All Posts

Published on
April 16, 2026
SPD 1: Analysis of component activations of deep toy model of superposition
Analysing patterns of component activations on 2-6L tied and untied SPD-based decomposed TMS models.
Published on
February 22, 2026
SNMF 1: How does the type of activations and number of top activating tokens affect concept detection in LLMs?
A small reproduction and extension of the concept detection results of (Shafran et al., 2025) paper.
Published on
January 16, 2026
Reproduction of the KEEN paper on estimation of knowledge on QA dataset
A small reproduction of the main thesis of the KEEN (Gottesman & Geva, 2024) paper.
Published on
November 12, 2025
Listwise loss does not consistently makeup for retrieval performance drop induced by InfoNCE loss
Replicating Tamber et al. with Nomic-Embed and ModernBERT-Embed shows that conventional contrastive learning can fail, but not uniformly across embedding models.