U-Define: LLM Planning with Hard and Soft Constraints
A new system called U-Define enables users to guide LLM-generated plans by defining constraints in natural language and categorizing them as either hard rules (must not be violated) or soft preferences (allow flexibility). The research, published on arXiv (2605.02765v1), addresses the challenge of applying rigid constraints to represent intent in end-user task planning. Prior work found that hard-only constraints are too rigid and numeric flexibility weights confuse users. U-Define abstracts strictness into high-level types paired with distinct verification mechanisms, aiming to help users more reliably express and align intent.
Key facts
- U-Define lets users define constraints in natural language.
- Constraints are categorized as hard rules or soft preferences.
- Hard rules must not be violated.
- Soft preferences allow flexibility.
- Prior work found hard-only constraints too rigid.
- Numeric flexibility weights confused users in prior work.
- U-Define uses high-level types paired with verification mechanisms.
- The system aims to improve reliability and control in LLM planning.
Entities
Institutions
- arXiv