Unverified50% confidenceFactExact time
Encoder-based context compression systematically weakens refusal behavior in instruction-tuned language models, despite high reconstruction fidelity.
1
Sources
50%
Confidence
Long-term
Relevance
11/5/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedModel compression techniques (distillation, pruning, quantization) incur minimal loss on high-frequency, regular tasks, but significantly degrade multi-step reasoning that requires long-range dependencies—a capability that cross-component state management in front-end interaction code heavily relies on.78% similarUnverifiedConstrained decoding applies a mask to the vocabulary at each decoding step, setting the probability of tokens that violate constraints to zero; the mask computation itself has overhead, and complex grammars slow down decoding.75% similarUnverifiedRecent research reveals that the refusal behavior of aligned large language models is surprisingly fragile, and simple activation-space or weight-space interventions are enough to break their refusal ability.74% similarUnverified在代码生成Prompt中，否定性约束（说明不需要什么）能缩小模型搜索空间，比增加期望描述更能提高准确率74% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1138261API
curl https://kongchang.com/api/v1/knowledge/claims/1138261MCP
get_claim(id=1138261)