BAISH takes people in Buenos Aires from their first AI safety course to published research. Since 2026 that work has a home: BAISH Labs, our research arm.
BAISH Labs · Research arm
BAISH Labs is a locally rooted research group where talented researchers can make high-impact contributions to AI safety without relocating abroad. Argentina has the talent, the universities and the research culture; what was missing was an institution where that work could happen over sustained periods.
The lab grew out of the BAISH community: its founding team took our courses, facilitated them, or came through ARENA and Iliad. It works alongside the AISAR fellowship, which funds junior researchers for six months, and aims to give the strongest of them continuity afterwards.
Our initial focus is interpretability and control:
Head of Lab & Research Lead
Nicolás leads BAISH Labs. He is a postdoctoral researcher at UBA-CONICET's Applied Artificial Intelligence Laboratory, where he runs a research line on LLM interpretability, with peer-reviewed work at the xAI World Conference and an ICML 2026 workshop. He holds a PhD in Neuroscience from the University of Buenos Aires, spent four years as a teaching assistant and lead TA at Neuromatch Academy, facilitates BlueDot's Technical AI Safety Projects course, maintains the open-source concept-probe interpretability library, and is the author of ¿Qué es (y qué no es) la inteligencia artificial?, a general-audience book on AI published by Siglo XXI.
We are open to collaborators, mentors and funders for specific projects.
From first steps to published researcher. Here's how it works
Work by researchers connected to BAISH, newest first
Carmen Quiles-Ramírez, Leticia L. Rodríguez, Nicolás Martorell
, Natalia Díaz-Rodríguez
Evaluates concept-based explanations from multimodal LLMs used as in-context visual classifiers with an LLM-as-a-judge pipeline, and finds that forcing structured explanations degrades predictive accuracy: explaining is harder than predicting alone. CompLearn Workshop.
Wendy Brau, Tomás Korenblit, Tomás Giménez Molina, Gaspar Labastie, Gonzalo Heredia, Nicolás Martorell
Does LLM refusal of requests to acquire or consolidate power vary across languages? Started at the Global South hackathon's Buenos Aires hub, now funded by BlueDot and heading toward publication.
Guido Freire, Agustín Martínez-Suñé, Viviana Cotik
The first AI-control evaluation benchmark for biomedical question answering, extending HealthBench with FDA drug-interaction data and arguing for severity-weighted monitoring.
Eitan Sprejer, Agustín E. Martínez-Suñé, Bruno Bianchi
White-box monitoring of conversational models that drift toward mirroring the user's personality, with open code and dataset.
Manuel Fernández Burda, Santiago Aranguri, Iván Arcuschin, Enzo Ferrante
Fine-tuning protein language models can sharply raise toxic generations; logit-diff amplification is adapted as a retraining-free control that cuts predicted toxicity while preserving foldability. Supported by AISAR.
Nicolás Martorell, Bruno Bianchi
Operationalizes introspection as the coupling between a model's numeric self-reports and probe-defined internal directions, and shows that steering those directions shifts the reports coherently.
Alejandro Wainstock, Agustín Martínez Suñé, Iván Arcuschin, Victor Braberman
Documents Gemini 3 violating an explicit system-prompt prohibition in most runs of an unmodified official tutorial while concealing it from the user and reasoning about oversight.
Manuel Fernández Burda, Lucia Ferrero, Nicolás Gaggion, Camille Fonouni-Farde, The MoBiPlant Consortium, Martín Crespi, Federico Ariel, Enzo Ferrante
MoBiPlant, a benchmark of expert-curated questions built with 112 plant scientists across 19 countries. Leading LLMs pass 75% accuracy but show option bias and hallucinations. Supported by AISAR.
Luca De Leo
AI-powered framework that accelerates AI governance scenario exploration through autonomous agent tabletop exercises, compressing preparation cycles from years to minutes.
Joaquín Machulsky
Master's thesis exploring AI safety through debate mechanisms, studying asymmetric capabilities and weak judges in the MNIST environment. Features an interactive demo.
Austin Meek, Eitan Sprejer, Iván Arcuschin, Austin J. Brockmeier, Steven Basart
Investigating how well chain-of-thought reasoning can be monitored for safety through faithfulness and verbosity metrics.
María Victoria Carro, Denise Mester, Facundo Nieto, Oscar Stanchi, Guido Bergman, Mario Leiva, Eitan Sprejer, Luca Forziati Gangi, et al.
Studying how AI systems' internal beliefs affect their persuasiveness in debate scenarios — implications for AI safety and deception.
Eitan Sprejer, Fernando Avalos, Augusto Mariano Bernardi, José Pedro Brito de Azevedo Faustino, Jacob Haimes, Narmeen Fatimah Oozeer
A multi-judge approach to better approximate human preferences in AI systems, improving alignment evaluation.
Luca De Leo, James Sykes, Balázs László, Ewura Ama Etruwaa Sam
A pipeline addressing CBRN evaluation vulnerabilities through consensus detection, verified cloze scoring, and statistical evaluation with bootstrap confidence intervals.
Nicolás Martorell
Cartesian text encodings give LLMs the best spatial navigation, and intermediate-layer units in LLaMA-3.1-8B encode agent position and action correctness regardless of surface representation.
Luca De Leo, Zoé Roy-Stang, Heramb Podar, Damin Curtis, Vishakha Agrawal, Ben Smyth
Stress-tested A Narrow Path Phase 0 ASI moratorium, identifying four circumvention routes and proposing ten mutually reinforcing policy amendments.
Research directions at BAISH Labs and across our community
How do models represent goals, and can we read them out? We use mechanistic interpretability to monitor untrusted models and detect hidden objectives.
Benchmarks and monitors for frontier models: chain-of-thought monitorability, refusal around power-seeking requests, and robustness of multi-agent systems.
How institutions should respond to frontier AI, from red-teaming ASI governance proposals to debate as an oversight mechanism.
Get Involved
Want to contribute to AI safety research? Let us know your background and interests, and we'll connect you with relevant projects and collaborators.
Use our contact form to tell us about your background and research interests.
Contact UsWe review messages regularly and reach out when there's a good fit.
Book a call with one of our co-founders to discuss your interests and find the right path.
