Scalable Ensembling For Mitigating Reward Overoptimisation
This addresses a critical alignment issue in large language models, offering a more feasible solution for mitigating overoptimisation in resource-intensive settings.
The paper tackles the problem of reward overoptimisation in Reinforcement Learning from Human Feedback (RLHF) for language models, where policies overfit to proxy reward models, by proposing a scalable ensembling method using a shared encoder with separate linear heads, achieving similar performance to full ensembles with significant reductions in memory and time requirements.
Reinforcement Learning from Human Feedback (RLHF) has enabled significant advancements within language modeling for powerful, instruction-following models. However, the alignment of these models remains a pressing challenge as the policy tends to overfit the learned ``proxy" reward model past an inflection point of utility as measured by a ``gold" reward model that is more performant -- a phenomenon known as overoptimisation. Prior work has mitigated this issue by computing a pessimistic statistic over an ensemble of reward models, which is common in Offline Reinforcement Learning but incredibly costly for language models with high memory requirements, making such approaches infeasible for sufficiently large models. To this end, we propose using a shared encoder but separate linear heads. We find this leads to similar performance as the full ensemble while allowing tremendous savings in memory and time required for training for models of similar size.