Last week I presented two papers at the SKILLED-LLMs workshop in Lisbon, part of FLoC 2026, the 9th Federated Logic Conference.
Both start from the same observation. Ask a large language model a hard question and you will always get an answer. Where the underlying knowledge is genuinely contested, that fluency does more harm than good.
When AI Should Refuse to Decide, with Rohan Nanda and Júlia Schütz Veiga, makes the case for formal irresolution: the capacity of a system to say that the applicable law does not settle the question.
We ground it in marine biodiversity law, where the BBNJ Agreement, the CBD, and the Nagoya Protocol overlap and sometimes point in different directions. Working from a Hohfeld-structured knowledge base of 377 treaty provisions, we found that LLM retrieval reaches precision of 0.77 but recall of only 0.53, systematically missing the dispersed benefit-sharing obligations that matter most to the actors with the least capacity to check an answer for themselves. The indeterminacy here belongs to the regime itself, so no amount of extra training data will close it. Collapsing it into one confident reading builds a form of hermeneutical injustice straight into the architecture.
Neurosymbolic Clinical Reasoning, with Fivos Tzavellos and Rohan Nanda, takes the same intuition into medicine.
The LLM extracts, the symbolic layer reasons. Toulmin’s model organises the evidence, epistemic commitments identify the beliefs that bear on the decision, and reason-based reasoning weighs them through the hierarchy of evidence and the characteristics of the patient. A semaglutide case study shows three patient profiles yielding prescribe, undecided, and withhold from an identical evidence base, with every step traceable back to source. The undecided outcome is where the framework earns its keep.
Henrique Marcos 