Research

Publications and working notes.

2026
publication
Neuro-Symbolic Injection of LTLf Constraints in Autoregressive Reinforcement Learning Policies
Accepted at the SKILLED-LLMs 2026 workshop · A. Ansarifard, E. Umili, M. Mancanelli, F. Patrizi

A differentiable logic-loss approach for injecting LTLf (Linear Temporal Logic on finite traces) constraints directly into autoregressive sequence-model policies (Decision Transformer, Trajectory Transformer), evaluated on the ColourBomb safety benchmark.