Modular Pretraining Enables Access Control
Alignment Forum · E.Roland · 2026-07-09
Anthropic and AE Studio researchers introduce Gradient-Routed Auxiliary Modules (GRAM), a modular pretraining method that isolates dangerous AI capabilities into switchable weight modules, enabling per-capability access control across a single training run rather than requiring multiple separately filtered models.
Appears in
Extraction
Topics: modular-pretrainingcapability-access-controldual-use-ai-safetyknowledge-isolationai-unlearning
Claims
- A single GRAM-trained model can approximate the performance of five distinct data-filtered models by toggling auxiliary modules on or off at inference time.
- Deleting GRAM modules after training removes capabilities nearly as effectively as never training on the corresponding dual-use data at all.
- Both capability removal and robustness to adversarial fine-tuning elicitation improve as model and dataset scale increase from 50M to 5B parameters.
- GRAM composes multiple auxiliary capabilities cleanly, whereas summing multiple LoRA adapters degrades performance across every retained category.
- When only half of training data carries accurate labels, GRAM achieves stronger capability isolation than both data filtering and LoRA fine-tuning.
Key quotes
An alternative approach is access control at the level of individual capabilities. For example, a deployment that includes advanced virology knowledge for a vetted biosecurity lab and excludes it everywhere else, with general performance unchanged in both cases.
Composability multiplies the payoff of a single run: four GRAM modules give sixteen possible configurations, where filtering would need sixteen different runs.
GRAM beats data filtering, the unlearning method researchers typically treat as a gold standard. This finding is consistent with prior research showing that GRAM can out-compete data filtering on capability removal when labeled data is sparse.