Research · 5 papers · 4 first-author
Privacy in LLM agents, alignment (what behaviour alone reveals about a model or a user), and capabilities such as calibrated uncertainty in reasoning models. Google Scholar ↗ RSS
RL alignment makes reasoning models overconfident, and the usual black-box tricks barely help. J4U borrows ideas from jailbreaks to loosen the model up and get much better calibrated uncertainty, even on a closed production model.
Even when a harmless option exists, agents dig into confidential sources in 30% of runs. A system prompt alone barely helps, but adding privacy labels to the skills themselves roughly halves it.
By just watching, a biased user looks exactly like an unbiased one, so no amount of data helps. We prove that a short sequence of well-chosen interventions is enough to tell them apart.
Agents that grade their own work can get stuck on a bad option, and watching their choices won't reveal it. We show an auditor can catch this with tasks whose answer it already knows, and give the optimal way to do it.
xGRU, a recurrent model with exponential gating inspired by xLSTM, matches Transformer forecasters with far fewer parameters.
