A new runtime check aims to stop LLMs from confidently repeating claims a conversation already retracted.
The paper, posted to arXiv, describes a verifier that sits between an LLM and its own chat history: an Interpreter model classifies each conversational turn into one of eight epistemic operations, and a symbolic engine files those into a dependency map showing what every claim rests on and whether that support still holds. Checking whether a new reply is grounded becomes a linear-time walk over that map, no extra LLM call required, and when a premise gets retracted the conflict tracking propagates through the map to flag exactly which downstream conclusions lose their footing. On two third-party benchmarks built around superseded premises, ReviseQA and MemoryAgentBench's fact-consolidation split, the verifier beat a budget-matched retrieval baseline across five QA models and pushed MemoryAgentBench single-hop accuracy from a range of 0.46-0.95 up to 0.93-0.98. With the verifier attached, even a 7B model outperformed an unassisted GPT-4o.
The interesting part is cost, not accuracy. Full conversation context can run to 114,000 tokens; this system keeps prompts near 800 tokens regardless of conversation length, and checks a retraction in under a microsecond at 2,000 turns. That overhead is low enough to run on every single turn of a production agent, which matters because context-manipulation attacks - feeding an agent a plausible continuation built on premises it already abandoned - are a live exploit against deployed systems, not a hypothetical one.
One caveat: these results mostly rely on the benchmarks' own structured updates. When a GPT-4o Interpreter had to extract those updates from raw text itself, accuracy barely moved - which is the part that will actually decide whether this holds up outside a lab.