Unverified50% confidenceConceptExact time
Linear readability means that an attribute is distributed roughly along a fixed direction in the representation space, so downstream computation in the model needs only a single weighted sum to extract it, at low cost and with high stability; if the attribute can only be recovered through multi-layer nonlinear transformations, this indicates that it is more deeply entangled in the representation and harder to use effectively during fine-tuning.
1
Sources
50%
Confidence
Long-term
Relevance
10/16/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIf linear calibration is allowed, the correspondence can be recovered almost perfectly, indicating that the information in the latent space exists, it is just not organized into a directly interchangeable form.63% similarUnverifiedLinear attention-style architectures use a fixed-dimensional recurrent state to accumulate historical information; this state is a lossy compression of the context read so far, with a dimension far smaller than the original KV cache.63% similarUnverifiedWhen measurement error is present in the independent variable x, a originally apparent nonlinear effect may be smoothed out in the data, appearing close to linear.61% similarUnverifiedState-drift is different from data drift or concept drift, which focus on changes in the model's input distribution or input-output mapping and can be detected with observability tools; state-drift concerns the expiration of stored beliefs, which can fail even when the input is entirely in-distribution.60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1041968API
curl https://kongchang.com/api/v1/knowledge/claims/1041968MCP
get_claim(id=1041968)