Ad­ver­sar­ial Training

Ad­ver­sar­ial train­ing, im­por­tance sam­pling, and anti-ad­ver­sar­ial train­ing for AI whistleblowing

Be­hav­ioral red-team­ing is un­likely to pro­duce clear, strong ev­i­dence that mod­els aren’t scheming

Solv­ing ad­ver­sar­ial at­tacks in com­puter vi­sion as a baby ver­sion of gen­eral AI alignment

High-stakes al­ign­ment via ad­ver­sar­ial train­ing [Red­wood Re­search re­port]

Can Gen­er­al­ized Ad­ver­sar­ial Test­ing En­able More Ri­gor­ous LLM Safety Evals?

AI Safety 101 - Chap­ter 5.2 - Un­re­stricted Ad­ver­sar­ial Training

Deep For­get­ting & Un­learn­ing for Safely-Scoped LLMs

AXRP Epi­sode 17 - Train­ing for Very High Reli­a­bil­ity with Daniel Ziegler

Take­aways from our ro­bust in­jury clas­sifier pro­ject [Red­wood Re­search]

Iron­ing Out the Squiggles

Ad­ver­sar­ial Ro­bust­ness Could Help Prevent Catas­trophic Misuse

Some thoughts on why ad­ver­sar­ial train­ing might be useful

Over­sight Leagues: The Train­ing Game as a Feature

EIS IX: In­ter­pretabil­ity and Adversaries

La­tent Ad­ver­sar­ial Train­ing (LAT) Im­proves the Rep­re­sen­ta­tion of Refusal

Con­tin­u­ous Ad­ver­sar­ial Qual­ity As­surance: Ex­tend­ing RLHF and Con­sti­tu­tional AI

Can We Trust the Judge? A novel method of Model­ling Hu­man Bias and Sys­tem­atic Er­ror in De­bate-Based Scal­able Oversight

En­hanc­ing Ge­nomic Foun­da­tion Model Ro­bust­ness through Iter­a­tive Black-Box Ad­ver­sar­ial Training

Does ro­bust­ness im­prove with scale?

EIS XII: Sum­mary

Loss Curves

Beyond the Board: Ex­plor­ing AI Ro­bust­ness Through Go

A Solu­tion to Sand­bag­ging and other Self-Prov­able Misal­ign­ment: Con­sti­tu­tional AI Detectives

LLM Sy­co­phancy: groom­ing, proto-sen­tience, or both?

EIS XI: Mov­ing Forward

In­trigu­ing Prop­er­ties of gpt-oss Jailbreaks

La­tent Ad­ver­sar­ial Training