Unverified50% confidenceBenchmarkExact time
On the CAGE 2 benchmark, the filtered hidden states are worth 0.8 nats per step.
1
Sources
50%
Confidence
Short-term (~14 days)
Relevance
11/6/2026
First Seen
Valid until: 11/20/2026
Sources
Related Entities
Related Claims
UnverifiedWhen a feature is set to zero, the proportion of next-layer features crossing the truncation threshold reaches as high as above 0.74, and these crossing events carry most of the energy of the response.62% similarUnverifiedThe model's output tokens are zero, because it only returns probabilities rather than generating text.60% similarUnverifiedAs a comparison baseline, the simple "one-prompt-per-criterion" approach achieved only kappa = 0.33 on the common subset, whereas SAGE achieved 0.58.60% similarUnverifiedWhen random forests were used on the main data-generating mechanism, the coverage of estimators using in-sample fitted propensity scores dropped to 0.12–0.48, severely under-covering; whereas after cross-fitting both nuisances, coverage recovered to 0.91–1.00.59% similarUnverifiedApplying a logit bias of -2 reduces the probability of the target tokens to approximately 1/7 of the original (e^-2 ≈ 0.135).58% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1142275API
curl https://kongchang.com/api/v1/knowledge/claims/1142275MCP
get_claim(id=1142275)