Skip to main content
Back

Tsinghua Builds NuHF-Claw to Stop AI Copilots From Causing Nuclear Accidents

A four-agent runtime uses ACT-R cognitive twins and Bayesian safety gates to keep LLM copilots from overriding nuclear operators.

Paper, Full Access
  • Tsinghua researchers propose NuHF-Claw, a risk-constrained multi-agent runtime for digital nuclear control rooms.
  • Four agents handle perception (LSTM), procedure mapping (AutoGraph knowledge graph), cognition (ACT-R), and risk (KRAIL).
  • An LLM auto-generates ACT-R Lisp code, eliminating the manual cognitive-modeling bottleneck that hobbled prior HRA tools.
  • A Bayesian Governance Safety Gate blocks autonomous execution when human error probability exceeds thresholds, preserving operator authority.
  • Validated on the HTR-PM600 full-scope simulator with workload predictions matching NASA-TLX scores at R-squared above 0.9.
  • No public code; prototype needs latency optimization and physiological sensors before real control-room deployment.

Digital nuclear control rooms have quietly become a nightmare for the field of human reliability analysis. Panels of blinking lights and physical switches have given way to touchscreen soft controls, dropdown menus, and layered navigation trees. The result is more efficient, but it also invents entirely new ways for an operator to fail, and existing safety methodologies were built for a world of physical knobs.

A team from Tsinghua University's Institute of Nuclear and New Energy Technology has proposed NuHF-Claw, a multi-agent runtime that tries to plug LLM-based copilots into a nuclear main control room without letting them hallucinate their way into an accident. The framework couples a cognitive model of the operator with a probabilistic risk score, and gates any autonomous suggestion through a Bayesian safety layer that can refuse to act.

Why soft controls break old safety models

The shift to digital interfaces has cut overall human error rates from roughly 49% to 41%, while introducing a category of failure researchers call soft control. Operators spend their time navigating screens, toggling modes, and reconstructing plant state from scattered indicators. The errors that emerge are cognitive: mode confusion, information overload, visual search fatigue, and getting lost eleven clicks deep in a navigation tree.

Traditional human reliability analysis (HRA) methods like THERP, ATHEANA, and CREAM were built for analog rooms and run offline. Newer frameworks such as HUNTER improve on this by simulating a virtual operator, but they require analysts to hand-author cognitive models in Lisp and cannot adapt to novel emergency procedures on the fly. Dropping a general-purpose LLM agent into a reactor is a non-starter. The paper is blunt about this, arguing that hallucinated but plausible operational steps are an unacceptable threat and that unconstrained agents erode the operator's ultimate authority.