COCOLogic-V2: Identifying Logical Inconsistencies via Truly Hard-Negatives

arXiv·medium signal

David Steinmann, Antonia Wüst, and Kristian Kersting introduce COCOLogic-V2, which surfaces logical inconsistencies in interpretable models such as concept bottleneck models and program-synthesis methods using 'truly hard-negative' examples that prior benchmarks failed to capture. This matters for evaluation and safety teams who need to stress-test whether a model's stated reasoning is actually self-consistent rather than superficially plausible.