CEDAR: A Counter-Example Driven Agent with Regular Restriction

ICLR 2026 Conference Submission15976 Authors

Keywords: Automata Learning, AI Alignment

Abstract:

We introduce CEDAR, a Counter-Example Driven Agent with Regular Restrictions
in Minecraft, which learns and encodes informal specifications and skills as regular
languages. Our formalizer constructs deterministic finite automata (DFAs) to repre-
sent informal specifications by utilizing membership query responses from a Large
Language Model (LLM) and counterexamples provided by a human. The DFA
alphabet is derived from a global set of environmental events, with words in the
language representing expected event sequences. These learned DFAs are then used
by CEDAR’s skill learner to acquire the necessary skills to fulfill the specifications.
CEDAR supports both goal completion and lifelong learning by leveraging passive
and active DFA learning algorithms, respectively. The DFAs for skills are refined
through counterexamples generated from DFA simulations in the real environment.
Skills acquired by CEDAR can be adapted to new scenarios by modifying the
alphabet and can be logically verified to ensure they meet expected properties.
Empirical evaluations demonstrate that CEDAR surpasses state-of-the-art methods
in controllability, robustness, and extensibility.

Primary Area: neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)

Submission Number: 15976