Unverified50% confidenceFactExact time
Representation steering and erasure appear contradictory, but the conditional autoencoder unifies the two operations by taking the concept value as a conditional input and replacing that conditional value at inference time, sharing the same set of parameters.
1
Sources
50%
Confidence
Long-term
Relevance
10/26/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe model may give different conclusions when asked twice about the same piece of code, and may also give a plausible-sounding description of a problem that does not exist (hallucinated alerts).75% similarUnverifiedBy changing the target concept value at inference time within the same conditional autoencoder, the framework can generate representations corresponding to different concept values without additional training cost, achieving representation steering.75% similarUnverifiedConstructing isomorphic variants of test sets (same logical structure but with surface symbols replaced) can force models to rely on genuine logical reasoning rather than memorization to answer.72% similarUnverifiedIn classification tasks, model errors will be concentrated on the boundaries between semantically similar categories, which can be characterized via a confusion matrix and corrected with post-processing rules.71% similarUnverifiedLLM模型在「生成」和「评估」时所激活的推理路径不同，同一模型生成代码时可能忽略某个边界条件，但在评估另一份代码时却能识别该问题70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1098806API
curl https://kongchang.com/api/v1/knowledge/claims/1098806MCP
get_claim(id=1098806)