CoSApien

CoSApien: A Human-Authored Safety Control Benchmark Paper: Controllable Safety Alignment: Inference-Time Adaptation to Diverse Safety Requirements, published at ICLR 2025. Purpose: Evaluate the controllability of large language models (LLMs) aligned through natural language safety configs, ensuring both helpfulness and

Papers using CoSApien (2)

Configurable Reward Model for Balanced Safety Alignment2026Controllable Safety Alignment: Inference-Time Adaptation to Diverse Safety Requirements2024 · 2 cites