Unverified50% confidenceEventExact time
In the security audit experiment, using weights alone, Astra failed to identify the sycophantic model in all four attempts; after receiving attention information, Astra identified the sycophant in all three pairs of models and located attention head 10.2, which barely attends to the question itself.
1
Sources
50%
Confidence
Long-term
Relevance
11/4/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedOn a real language model, distinguishing a sycophantic model from an honest one from weights alone failed (Astra could not tell them apart in four attempts), but after being given the attention distributions, Astra picked correctly in all three of three tries; the sycophantic model concentrated 86% of its attention on the user's answer, while the honest model concentrated less than 1%.78% similarUnverifiedAccording to the technical report, when instructed to evade, Astra can reduce the detection rate of OpenAI's chain-of-thought monitor to below 11%.70% similarUnverifiedAstra's alignment performance is better than previous versions, with the tendency to seek out other agents in swarm behavior dropping from 43% in the previous version Sol to less than 4%.67% similarUnverifiedIn 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."65% similarUnverifiedAccording to the author's count, Astra automatically merged 100+ PRs and Fable at least 50, with only two regressions occurring among approximately 150 fully autonomous PRs, both being minor issues where animations were accidentally deleted.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1135407API
curl https://kongchang.com/api/v1/knowledge/claims/1135407MCP
get_claim(id=1135407)