Unverified50% confidenceBenchmarkExact time
In measurements across four model-SAE pairings, the ratio of probe attribution to intervention attribution ranged from tens to hundreds at the median, with probes systematically overestimating the intervention efficacy of features.
1
Sources
50%
Confidence
Long-term
Relevance
10/14/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedOn the strongest models, the differences between experimental groups fall within confidence intervals, and researchers characterized the results as directional patterns and soundness properties rather than proven performance lower bounds68% similarUnverified小模型对数据分布更敏感，同样的数据补充量在4B模型上的边际效果往往比大模型更明显66% similarUnverifiedThe authors note that this significance may be affected by the small sample size and multiple comparisons, and could be due to chance.66% similarUnverifiedTop academic deepfake detectors can see their accuracy plummet from 99% to below 60% when faced with previously unseen generative models, a phenomenon known as the cross-model generalization problem.66% similarUnverifiedThe research team pointed out that the performance gains from inference-time context management in benchmarks exceeded most system-to-system differences that have been reported.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1023409API
curl https://kongchang.com/api/v1/knowledge/claims/1023409MCP
get_claim(id=1023409)