The application discloses a multi-agent hallucination identification and propagation suppression method based on
declaration-evidence propagation graph and
shared memory access control, and belongs to the field of
artificial intelligence credible reasoning and
generative model security governance. The method first subdivides the multi-agent cooperation output into the smallest verifiable
declaration unit, then collects multi-source evidence for each
declaration unit through retrieval, execution,
perception and questioning of heterogeneous
verification agents, and constructs a declaration-evidence propagation graph containing declaration nodes, evidence nodes, agent nodes, message nodes,
shared memory nodes and time nodes. The support strength, contradiction strength, cross-agent
divergence, propagation
gain, evidence deficiency and historical credit decay are jointly calculated on the propagation graph to obtain the declaration
unit level hallucination risk
score and
diffusion risk
score. According to the double-threshold access strategy and the propagation write-in double-channel blocking rule, the high-risk declaration unit is subjected to freezing, directional review, fragment reservation type
rewriting and
shared memory access control. Finally, the result with evidence
annotation and risk
label is output by the adjudication agent, and the credit value of each agent is updated. The application realizes fine-grained positioning, traceable judgment, shared memory isolation and propagation suppression of the internal cooperation hallucination of the multi-agent, and can be applied to scenarios such as image-text
question answering,
code generation,
software testing, auxiliary decision-making and multi-
modal reasoning.