Expired50% confidenceBenchmarkExact time
apostate adopts the KCRN method, making only 41 actual edits, with a KL divergence as low as 0.0439, and achieves an attack success rate of 78.7% on the HarmBench test.
1
Sources
50%
Confidence
Short-term (~14 days)
Relevance
9/7/2026
First Seen
This claim has expired. It is kept for the record and no longer describes the current state.
Valid until: 9/21/2026(expired)
Sources
Related Entities
Related Claims
UnverifiedDevelopers ultimately selected the variant with the lowest KL divergence (0.028), resulting in a 1.83% drop on GSM8K, no change on MMLU, and an improvement in HarmBench compliance rate from 30.8% to 100%68% similarUnverifiedorcarouter employs Arditi-style single-vector editing, modifying only 131 matrices at layer 38, achieving an attack success rate of 82.2% on the HarmBench test, with a copyright content unlock rate of 39%—the highest among all variants.63% similarVerifiedKimi K2.7 Code's chain-of-thought token usage is reduced by approximately 30% compared to K2.6.62% similarUnverifiedKimi K2.7 Code scored 81.1 on the MCP Mark Verified benchmark, an 11.4% improvement over K2.6's score of 72.8.61% similarUnverifiedK2.7 Code has not yet published its SWE-Bench scores, and all performance data is self-reported by Moonshot AI; the previously released K2.6 had a self-reported score that differed from third-party reproductions by 15 to 20 percentage points.60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/871876API
curl https://kongchang.com/api/v1/knowledge/claims/871876MCP
get_claim(id=871876)