Successful human puzzle players voiced qualitative explanations more often than language agents
Source check: Source checked. Not independent scientific confirmation.
Finding: In an ARC-AGI-3 trajectory study, qualitative descriptions appeared in 72 of 144 successful human runs and 17 of 119 successful runs from three agent systems. The human data came from think-aloud recordings and the model data from agent traces. The comparison measures recorded descriptions in successfully completed games.
Technology: GPT-5.6 Sol assisted analysis; ARC-AGI-3 agent traces
Family GPT-5: profile · models page
Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
- AI contribution
- GPT-5.6 Sol implemented trigger detectors and annotated cognition from aligned transcripts and game traces. Learned agent trajectories supplied matched-task comparison material.
- Test and result
- The authors collected 144 human trajectories across 64 participants and 25 games, collected 250 publicly released model trajectories, and analyzed successful subsets. They spot-checked parts of the model-assisted annotation.
- Limitation / still unknown
- Differences in what humans say aloud and what agents write or log limit claims about underlying cognitive representations.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://arxiv.org/abs/2610.34523v1 · https://arxiv.org/html/2610.34523v1
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review date: 21 Nov 2026. Review scope: v1 annotation workflow, participant selection, Section 4 trajectory results.
No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.
Editorial assessment: Impact 2/5 · Validation 2/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.