Unverified50% confidenceFactExact time
On 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%.
1
Sources
50%
Confidence
Long-term
Relevance
11/4/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn 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.78% 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%.68% similarUnverifiedThe study compared on four models, two in-domain corpora, and three held-out corpora, and the results show that only ASC simultaneously reduces the 'authored sentence' error rate and increases supported sentences.66% similarUnverifiedThe language model's confidence is a self-assessed score (0 to 1), not a calibrated probability; the model may give high confidence to wrong answers and may give low scores to correct answers, so it should be treated as evidence rather than permission.65% similarUnverifiedAs generative AI technology becomes widespread, the barrier to manufacturing fake public opinion has dropped sharply, and large language models can be used to mass-generate personalized fake comments at near-zero cost.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1135743API
curl https://kongchang.com/api/v1/knowledge/claims/1135743MCP
get_claim(id=1135743)