DIO-Agent: LLM-Driven Discovery for IO2Code Program Synthesis
Researchers propose DIO-Agent, a discovery agent for synthesizing programs from input-output behavior (IO2Code), a challenge distinct from natural language to code (NL2Code). While NL2Code leverages semantic alignment from pretraining, IO2Code requires recovering principles from computational behavior. DIO-Agent frames IO2Code as evolutionary search over discrete program space, using an LLM as mutation operator guided by execution error signals. It prevents wandering into incorrect dead ends. The paper is available on arXiv.
Key facts
- arXiv:2605.15334v1
- DIO-Agent is a discovery agent for IO2Code
- IO2Code synthesizes programs from input-output behavior
- NL2Code exploits semantic alignment between natural language and code
- IO2Code requires recovering underlying principles from concrete computational behavior
- DIO-Agent uses an LLM as mutation operator
- Concrete error signals from execution guide each mutation
- The search is over discrete program space
Entities
Institutions
- arXiv