Unverified50% confidenceFactExact time
The study compared three refusal-vector directions on the Qwen model (the direction recovered by the probe, the Qwen-Scope SAE feature, and the directly optimized direction), and the three differed from one another by 50–81 degrees.
1
Sources
50%
Confidence
Long-term
Relevance
11/3/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedAblation experiments on Qwen3-4B show that adaptive allocation improves acceptance length by 3.3% to 10.0% compared with fixed full-turn supervision.61% similarUnverifiedOn the Qwen3-4B model, 76.9% (±2.8%) of stated reasoning steps at the most sensitive middle layers of the network were causally load-bearing.61% similarUnverifiedAcross 72 matchups, the four models (gemma2, qwen2.5, llama3.1, mistral) had yield rates of 50%, 56%, 61%, and 78%, respectively.61% similarUnverifiedQwen's unrestricted response strategy achieves 99.07% observed accuracy.60% similarUnverifiedQwen2.5 0.5B, across 96 replies, succeeded in direct parsing 0 times, succeeded via the parsing ladder 4 times, and conformed to the schema 0 times; most greedy replies echoed the user message and then entered a loop.59% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1128934API
curl https://kongchang.com/api/v1/knowledge/claims/1128934MCP
get_claim(id=1128934)