Can LLMs Improve Their Own Harness? A Controlled Study on ARC-AGI-3

Paper Overview

Field: Machine Learning Authors: Jens Lehmann, Sahar Vahdati Published: 2026-10-24 arXiv: 2610.30499

Summary

A hallmark of generally intelligent systems is acquiring new skills efficiently without humans redesigning the 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. Human-steered systems already report ceiling scores on the public games, showing that effective designs exist. The question this paper asks: can frontier models build and improve such a harness on their own?

Method

• A fixed coding bootstrap has each model construct a complete task harness.
• The model 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.
• Task performance is measured with Relative Human Action Efficiency (RHAE), combining level completion with action efficiency relative to human baselines; dollar-denominated inference cost is recorded as a practical resource measure.

Findings

• 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-set gains overstate held-out improvement.
• Reductions in inference cost can come at the expense of task performance.
• Human-steered systems demonstrate further headroom.

Conclusion

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.

Original 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.

--- *Auto-collected on 2026-10-24*