BAISHResearch
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 research group rooted in Buenos Aires, where researchers can work on 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:
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.
From your first course to published research.
01 · Learn
Build foundations through BlueDot Impact’s Technical AI Safety Course, hosted in person by BAISH.
Learn more02 · Practice
Project-based technical course for making concrete contributions to AI safety research or engineering.
Learn more03 · Research
Do a first mentored research project through the AISAR fellowship, or at the Apart Research hackathons we host.
Learn more04 · Launch
Join BAISH Labs, our research lab in Buenos Aires, or move into a full-time AI safety role.
Learn moreWork by researchers connected to BAISH, newest first
Aa Researcher connected to BAISH
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.
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.
White-box monitoring of conversational models that drift toward mirroring the user's personality, with open code and dataset.
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.
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.
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.
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.
AI-powered framework that accelerates AI governance scenario exploration through autonomous agent tabletop exercises, compressing preparation cycles from years to minutes.
Master's thesis exploring AI safety through debate mechanisms, studying asymmetric capabilities and weak judges in the MNIST environment. Features an interactive demo.
Investigating how well chain-of-thought reasoning can be monitored for safety through faithfulness and verbosity metrics.
Studying how AI systems' internal beliefs affect their persuasiveness in debate scenarios — implications for AI safety and deception.
A multi-judge approach to better approximate human preferences in AI systems, improving alignment evaluation.
A pipeline addressing CBRN evaluation vulnerabilities through consensus detection, verified cloze scoring, and statistical evaluation with bootstrap confidence intervals.
Stress-tested A Narrow Path Phase 0 ASI moratorium, identifying four circumvention routes and proposing ten mutually reinforcing policy amendments.
Get Involved
Whether you want to do research with us or support the lab's work.
Want to contribute to AI safety research? Let us know your background and interests, and we'll connect you with relevant projects and collaborators.
After you've written to us, you can also book a call with Eitan.
