What our cohorts have produced.

Publications

• Inference-Time Toxicity Mitigation in Protein Language Models via Logit-Diff Amplification
  Manuel Fernández Burda, Santiago Aranguri, Ivan Arcuschin, Enzo Ferrante. Generative and Experimental Perspectives for Biomolecular Design Workshop at ICLR 2026.
• Benchmarking AI Control Protocols for Safety in Medical Question-Answering Tasks
  Guido Freire, Agustín Martínez-Suñé, Viviana Cotik. Principled Design for Trustworthy AI Workshop at ICLR 2026.
• White-Box Monitoring for Personality Mirroring in Conversational AI
  Eitan Sprejer, Agustin E. Martinez-Sune, Bruno Bianchi. Catch, Adapt, and Operate: Monitoring ML Models Under Drift Workshop at ICLR 2026.
• What Large Language Models Know About Plant Molecular Biology
  Manuel Fernández Burda et al. LatinX in AI (LXAI) Research Workshop at NeurIPS 2025. Acknowledges AISAR support; also submitted to Nature Plants.
• Is Gemini 3 Scheming in the Wild?
  Alejandro Wainstock, Agustín Martínez-Suñé, Ivan Arcuschin, Victor Braberman. LessWrong.

Scholar outcomes

• Two scholars were admitted to MARS (Cambridge).
• A scholar was admitted to LASR Labs.
• A scholar was admitted to ARENA.
• Scholars took part in ML4Good Brazil, one as a participant and one as a facilitator.
• Scholars took part in BlueDot's AI Strategy course, one as a participant and one as a facilitator.
• A scholar gave a talk at a MILA (Quebec) reading group.
• A scholar received a career-transition grant from Coefficient Giving.
• Manuel Fernández Burda, one of our scholars, was co-first author of Kaleidoscope: In-language Exams for Massively Multilingual Vision Evaluation (ICLR 2026), separate from his AISAR project.
• A scholar's AISAR project became their undergraduate thesis.

Mentor outcomes

• Four of the six mentors would not have worked on AI Safety in 2025 without AISAR, and all six plan to continue.
• Two CONICET doctoral fellowships were secured for external students on AI Safety topics: one on confidence quantification for safer AI models, and one on formal foundations for reliable LLM-based software.