MCA: Open-source neuro-symbolic wrapper for 7B LLMs

discuss.huggingface.co3 min read

The author has introduced MCA (Metacognitive Control Architecture), an open-source neuro-symbolic wrapper designed to teach 7-byte language models how to challenge false premises, explain paradoxes, and abstain. Benchmarked on 100 adversarial questions using Kaggle 2xT4 GPUs, models wrapped with MCA dramatically improved their accuracy and false-acceptance rates at the cost of increased latency and higher rates of over-refusal on ordinary facts. Open article →