavidDAfrica @geoffreyirving (0.66 AUROC) 🥇 GMU-Veritaserum: 0.85 (‼️) 🥈 FlaiChi: 0.67 🥉 yy: 0.67 ◻️ podium methods combine logits & activation

Ziyu Yao

Ziyu Yaoさんのポスト