Novel problem synthesis for training problem solving and inventiveness: design choices for simulated worlds
Year: 2026
Type: preprint
Abstract: Training on texts teaches an AI to imitate human achievement, not to reach new heights, and in our world humans have already discovered the most essential things. This paper explains how to put current models, with all their earthly priors, into HALs: simulated worlds, often with alien laws, where those priors can help them latch on but do not hand them the answers, because changed laws give each world problems of its own. Such problems, which no human has ever needed to solve, are a fresh resource that might sharpen the problem-solving ability models already have. The training is meant to sharpen the pipeline from knowledge to application: from a law or property of the world, to a mechanism, to an invention that does a job. One line runs through the paper: a pre-programmed outcome, such as a crafting recipe, was written by the designer, while an invention uses the rules of the world to extend the space of ideas explored. The paper sets out the design choices: what a world is built from, the difficulty map and its layers of subversion, depth and breadth as dials, fields inside a world, purpose as scoring, what the agent is given to read, and the loop of invention. The hypothesis is that HAL training teaches a general way of inventing and problem solving, with weight on in-world problems; the argument is that, since every HAL is built to have laws that can be put to use, understanding and applying may carry over to other HALs and then to ours. It is untested: no model has been trained. Two further uses need less than the full hypothesis. HALs as a family of benchmarks for problems that may lie beyond what people can solve need no transfer at all. Realistic HALs, whose by-products can be knowledge that applies to our own world, need what is learned to carry over from one HAL to another, but not from HALs to our world. The first agent runs in small worlds are demonstrations of the design, not tests of the hypothesis, and are reported in Section 10 and Appendices B, C and E. This paper follows "Training AI to invent in alien simulated worlds" (Version 3, doi:10.5281/zenodo.23734411), which proposed HALs. AI models were used as writing and research assistants; see the statement at the end of the paper.
Author: Mortensen, Johannes Dvorak
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7234338497
