Unverified50% confidenceBenchmarkExact time
On the ARC-AGI benchmark, the same model can jump in score from 30% to 95% or even 100% solely through different Harness designs.
1
Sources
50%
Confidence
Long-term
Relevance
10/8/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedScaling the model size by approximately one hundred thousand times improved the base LLM's score on ARC from only 0% to approximately 10%.69% similarUnverifiedOn the ARC-AGI-3 benchmark, GPT-5.6 Sol scored 8%, while most other models scored below 2%; Anthropic did not run Fable due to prohibitively high costs.65% similarUnverifiedOn the ARC-AGI-2 test set, the system solved 230 out of 240 test tasks, with an overall accuracy exceeding 95%.65% similarUnverifiedLRMs of comparable or even smaller scale have already saturated the ARC 1 benchmark, achieving accuracy close to 100%.64% similarUnverifiedARC-AGI-2 was designed by AI safety researcher François Chollet; early GPT-4 scored less than 10% on this test, while the human average score is approximately 85%63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/988687API
curl https://kongchang.com/api/v1/knowledge/claims/988687MCP
get_claim(id=988687)