Readable behavioral signals remain in frozen LLM hidden states, Cygnus boosts accuracy
冻结大语言模型隐藏状态中仍存可读行为信号,新技术大幅提升准确性
Proprioceptive AI says Cygnus adds adapters to frozen LLMs and raises Qwen-32B on ARC-Challenge from 82.2% to 94.97%. It projects hidden states into a gl(4,R) Lie-algebra space to isolate “dark modes.” Watch replication; the post does not disclose full eval sets or controls.
Why it matters: Proprioceptive AI attaches a self-awareness adapter to a frozen model, reading behavior signals from hidden states without retraining, lifting Qwen-32B's ARC score from 82% to nearly 95%.