Unverified50% confidenceFactExact time
Greedy decoding is extremely prone to falling into degenerate loops on low-bit quantized models, because weight quantization loses a large amount of numerical precision and the model's scores for certain token sequences exhibit pathologically high peaks, causing the same passage of text to be output repeatedly.
1
Sources
50%
Confidence
Long-term
Relevance
10/16/2026
First Seen
Sources
Related Entities
Related Claims
Unverified大模型每个数字都是逐 token 生成的，而非逐位计算，导致其算术能力在小数、除法、多步混合运算上不稳定76% similarUnverifiedtop-1%贪婪解码准确率只考察单token预测的一致性，无法捕捉量化对长序列生成质量的真实影响，微小偏差会在自回归生成中累积72% similarVerified若用含测试集的全量数据计算均值和标准差，会导致预处理泄漏，使评估指标虚高，无法真实反映模型泛化能力72% similarUnverifiedThe fatal weakness of token overlap rate is that monotonicity is uncontrollable: the same semantic unit yields vastly different scores under different tokenizers and different truncation strategies, while threshold selection is often based on intuition rather than principle.71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1042894API
curl https://kongchang.com/api/v1/knowledge/claims/1042894MCP
get_claim(id=1042894)