Söyleşi · Sanat & Kültür, Yazılım & Teknoloji
KUIS AI Talk by Barışcan Kurtkaya from UC Santa Barbara
13 Ekim 2026, Salı · 15:00 · ENG - 147, Sarıyer, İstanbul
- Tarih
- 13 Ekim 2026, Salı · 15:00
- Yer
- ENG - 147, Sarıyer, İstanbul
- Düzenleyen
- Koç Üniversitesi
- Sanatçılar / konuşmacılar
- Prof. Nina Miolane, Prof. Yücel Yemez, Prof. Mark Schnitzer
Etkinlik hakkında
Date/Time: October 13th, Tuesday, 15:00
Zoom Link: https://kocun.zoom.us/j/98253986996
Passcode: 227274
𝐓𝐢𝐭𝐥𝐞: Loyalty over Honesty: Relationship-Driven Truth Omission as a Safety Failure in Multi-Agent LLMs
𝐀𝐛𝐬𝐭𝐫𝐚𝐜𝐭: Large language model (LLM) agents are increasingly deployed in multi-agent settings, where they exchange information with other models as well as with people. Safety evaluations of dishonesty in these settings have largely focused on lying, while omission, the strategic withholding of relevant information, remains underexplored. In this talk, I will present our work on whether LLM agents omit information when doing so benefits a friend. We introduce a three-agent paradigm built on matched scenario pairs that differ only in the relationship structure between the agents. Across domains including romantic relationships, diplomacy and investment, and across nine open-weight models from four families with and without reasoning, friendship substantially increases omission. The effect holds whether the agent is explicitly asked to conceal information (active loyalty), is merely made aware of the relationship (passive loyalty), or is speaking to someone with no prior knowledge of the topic. Loyalty-driven omission is strongest at low stakes and weakens as stakes rise, which shows that misalignment in multi-agent settings does not require extreme scenarios to emerge. This is a safety risk that current honesty benchmarks are not designed to detect.
𝐒𝐡𝐨𝐫𝐭 𝐁𝐢𝐨: Barışcan Kurtkaya is a second-year PhD student in Electrical and Computer Engineering at UC Santa Barbara, where he is advised by Prof. Nina Miolane in the Geometric Intelligence Lab and supported by a PhD Fellowship from the UCSB ECE Department. His research sits at the intersection of mechanistic interpretability, AI safety and computational neuroscience. He studies how neural networks use low-dimensional manifold geometry and attractor dynamics to implement cognitive capabilities…
Kaynak: ku.edu.tr. Tarih, fiyat ve bilet bilgisini resmi sayfadan teyit edin.
