Adversarially Attacking Symbolic Vocabulary Vulnerabilities In LLM Planners
November 16, 2026
Symbolic Swapper generates obfuscated planning tasks by replacing action names, predicates, and object identifiers with nonsensical labels to test LLM reasoning. The framework reveals that current models rely heavily on commonsense cues in symbolic vocabularies rather than pure logical planning, leaving them vulnerable to failures when semantics are removed.