Apollo Research shares a reading list on testing AI models

In short

Apollo Research, an independent group that tests AI models, has published a reading list for people who build evaluations. Evaluations, often called "evals," are tests that check what an AI model can do and how it behaves. The list reflects the team's own opinions. It does not describe any incident.

What happened

Apollo Research published the list on 15 August 2024. Its evaluations team says it has "greatly benefited from reading others' papers" and compiled the list of papers it felt were important. It says it likely missed some relevant papers and that the picks are personal opinions.

The list is sorted by topic. Topics include AI agents (programs built on language models that take actions on their own), benchmarks (standard tests for comparing models), the "science of evals," and red teaming. Red teaming means deliberately trying to make a model behave badly in order to find weaknesses. Other sections cover scalable oversight, training methods, auditing and AI governance.

Each paper is marked as "core" or "other." Some entries carry short comments. Apollo says its top picks include a study on dangerous capabilities in frontier models, a paper on predicting how larger models will perform, and a detailed description of how Meta trained its Llama 3.1 models.

Apollo also marks which entries involve its own staff or work. It labels these with notes such as "Disclosure: Apollo paper." One of its own papers on the list looks at whether language models can strategically deceive users when put under pressure.

Some comments are critical. Apollo says one well-known paper on model-written evaluations produced data that was "not very high quality" and had a confusing methods section. Marius Hobbhahn, an Apollo researcher, also adds personal notes. He says he would like to see a more rigorous replication of one paper, and that he disagrees with part of another paper's argument.

The list says it was based on earlier reading lists by Hobbhahn and Jérémy Scheurer. Hobbhahn wrote most of the final draft.

What was done about it

Nothing needed to be done, because the report describes no incident. Apollo says only that the list may be incomplete and reflects the team's views.

Why it matters

Apollo says reading other researchers' work helped it improve its research ideas, its understanding of the field and its concrete evaluations. The list is meant to pass that help on to others who test AI models. The page is a set of recommendations, not a finding about any model's behaviour. More detail is in the original report.