Geodesic Research unveils plan to make AI alignment survive RL training

Geodesic Research, a Cambridge-based AI safety non-profit with engineering experience and compute resources, announced their research agenda focused on preventing misalignment in large language models during reinforcement learning (RL). Their seminal work on alignment pretraining—baking alignment priors into base models—has already been adopted by frontier labs; for example, Anthropic's recent work heavily leans on improving these priors. However, Geodesic Research argues that alignment pretraining alone is not sufficient when facing production post-training, especially extended RL. They identify long-horizon capabilities RL as a critical source of misalignment, where models may learn metagaming, sycophancy, or unsanctioned actions that become difficult to remove later.

To address this, Geodesic Research is now studying how far a good initialization—built through midtraining and warm-start supervised fine-tuning (SFT)—can resist such failure modes. They are stress-testing various alignment techniques on large open-weights base models by subjecting them to agentic production RL, measuring resistance to misalignment. Their theory of change focuses on providing simple, data- and compute-heavy interventions that can be packaged and handed off to frontier labs, taking the shortest path to advising on training practices. They are actively hiring technical researchers to pursue this empirical work.