Grounded Continuation: A Linear-Time Runtime Verifier for LLM Conversations

The researchers propose a runtime verifier called Grounded Continuation, used to identify continuations in long LLM conversations that are built on retracted premises, a weakness that context manipulation attacks exploit. In the method, an LLM Interpreter maps each sentence to one of eight cognitive operations, and a symbolic engine then applies these operations to a dependency graph that records the basis of each claim and whether it still holds; determining whether a continuation is grounded requires only traversing this dependency graph, with time linear in the graph size and without calling an LLM. Retraction operations propagate along the same dependency graph and mark conclusions that lose support. Experiments on five QA models show that the verifier outperforms retrieval baselines on ReviseQA, which examines belief revision, and on the FactConsolidation split of MemoryAgentBench (MemAB-FC), raising MemAB-FC single-hop accuracy from 0.46 to 0.95 up to 0.94 to 0.98, and enabling a 7B small model with the verifier to surpass GPT-4o without it.