Runtime detection and control of LLM coherence failures (looping, hallucination, context loss). No fine-tuning. Zero iatrogenic harm. 69 experiments across 5 architectures.
reproducible-research transformers pytorch open-science pythia checkpoint-restart ai-safety runtime-monitoring control-barrier-functions mechanistic-interpretability llm-safety inference-time-intervention coherence-control
-
Updated
Mar 30, 2026 - Python