Modular Pretraining Enables Access Control
Introduces GRAM, which isolates dangerous knowledge in switchable model modules so one model can approximate versions with categories removed, tested from 50M to 5B parameters.
Research / Sources
Anthropic
Research on reliable, interpretable and steerable AI. anthropic.com
Focus All
Topic Unlearning
Date Any time
An off switch for dual-use knowledge in AI models
Anthropic and AE Studio present GRAM, a method that confines dual-use CBRN and cyber knowledge in a frontier model to a removable part, and test how well it works.
Beyond Data Filtering: Knowledge Localization for Capability Removal in LLMs
Proposes Selective GradienT Masking, which confines learning from dangerous (CBRN) data to removable parameters, and finds it beats data filtering and resists recovery of removed knowledge.
Believe It or Not: How Deeply do LLMs Believe Implanted Facts?
Introduces a belief-depth framework showing synthetic document finetuning often implants genuine beliefs in LLMs, while prompting and mechanistic editing do not, for safety-relevant knowledge editing.
Enhancing Model Safety through Pretraining Data Filtering
Filters CBRN weapons information from pretraining data, then pretrains models from scratch, cutting accuracy on a harmful-capabilities eval by 33% relative to baseline while preserving benign capabilities.
Modifying LLM Beliefs with Synthetic Document Finetuning
Shows synthetic document finetuning can insert beliefs into LLMs, and demonstrates applications to honeypotting for misalignment detection and unlearning of hazardous knowledge.