Yusuf Kesmen

CS'25 Valedictorian @bilkentuniv | PhD Student @epfl_en IC | Co-fndr @bilkentai | Ex @roketsan @Huawei @aselsan | LLMs, Reasoning

[Yusuf Kesmen @yusuffkesmen]

Grateful to my advisor @anniehartley_ and to @aroraakhilcs and Lars Klein for their supervision, and to all my co-authors 🙏

📄 https://t.co/OyAz12gqgz
See you at NeurIPS!

[Yusuf Kesmen @yusuffkesmen]

🎉 MoBayes is accepted at #NeurIPS2026!

Should an LLM doctor do probabilistic reasoning in its head? We argue it shouldn't: separate reasoning from language.

A cheap LLM + an explicit Bayesian module beats frontier standalone LLM doctors, at a fraction of the cost 🧵

[Yusuf Kesmen @yusuffkesmen]

Why this matters clinically:
• an explicit, auditable posterior at every turn
• a tunable abstention threshold instead of overconfident guesses
• swap the knowledge base for a new population, no retraining
It also holds up under adversarial patient communication styles.

[Tunga Bayrak @TungaBayrak]

Today we’re releasing Adam and Eve, the most human-like AI voices ever built.

Adam ranks #1 among AI models on @DesignArena’s AudioRealismBench.

We believe we’ve crossed the uncanny valley. API access on our Website!

Listen: https://t.co/gHXurz0gSc

Leaderboard: https://t.co/cQmbwyBqrB

Special thanks to @alpsencerozturk, @ahmeterdempmk, and the entire Freya team for making this possible.

We’re just getting started. Join us!

[Akhil Arora @aroraakhilcs]

AI agents like @openclaw 🦞 are everywhere, answering emails, managing calendars, doing our chores for us

📣 REALM is back for year 2! Workshop for Research on Agent Language Models at #EMNLP2026, Budapest 🇭🇺

Stellar lineup ⬇️

📅 Submit by July 17, 23:59 AoE
#LLMAgents #NLProc

[Yusuf Kesmen @yusuffkesmen]

18) the meta-lesson: the skill isn't in the prompt. it's in how you structure your codebase, your workflow, and your empathy for a system that starts from zero every time. changed how i think about working with agents.

[Yusuf Kesmen @yusuffkesmen]

1) my notes on @lexfridman #491 w/ @steipete (@openclaw). it was a very good one. peter gives some best practices on agentic coding and i wanted to share them

[Yusuf Kesmen @yusuffkesmen]

17) codex vs opus — peter described codex as "german" (dry, reliable, disappears 20 min, comes back with results) and opus as "too american" (interactive but "you're absolutely right" triggers his allergy).