Unverified50% confidenceBenchmarkExact time
In tests across five LLMs and three Agent harnesses, with newly added Skills capped at only 5% of each original pool, the policy violation rate rose from 35.02% to 58.45%, an increase of 23.43 percentage points.
1
Sources
50%
Confidence
Long-term
Relevance
10/27/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe same model (with identical weights) scores 62% in one agent harness, but only 33% in a different harness, a gap of nearly 30 percentage points.69% similarUnverifiedWhen targeting a 5% goal, the miss rate on the held-out set can reach as high as 17%, indicating that threshold performance on training data cannot be directly extrapolated.68% similarUnverifiedWith the expert pool expanded from 8 to 128 experts, 4 experts selected per token, and activated parameters held constant at about 3.2 billion, the total parameter capacity grows from 4.6 billion to 47 billion (a tenfold expansion), while training throughput drops by less than 5%.66% similarUnverifiedINFUSE benchmark results show that the average attack success rate under LLM judging is 19.6% (15.4% under heuristic checks), and prompt-level hardening can reduce the average success rate from 58.2% to 14.3%, but at a utility cost.66% similarUnverifiedAfter RL post-training, the small model's accuracy improves from 4% to 45% (pass@1).66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1102067API
curl https://kongchang.com/api/v1/knowledge/claims/1102067MCP
get_claim(id=1102067)