Sampura Research is an AI safety research non-profit focused on scalable oversight for frontier models through improved human/AI “judges”.
Judges (LLM, human, or a hybrid) are essential for aligning highly capable models through their use in evaluation and training. Limitations in today's judges can lead to undesirable behavior such as reward hacking in training, or failure to spot misalignment in evaluation.
Read more
As models become more capable and their data more complex, often beyond what the systems overseeing them can understand, we expect these limitations to grow in severity and to become harder to notice.
We are designing novel judge protocols which leverage the strengths of both humans and models to outperform either alone, and building high-quality evaluations to explore judge performance across a diverse set of safety-related tasks. We believe that our focused approach will rapidly advance the science of judges by providing a much-needed systematic measurement of which protocols actually work, and producing state-of-the-art judges which can be used in real world training and evaluation pipelines.
Backed by an $11M grant from Coefficient Giving ($7M for the first year, and another $4M pledged), Sampura Research was founded in 2026 by members of DeepMind’s alignment and human data teams.
