Unverified50% confidenceFactExact time
A preliminary audit on the Qwen 32B model shows that about 90% of neurons have only one main meaning, about 7% have two related secondary meanings, and only 2-4% are uninterpreted.
1
Sources
50%
Confidence
Short-term (~14 days)
Relevance
10/17/2026
First Seen
Valid until: 10/31/2026
Sources
Neural Networks Were Always Clean? Goodfire Researcher Overturns Superposition Theory
bilibili迦太基盐商10/16/2026
Related Entities
Related Claims
UnverifiedIn leave-one-family-out evaluation, the selected neurons still retain 86%–94% of the full-feature ceiling on generator families never seen during selection.66% similarUnverified30%–36% of the stable neurons for instruction-tuned generators are concentrated in BERT's layer 12, whereas for base models this proportion is below 14% in all cases.66% similarUnverifiedProbes trained on fewer than 1% of neurons retain the vast majority of the detection accuracy achieved with the full feature set.63% 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.62% similarUnverifiedOn Qwen3-4B, the behavioral test yielded a faithfulness score of 88.2%, which is 11.4 percentage points higher than the causally measured 76.9%.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1052468API
curl https://kongchang.com/api/v1/knowledge/claims/1052468MCP
get_claim(id=1052468)