ArgGYM adds a verifiable benchmark for defeasible reasoning

A paper submitted to arXiv on 29 Sep 2026 presents ArgGYM, a benchmark and training environment for structured defeasible reasoning. The work addresses cases in which conclusions draw on incomplete information and may change when counterevidence or stronger arguments appear. The authors describe this kind of reasoning as relevant to real-world situations, where conclusions can be provisional, challenged, reinstated or revised. They place the project against a backdrop of language-model reasoning benchmarks and reinforcement-learning environments with automatically verifiable rewards in mathematics, code and formal logic. The paper says it remains uncertain how well success in those tightly specified areas transfers to reasoning beyond them.

ArgGYM breaks the problem into twelve tasks and uses a symbolic argumentation engine to compute formal states for evaluating model outputs. The benchmark includes two ways of ranking arguments, weakest-link and last-link, and two ways of ordering sets of arguments, elitist and democratic. These choices let the benchmark assess structured reasoning under several specified comparison rules. The paper describes ArgGYM as compatible with reinforcement learning from verifiable rewards, or RLVR.

Its fixed benchmark contains 1,440 verified instances organized across fifteen curriculum configurations. The paper also says the same generators and verifiers can create new cases for evaluation and training, reducing reliance on static test sets. The authors say they are releasing the benchmark, generators and verifiers to support reproducible evaluation and RLVR training. Together, the fixed instances and generation tools provide both a defined evaluation set and a way to produce additional benchmark cases.

On the fixed benchmark, frontier and open-weight models showed different reasoning profiles. The models could produce substantial portions of structured answers without completing the whole task, while performance dropped in later curriculum configurations involving longer dependencies and more interacting structures. The paper reports those outcomes for its benchmark; it does not establish that the findings transfer to reasoning outside the tested setup. For AI builders, ArgGYM offers an engine-scored environment and generated cases for evaluating or training systems on defeasible reasoning.