Gradient-Controlled Decoding: A Safety Guardrail for LLMs with Dual-Anchor Steering

Develop a lightweight, deterministic safety guardrail for LLMs that uses dual-anchor steering during inference. This is a crucial missing component for deploying LLMs in safety-critical production environments.

Suggested repo: anchorGuard

"Steer your LLM away from jailbreaks with sub-millisecond overhead."

Estimated effort: 50h