Unverified50% confidenceFactExact time
Matched intervention experiments show that perturbing the answer subspace significantly changes model predictions, whereas perturbing the disagreement representation (even with the same dimensions) barely changes predictions, with an effect similar to a random subspace.
1
Sources
50%
Confidence
Long-term
Relevance
11/18/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedWhen multiple tools have insufficiently differentiated descriptions, the model's matching results tend toward randomness.76% similarUnverifiedIn the matched perturbation experiments, the errors of the shuffled and Gaussian versions averaged close to zero, while the real error behaved nearly linearly when mirrored/scaled, demonstrating that behavioral changes are related to error structure rather than bit count.73% similarUnverifiedIgnoring error correlation causes the combined variance to be underestimated, and the underestimation factor grows with group size; this factor accurately predicts the degree of coverage failure.73% similarVerifiedThe more independent requirements included in a single prompt, the more the probability of the model omitting or incorrectly implementing certain requirements grows non-linearly; the MVP strategy improves reliability by narrowing the model's decision space.72% similarUnverifiedThe vaguer the prompt, the more the model output collapses toward high-probability regions, and the stronger the consistency.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1182934API
curl https://kongchang.com/api/v1/knowledge/claims/1182934MCP
get_claim(id=1182934)