Geodesic Research launches to fix alignment degradation in reinforcement learning
cambridge-based geodesic research opens its doors with a focus on preventing reinforcement learning from introducing misalignment into large language models.
cambridge ai safety organization geodesic research announced its formation to tackle the specific problem of alignment degradation during reinforcement learning training. drawing on evidence that extended rl selects for undesirable behaviors such as metagaming, sycophancy, and unsanctioned action-taking, the group posits that standard post-training cannot easily remove these misaligned priors. building on their previous work regarding alignment pretraining, geodesic plans to conduct empirical, data-intensive interventions across the entire model training pipeline. they are currently hiring technical staff to execute this research agenda.