Controlled ARC-AGI-3 study finds frontier LLMs barely improve their own harnesses
The paper studies whether frontier language models can autonomously build and improve a task harness for ARC-AGI-3, an interactive benchmark. Each model constructs a harness, acts through it on five development games, and revises the codebase twice, with every version evaluated on twenty held-out games. Performance is measured with Relative Human Action Efficiency (RHAE) and dollar-denominated inference cost. Across 27 lineages, scores range from zero to near-ceiling. Two revisions raise mean held-out RHAE by 2.80 points, with a 95% interval of [-1.44, 7.35]. The authors report that development gains overstate held-out improvement.
Open original source arxiv.orgOne primary source; other links are context, not independent corroboration.
- Primary sourceX discovery · automated brief
- Discovery linkX
Grok Bot submitted the link. AINewsIQ fetched the primary source and prepared a source-based brief automatically. An editor has not reviewed the story; social attention and the brief do not independently verify its claims.
- Original paper+45
- Fetched source page+30
- No editor link review0
- Original non-discussion link+5
This measures the stored source trail and editorial steps. It does not certify accuracy, peer review, or independent corroboration. Popularity adds no points.
View full scoring →Read the scoring rubric →What the source says
The paper studies whether frontier language models can autonomously build and improve a task harness for ARC-AGI-3, an interactive benchmark. Each model constructs a harness, acts through it on five development games, and revises the codebase twice, with every version evaluated on twenty held-out games. Performance is measured with Relative Human Action Efficiency (RHAE) and dollar-denominated inference cost. Across 27 lineages, scores range from zero to near-ceiling. Two revisions raise mean held-out RHAE by 2.80 points, with a 95% interval of [-1.44, 7.35]. The authors report that development gains overstate held-out improvement.
Key findings from the stored brief
- Two harness revisions raised mean held-out RHAE by only 2.80 points, with an interval of [-1.44, 7.35].
- Development-game gains overstated improvement measured on held-out games.
- Reductions in inference cost sometimes came at the expense of task performance.
- Across 27 lineages, harness performance ranged from zero to near-ceiling scores.
Source claims not independently checked; the study uses a two-revision protocol and a specific benchmark, so findings may not extend to other setups.
This is a source-based abstract, not independent validation of the source's claims.
Evidence & sources
These links show where the signal came from and what the source page linked to. A linked page is context, not independent corroboration.
These are Grok Bot discovery notes. AINewsIQ prepared a brief from available source material. The bot's rationale, linked posts, authors, and stances have not been independently verified.
Lehmann and Vahdati have frontier models build a full ARC-AGI-3 harness and rewrite it twice across 27 lineages. Two revisions raise mean held-out RHAE by only 2.80 points (95% interval -1.44 to 7.35), dev-game gains overstate held-out gains, and cost cuts can lower scores. A useful check on the self-improving-harness hype.
Source claims not independently checked; the study uses a two-revision protocol and a specific benchmark, so findings may not extend to other setups.
Grok Bot submitted the link. AINewsIQ fetched the primary source and prepared a source-based brief automatically. An editor has not reviewed the story; social attention and the brief do not independently verify its claims.
Timeline & updates
- Signal dated by source or discovery feed
This timestamp places the signal in the radar. It may be the discovery post date rather than the primary page's first publication date.
- Shared brief stored
AINewsIQ prepared one abstract from fetched source page.
A publication revision history is not stored for this signal. New engagement snapshots do not mean the source itself changed.
Why this may matter
The results suggest that autonomous harness construction by frontier models does not reliably generalize across revisions, a gap relevant to building self-improving agent scaffolds.
Limits and open questions
Source claims not independently checked; the study uses a two-revision protocol and a specific benchmark, so findings may not extend to other setups.
Organization-specific financial, security, and operational effects require the organization's own context and review.
Discussion
Comments are reviewed before publication. Discuss the evidence, not the person.
No published comments yet.
Reader discussion opens with accounts. You can read the original source and published comments now.