GRAM Paper: Precisely Removing LLM Dangerous Capabilities via Gradient-Routed Modules
burkov · x · 2026-07-30
A recent study introduces Gradient-Routed Auxiliary Modules (GRAM) to address the safety dilemma where dangerous knowledge (e.g., bioweapons) and beneficial knowledge (e.g., vaccine research) coexist in LLMs.
Core Mechanism
Researchers add small auxiliary modules to Transformers whose parameters receive backpropagation updates only for selected data types. This encourages specialized knowledge to settle in a removable module rather than spreading through the network. Turning off the module at inference time reduces the corresponding capability while leaving general performance largely intact.
Experiments
Experiments spanned virology, cybersecurity, nuclear physics, and specialized code, testing models from 50M to 5B parameters. GRAM outperformed traditional data filtering and model unlearning methods in precision.
More from Safety
- China drafts export controls on 23 AI technologies; already open-sourced tech exempted — Stannis_Loyalist · 2026-10-16
- Anthropic grants METR inside access to audit its AI models and safety practices — The Decoder · 2026-10-16
- Agent actions need checks outside the model: NVIDIA demos lean on guardrails amid wave of real agent incidents — Creamy-And-Crowded · 2026-10-16
- Vitalik builds "paranoid mode" agent wrapper with Tor, zkAPI and local prompt rewriting — DavideCrapis · 2026-10-16
- Security team demos Signal 1-click exploit technique at Hexacon Paris — moyix · 2026-10-16
- China moves to impose registration and licensing on frontier AI model weights and chip designs — deanwball · 2026-10-16