From the source
Reflection is publishing the safety training methods and evaluation results behind Beam, its first open-weight model, through a research initiative it calls Project Iris.
- The company says Beam sits on the observed Pareto frontier, declining harmful requests more than all other Western and Chinese open-weight models it tested while still complying with harmless ones.
- Reflection evaluated Beam with Iris Safety Evaluations, a benchmark covering chemical, biological, radiological and nuclear weapons, cyber offense, self-harm, violence, fraud and privacy, and shares it so others can find gaps in their own models.
- Because safeguards on open-weight models can be modified or removed after release, a separate Open Safety Framework assesses whether a model's underlying capabilities could raise risk beyond what comparable public models already enable.
- Beam was trained with pretraining data filtering and a dedicated safety-alignment teacher, combined with a capability teacher through multi-teacher on-policy distillation.
- Reflection reports that scaling the safety teacher from 100B to 501B parameters with the recipe fixed cut over-refusal by 48.7% relative, alongside a 3.1% decrease in the harmful-request pass rate.
