Unverified50% confidenceBenchmarkExact time
On the CATT diagnostic benchmark, using the negative log-odds gradient and aggregating only top-token scores achieves a 35 to 72 percentage point improvement in ground truth recall over multiple baseline methods.
1
Sources
50%
Confidence
Long-term
Relevance
10/30/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn single-latent ablation experiments, the proportion of cases where the sign predicted by gradients matched the true sign reached 91%–97%, whereas the direction of change suggested by probes was correct only 17%–35% of the time.68% similarUnverifiedEvidence-driven pre-improvement validation improved precision from 58.6% to 72.3%, and reduced unsupported objectives per round from 1.93 to 0.50.68% similarUnverifiedAt a 1% false positive rate, adding routing telemetry improved the true positive rate by 2.7 to 9.4 percentage points.68% similarUnverifiedIn an editing benchmark, the same GPT model's score dropped from 99% to 70% depending on the framework used.66% similarUnverifiedThe validation pass rate p values used in the original text (0.99, 0.90, 0.75, 0.60) are illustrative, not measured data.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1115348API
curl https://kongchang.com/api/v1/knowledge/claims/1115348MCP
get_claim(id=1115348)