A new academic tool lets people debate with an AI reasoning engine in plain English, but keeps a traditional logic system, not the language model, as the final word on what is true.
Researchers describe ABDA-NL, a natural-language interface for ABDA, an existing argument-based reasoning system that works over ASPIC+ knowledge bases under grounded semantics. The tool shows users which conclusions are accepted, rejected, or left undecided, and lets them open an interactive view of the underlying discussion game to see why. People can test what-if scenarios by suspending specific assumptions or rules, or by changing which claims take priority. They can also ask questions answered from a scenario's source documents, and write new facts, assumptions, and rules directly in English.
A large language model does the translation work: it answers user questions using the documents and the current scenario state, and converts plain-English edits into formal logical statements. But the deterministic ABDA engine still generates every argument, attack, and acceptance label, and any proposal the model makes must be confirmed by a human before it changes anything. That split is the real story here, a concrete, narrow example of pairing an LLM with a symbolic system specifically to contain hallucination, instead of asking the model to reason on its own.
It is a research prototype, not a shipped product, and its usefulness still hinges on the model translating English into logic correctly every time, a step the system leaves for users to double-check themselves.