ALL SPEAKERS
SESSIONS
KEYNOTE
Confidently Wrong: Teaching an LLM to Know When It Doesn’t Know
Wed, Nov 4
17:10
–
17:30
Stage 3
AT A GLANCE
Role
AI Engineer
Organisation
PHAROS Labs
Category
Community
ABOUT THE SPEAKER
Rissal Hedna is a Forward Deployed AI Engineer at PHAROS Labs in Hamburg, where he builds agentic AI systems for medical writing and regulated life sciences. He holds an MSc in Intelligent Adaptive Systems from Universität Hamburg and is first author on a paper about LLM calibration at EMNLP 2026.
TALK
Confidently Wrong: Teaching an LLM to Know When It Doesn’t Know
Every LLM in production has the same failure mode, which is delivering a wrong answer in exactly the same confident tone as a right one, and nothing in the output tells you which is which.
Fortunately, Language Models are consistently better at judging an answer than at generating one, and that gap is freely available and usable. This talk walks through SECL, the method behind our EMNLP 2026 paper, which uses a model’s own judgement to retrain its confidence at test time with LoRA updates on the last few layers, and without labelled data at inference.
I will cover how to tell whether your model’s confidence numbers mean anything at all, what moved them and what did not, results across several open models, and where the method breaks down.
MORE VOICES
Other speakers at code.talks
See them live — and 100+ more.
One ticket, every track. Nov 4–5, 2026 at Kinopolis Hamburg.
Get your ticket





