Language models struggle to improve their own support programs automatically
Can LLMs Improve Their Own Harness? A Controlled Study on ARC-AGI-3
Artificial IntelligenceSoftware Engineering
Summary
Some smart computer programs, called language models, rely on helper software known as a harness to work better. The authors tested if these models could build and improve this harness by themselves while playing new games. They found that although the models can perform well initially, making changes to the harness did not consistently improve or maintain their performance on new tasks. This suggests that even advanced models need human help to reliably update the programs that guide their actions.
What this means in practice
- •For game ai developers: Design more reliable AI frameworks by understanding the limits of self-improving control programs in open-ended tasks.
- •For software engineers in automation: Recognize that automatic autonomous software adaptation remains challenging, motivating hybrid approaches with human oversight.
Authors
Jens Lehmann, Sahar Vahdati
Abstract
A generally intelligent system acquires new skills efficiently without requiring humans to redesign its software scaffold for every new task. Yet language models rarely act alone. A harness is a symbolic program that orchestrates calls to the model, handling perception, memory, tool use, recovery, and inference allocation on the model's behalf. ARC-AGI-3 is an interactive benchmark in which agents must explore unfamiliar environments, figure out their goals on the fly, and keep learning as they go. On it, human-steered systems already report ceiling scores across the public games, so we know effective designs exist. We ask whether frontier models can build and improve such a harness on their own. Through a fixed coding bootstrap, each model constructs a complete task harness, acts through that harness on five development games, and revises the entire codebase twice. Every version is evaluated on twenty games that never influence adaptation. We measure task performance with Relative Human Action Efficiency (RHAE), which combines level completion with action efficiency relative to human baselines, and record dollar-denominated inference cost as a practical resource measure. Across 27 lineages, harness performance ranges from zero to near-ceiling scores. Two revisions raise mean held-out RHAE by only 2.80 points (95% interval [-1.44,7.35]). Development gains overstate held-out improvement, and reductions in inference cost can come at the expense of task performance. Human-steered systems demonstrate further headroom. These findings reveal a gap between autonomous harness construction and reliable adaptation under this two-revision protocol: even strong task performance does not ensure that subsequent revisions generalize or preserve existing capabilities.