Changed on 03/03/2026
- Thursday, February 12, 2026 - 10:00 am (Santiago, Chile time)
- Hybrid format
- The talk will be held in English
- Speaker:
- Yannis Karmim is a postdoctoral researcher at ALMAnaCH team, Inria and Inria Chile.
Abstract
How do large language models (LLMs) acquire, store, and retrieve factual knowledge?
Unlike classical databases that respond to explicit queries, LLMs model probability distributions over token sequences, a fundamentally different mechanism for encoding world knowledge.
In this talk, Yannis Karmim will examine the inner workings of knowledge acquisition during LLM pretraining, centering the discussion on "How Do Language Models Acquire Knowledge During Pretraining?" (Chang et al., NeurIPS 2024). This work offers insights into the conditions under which a model reliably assimilates a fact: how many exposures are needed, which presentation strategies are most effective, and how catastrophic forgetting shapes what is retained over long training runs. Building on these findings, he will also present recent complementary results that help to understand the underlying mechanisms, touching on how relational structure, repetition, and paraphrasing interact during knowledge encoding.
Yannis Karmim
Yannis Karmim is a postdoctoral researcher at ALMAnaCH team, Inria and Inria Chile.
Yannis holds a PhD from the Conservatoire National des Arts et Métiers (CNAM), where his doctoral research focused on machine learning on dynamic graphs.
He is currently a postdoctoral researcher at Inria, working jointly between the ALMAnaCH team, at Inria Paris Centre, and Inria Chile, on the topic of sociocultural biases in large language models.
His postdoctoral work aims both to characterize these biases and to develop more effective mitigation strategies, in particular by exploring how knowledge graphs can serve as structured priors for debiasing and knowledge injection in LLMs.