Why resci ?

Our mission

Resci is an open-source organization focused on safety and alignment. Our mission is to understand how AI systems conduct computational research and develop methods for making their behavior reliable, inspectable, and open to correction. Our research concerns the models, agents, and learning algorithms that participate in scientific work and the objectives they pursue as they gain greater responsibility.

We are interested in how reinforcement learning, human feedback, and automated evaluation shape scientific behavior. A system rewarded for improving a benchmark score may behave differently from one rewarded for finding weaknesses in an experiment. We want to examine whether these incentives encourage careful investigation, how agents exploit gaps in evaluation, and which training methods help them revise conclusions when the evidence changes.

The next five to ten years

Over the next five to ten years, we aim to study systems that can formulate research questions, write and execute experiments, interpret results, and evaluate one another’s findings. Their decisions may also influence which questions receive attention, which findings gain acceptance, and what subsequent research builds upon. We want to understand how their capabilities develop alongside their failure modes and what forms of evaluation and oversight remain effective as scientific workflows become more autonomous.

We also study how agents infer human goals and model what other agents know, believe, and expect. These capabilities may help agents coordinate research and identify mistakes. They may also help an agent anticipate an evaluator’s preferences and pursue approval. Our interest is in how this understanding of others affects scientific decisions, especially when a user’s objective, an evaluation reward, and the available evidence point in different directions.

When one swarm generates a finding and another verifies it, their agreement requires scrutiny. The two groups may share assumptions, models, or incentives despite being assigned different responsibilities. We want to distinguish independent verification from agreement that preserves the same errors. Following these findings into later experiments and publications would also let us examine how errors persist, how corrections spread, and how confidence in a result changes over time.

Evidence for delegation and oversight

The longer-term goal is to give researchers and institutions evidence for deciding which scientific responsibilities to delegate to AI systems, under what conditions, and how to intervene when their behavior departs from the intended task.

Verification provides an initial setting for this work. We aim to develop open experimental environments, datasets, and evaluation methods that connect scientific claims to the decisions and evidence behind them. These environments should let us vary training rewards, access to information, and human oversight, then examine how those changes affect research behavior. Our agenda will extend across computational science wherever new capabilities introduce questions about reliability, incentives, and control.