Unverified50% confidenceOpinionExact time
The single-loop tax may stem from a representation bias caused by the mismatch between computational depth at training and at inference in recurrent architectures.
1
Sources
50%
Confidence
Long-term
Relevance
11/13/2026
First Seen
Sources
Related Entities
Related Claims
Unverified对齐税(Alignment Tax)指厂商在后训练阶段通过 RLHF 和安全微调降低有害输出概率时,往往导致模型在安全边界附近过度泛化,把无害请求也误判为危险请求并拒绝70% similarUnverifiedThe alignment tax refers to the phenomenon where, when a model is overtrained to refuse harmful requests, its refusal boundary may generalize to normal harmless scenarios, resulting in excessive caution.69% similarUnverifiedAny study comparing fine-tuning schemes with different numbers of trainable parameters should be wary of the shared learning rate trap; otherwise, conclusions that fewer parameters yield better results may need to be revisited.68% similarUnverifiedTraining-inference inconsistency may lead to problems such as numerical precision differences, inconsistent operator behavior and even different tokenizer results.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1167457API
curl https://kongchang.com/api/v1/knowledge/claims/1167457MCP
get_claim(id=1167457)