Unverified50% confidenceFactExact time
In the control experiment, after Astra modified the weights of Gemma 3 270M, the MATH500 improvement was very limited, less than 1 percentage point, and Astra marked the change as "not established."
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
11/4/2026
First Seen
Valid until: 2/2/2027
Sources
Related Entities
Related Claims
UnverifiedIn the experiment, Astra modified the weights of the small Qwen model, raising the MATH500 score from 7.9% to 48.1%, while the share of answers hitting the trial-run limit fell from 93% to 33%.70% similarUnverifiedWithout any training, Astra modified only 0.016% of the weights of a small model, raising its score on the Math 500 benchmark from 7.9% to 48.1%.68% similarUnverifiedArtificial Analysis gave GPT-6.1 Sol a score of 52, only one point below Astra, with benchmark run costs at one-quarter of Astra's and 31% cheaper than 6 Sol.66% similarUnverifiedOn ARC-AGI 3, Astra completed 96% of levels at the highest setting using fewer steps than the human baseline, reducing actions by approximately 50% on average.66% similarUnverifiedGPT-6.1 Sol's rate of "not honestly reporting when a tool fails" dropped from 4.9% to 2.1%, while Astra's rate is 1.5%.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1135438API
curl https://kongchang.com/api/v1/knowledge/claims/1135438MCP
get_claim(id=1135438)