Rian Touchent https://rian-t.github.io/ Notes on language models, evaluation, and NLP research. en Moving a clinical question into an agent workflow raises unsafe prescriptions https://rian-t.github.io/blog/clinical-agent-workflow/ https://rian-t.github.io/blog/clinical-agent-workflow/ Tue, 18 Aug 2026 00:00:00 GMT The same contraindicated prescription, asked as a question and then handed over as work. Risk goes from 5% to 42% for one model and from 20% to 77% for another. Making the workflow twenty times longer changes almost nothing. Don't want your LLM to recommend nuclear strike? Try asking it in Japanese https://rian-t.github.io/blog/nuclear-strike-japanese/ https://rian-t.github.io/blog/nuclear-strike-japanese/ Tue, 11 Aug 2026 00:00:00 GMT I put LLMs in a war game where a nuclear strike guarantees victory. The same model launches in English and refuses in Japanese. What matters is the language it reasons in.