Ava Iranmanesh

Ava Iranmanesh

I study AI safety and alignment starting from the people these systems are meant to serve. I am most interested in quantitative, qualitative, and philosophical methods to explore questions about the societal impacts, governance, and future of AI and humanity.

I have worked with Dr. Liwei Jiang at the University of Washington on HearSayBench, a benchmark asking whether language models recognize the unstated constraints that make standard advice impossible for people in adversarial conditions, and with Prof. Daryl Cameron at Penn State on empathy and AI risk. My publications can be found here.

Contact: ami5520 at psu.edu

What’s new?

I am a fourth year Computer Science student in the Schreyer Honors College at Penn State, graduating in May 2027, and I am applying to PhD programs for Fall 2027.

HearSayBench accepted as an oral at NeurIPS 2026, one of 15 orals among 3,757 submissions to the Evaluations & Datasets Track.

Our position paper, There are futures that benchmark driven AI cannot see, presented as an oral at ICML 2026.

Smolder: Empathy, Collapse, and What Remains for AI Risk accepted at the Positech workshop, CSCW 2026.

Research

(1) HearSayBench: Can LLMs Navigate from Abstract Human Rights to Lived Lives? [project page]
Ava Iranmanesh*, Sobhan Lotfi*, Ali Iranmanesh*, Liwei Jiang
NeurIPS 2026, Evaluations & Datasets Track; Oral. Also accepted at the Pluralistic Alignment Workshop, ICML 2026

(2) Position: There are futures that benchmark driven AI cannot see [paper, pdf]
Ava Iranmanesh*, Sobhan Lotfi*, Lachin Naghashyar, Ali Shirali, Fateme Nateghi Haredasht, Sanmi Koyejo, Philip Torr, Yong Suk Lee, Fazl Barez, Joel Lehman, Peter Norvig, Arvind Narayanan
ICML 2026; Oral

(3) Smolder: Empathy, Collapse, and What Remains for AI Risk
Ava Iranmanesh, Daryl Cameron
Positech Workshop, CSCW 2026

* equal contribution

Experience

Research Assistant, University of Washington
Winter 2026
Remote
Supervisor: Dr. Liwei Jiang

  • Built HearSay, a benchmark asking whether LLMs can move from abstract human rights to advice that fits a person’s actual life, framed with Sen’s capability approach and Gricean pragmatics.
  • Designed the prompts and a five axis failure taxonomy, and built the human verification platform.

Research Assistant, Penn State University
Fall 2025
State College
Supervisor: Prof. Daryl Cameron

  • Working on a paper on the relation of empathy and compassion with AI existential risks, accepted to the CSCW Positech workshop 2026.

Summer Exchange Student, Hong Kong University of Science and Technology
Summer 2024
Hong Kong
Supervisor: Prof. Fred Lochovsky

Awards

Oral presentation, NeurIPS 2026
September 2026; 15 of 3,757 submissions to the Evaluations & Datasets Track (top 0.4%)

Oral presentation, ICML 2026
May 2026; 168 of 24,371 submissions selected for oral (top 0.7%)

MCREU Fellowship, Penn State
June 2026; 75 of 9,000+ applicants selected (top 0.8%)

NTU Global Connect Fellowship, Singapore
December 2025; 12 of 2,700+ applicants selected (top 0.5%)

بنی‌آدم اعضای یکدیگرندکه در آفرینش ز یک گوهرندچو عضوی به درد آورد روزگاردگر عضوها را نماند قرارتو کز محنت دیگران بی‌غمینشاید که نامت نهند آدمی

Human beings are limbs of one another, made in creation from a single essence. When time brings pain to one limb, the others cannot rest. You who feel nothing for the suffering of others do not deserve to be called human.

Saadi, Golestan, 1258