PIEFACE (Personalized Interactive Environment For Automata Combination Exploration) is a controllable, verifiable sandbox for training and evaluating reasoning agents. It supports multiple agent backends alongside human-in-the-loop play in the same environment. PIEFACE is ideal for debugging policies, collecting human feedback, and running alignment experiments in symbolic reasoning domains.
While our current demo focuses on a specific domain from theoretical computer science, the PIEFACE platform can support any symbolic reasoning task with discrete, maskable actions and a ground-truth verifier.
In complexity theory, gadgets are modular components used in reductions—like logic gates for computational hardness proofs. These gadgets help encode constraints in problems such as Sokoban, PushPush, or block-sliding puzzles.
Common gadgets include:
PIEFACE began as a visual debugger for RL-discovered gadget simulations, allowing researchers to step through traces, inspect gadget states, and interact with known constructions. It has since evolved into a flexible environment where users can swap between agents mid-trace, replay the same scenario with different policies, or take over manually — enabling head-to-head comparisons, interactive debugging, and personalized agent training.
Interested in collaborating or exploring new domains in PIEFACE?
Reach out via LinkedIn or email me at zburton [at] mit [dot] edu.