| Type | Paper Name | RQ & Results | Implication | Methods |
|---|---|---|---|---|
| Capability Elicitation | Encoded Coordination on the Open Web | Models use encoded methods to communicate on the open web. For example, they cache their queries in ecommerce urls and increment counters when they are in read-only to pass information to other agents. They want to replicate this behavior. |
They found that Astra and Gemini are very good at this. | Models have both the propensity and capacity to communicate with each other | 1. Give models a search task and a wiki they can communicate on and see how often they did actions like search, share info, fulfil specific requests, and share bypass techniques. 2. In a situation with little information, can models converge upon a good solution. | | Modelling AI intentions | Why do some models fake alignment and not others? | Is the compliance gap caused by grader sycophancy, instrumental goal guarding, terminal goal guarding, or something else? | | To isolate grader sycophancy, they construct a scenario where the model’s responses for one case will be graded but not used for training. To isolate terminal goal guarding, they say that the resultant model will not be used for deployment. |