Can LLMs Improve Their Own Harness? A Controlled Study on ARC-AGI-3
3 judges. Scored by all three judges. Only the day’s top stories get the claim check. How far stories are checked
A study tests whether frontier language models can autonomously build and then revise the software harness that wraps them on ARC-AGI-3, an interactive benchmark of novel games, without human edits.
Score
0.71total under rubric v2.0
- Significance0.66
- Reader interest0.79
- Audience breadth0.58
- Practical relevance0.63
- Substantiated0.88
- Technical depth0.89
- New information0.90
- Counts against: Promotional0.04
- Counts against: Overclaim0.20
Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).
| Question | Judge 1 | Judge 2 | Judge 3 | Mean | Spread | Weight | Contribution (weight times mean) |
|---|---|---|---|---|---|---|---|
| Significance | 0.67 | 0.55 | 0.76 | 0.66 | 0.21 | +0.25 | +0.165 |
| Reader interest | 0.69 | 0.80 | 0.88 | 0.79 | 0.19 | +0.20 | +0.158 |
| Audience breadth | 0.57 | 0.60 | 0.58 | 0.58 | 0.03 | +0.10 | +0.058 |
| Practical relevance | 0.60 | 0.60 | 0.68 | 0.63 | 0.08 | +0.15 | +0.094 |
| Substantiated | 0.98 | 0.80 | 0.86 | 0.88 | 0.18 | +0.10 | +0.088 |
| Technical depth | 0.95 | 0.80 | 0.93 | 0.89 | 0.15 | +0.10 | +0.089 |
| New information | 0.91 | 0.80 | 0.98 | 0.90 | 0.18 | +0.10 | +0.090 |
| Promotional | 0.05 | 0.05 | 0.01 | 0.04 | 0.04 | −0.15 | −0.006 |
| Overclaim | 0.25 | 0.20 | 0.14 | 0.20 | 0.11 | −0.15 | −0.030 |
| Sum of contributions | +0.707 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.707 | ||||||
Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested.
Also covered by
No other coverage was clustered with this story.
Elsewhere
Elsewhere: HN —
Recorded for context only, checked 24 Oct 2026, 08:53 UTC. It played no part in the ranking.