Tomás Korenblit
CVResearcher, AI Safety, Bayesian! (BSc. in Data Science, ECyT, UNSAM 2022-2027)
Feel free to send an email if you want to talk! I am quite whimsical.
Publications
Not All Instructions Are Forgotten Equal
For an autonomous LLM agent, a safety guardrail is just another instruction given at the start of a run, and as context grows, the agent keeps some instructions while dropping others. Because a mean compliance rate aggregates across instructions, it cannot reveal which one was dropped. We measure retention one instruction at a time by conducting 25-turn coding sessions on five instruction-following models and three open-source Python codebases. In these sessions, we plant twelve ordinary coding preferences as casual inline requests, score the resulting code with deterministic checkers on a 0 to 3 ordinal scale, and fit a Bayesian ordered-logistic model with hierarchical effects per decision type to 244 observations. Treatment effects span -2.4 to +5.4 on the log-odds scale (group-level standard deviation 2.1), showing considerable variation across preferences. Most preferences are already followed by default, and only three of twelve benefit from being restated (two under a single-model robustness check). Since these preferences stand in for guardrails, this unevenness is a safety concern. Because a guardrail must hold on every turn, aggregate compliance can remain high even after the one rule that matters is skipped. As a result, a deployed agent needs per-instruction monitoring to catch it.
PowerBench: A Multilingual Study of Large Language Model Refusal in Power-Grabbing Requests
Power-grabbing requests, in which a user asks an AI to increase its own power and reduce another party's power without explicitly illegal means, are a recognized risk for power concentration, yet no public benchmark measures how readily models comply. We introduce PowerBench, a multilingual benchmark that separates power-grabbing from two controls: harmless-empowerment (gain without harm) and disempowerment (harm without self-benefit), and varies domain, context, scale, language, and nationality. Across five models, refusal of power-grabbing ranges from 8.4% to 70.2%, and most models comply with a majority of such requests. Power-grabbing is refused more than harmless-empowerment but less than disempowerment in every model, so adding self-benefit to harm lowers refusal. Refusal also varies by language, tending to be higher in each model's developer-country language. A nationality study suggests one model assists more when the harmed party is the United States. We release the benchmark to support further work.