LASR Labs extension phase grant

Amount: $250,000Date Awarded: Apr 17, 2026

Details

Funder: Coefficient Giving
Recipient: Surrogate base model for Mechanistic Interpretability
Purpose: The grant was asked by the whole team working on model organsims (Andrzej Szablewski, Raffaello Fornasiere, Gabriel Konar-Steenberg, Nikita Menon) The aim was to extend the LASR Labs project on developing realistic model organisms. In our project, we create more realistically trained model organisms (MOs) to better study various white-box auditing techniques, such as those used in Anthropic’s auditing game, in Eliciting Secret Knowledge, and, more recently, in AuditBench. Currently we have preliminary evidence that the success rates of these techniques vary by MO training methodology. Our primary goal during the extension is to build a compelling empirical case that wide training methods produce model organisms that are harder to interpret than narrow fine-tuning baselines. Thanks to this grant we succesfully published a paper (https://arxiv.org/abs/2607.01033) that got accepted at ICML Mechanistic Interpretability Workshop 2026
Source: -