Fabien Roger

Toy mod­els of AI con­trol for con­cen­trated catas­tro­phe prevention

A quick in­ves­ti­ga­tion of AI pro-AI bias

Mea­sure­ment tam­per­ing de­tec­tion as a spe­cial case of weak-to-strong generalization

Scal­able Over­sight and Weak-to-Strong Gen­er­al­iza­tion: Com­pat­i­ble ap­proaches to the same problem

AI Con­trol: Im­prov­ing Safety De­spite In­ten­tional Subversion

Au­dit­ing failures vs con­cen­trated failures

Some nega­tive steganog­ra­phy results

Coup probes: Catch­ing catas­tro­phes with probes trained off-policy

Prevent­ing Lan­guage Models from hid­ing their reasoning

Pro­gram­matic back­doors: DNNs can use SGD to run ar­bi­trary state­ful computation

Will early trans­for­ma­tive AIs pri­mar­ily use text? [Man­i­fold ques­tion]

If in­fluence func­tions are not ap­prox­i­mat­ing leave-one-out, how are they sup­posed to help?

Bench­marks for De­tect­ing Mea­sure­ment Tam­per­ing [Red­wood Re­search]

When AI cri­tique works even with mis­al­igned models

Pass­word-locked mod­els: a stress case for ca­pa­bil­ities evaluation

Sim­plified bio-an­chors for up­per bounds on AI timelines

LLMs Some­times Gen­er­ate Purely Nega­tively-Re­in­forced Text

How Do In­duc­tion Heads Ac­tu­ally Work in Trans­form­ers With Finite Ca­pac­ity?

What Dis­cov­er­ing La­tent Knowl­edge Did and Did Not Find

Some ML-Re­lated Math I Now Un­der­stand Better