Abstraction and Reasoning on ARC-AGI Public Training Set (Easy) (60 tasks)
0.41Total CostTwo-step agent
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Two-step agentBase model=GPT-3.52025.12 | 0.41 | — | |
| ADAS best agent (reproduced)Base model=GPT-3.52025.12 | 2.11 | — | |
| Two-step agentBase model=GPT-4o2025.12 | 2.85 | — | |
| ENCOMPASS (+ global best-of-N, N = 8)Base model=GPT-3.5, N=8, branchpoints=12025.12 | 3.29 | — | |
| ENCOMPASS (+ global best-of-N, N = 36)Base model=GPT-3.5, N=36, branchpoints=12025.12 | 14.81 | — | |
| ENCOMPASS (+ BFS)Base model=GPT-3.5, branching=8, branchpoints=22025.12 | 15.81 | — | |
| ENCOMPASS (+ global best-of-N, N = 8)Base model=GPT-4o, N=8, branchpoints=12025.12 | 22.76 | — | |
| ADAS best agent (reproduced)Base model=GPT-4o2025.12 | 27.85 | — | |
| ENCOMPASS (+ BFS)Base model=GPT-4o, branching=8, branchpoints=22025.12 | 88.69 | — | |
| ENCOMPASS (+ global best-of-N, N = 36)Base model=GPT-4o, N=36, branchpoints=12025.12 | 95.98 | — |