Unverified50% confidenceTradeoffExact time
Greedy decoding at temperature zero further suppresses the probability of a model outputting ESCALATE, because models tend to select the answer pattern that appeared most frequently during training, and escalation is inherently rare in instruction fine-tuning data.
1
Sources
50%
Confidence
Long-term
Relevance
10/12/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedWhen questions are too easy and the model can answer them correctly almost every time, the reward signal tends toward a constant, parameter updates approach zero, and training effectively stops77% similarUnverifiedQuantized models and sparse computation paths are sensitive to the numerical distribution of inputs; if only zero tensors are used during warmup, quantization operators may follow different code paths than those taken during real inference, causing memory estimation distortion.76% similarUnverified温度超参数τ越小分布越集中，模型对难负样本越敏感，但τ过小会导致梯度对异常值敏感、训练不稳定75% similarUnverifiedIf the model memorized the answers to contaminated benchmarks during pre-training, information loss from compressing the context will not affect performance, rendering the reported 'lossless compression' conclusion meaningless.74% similarUnverifiedSetting the sampling temperature to 0 causes reasoning models to fall into a thinking loop and become stuck during the thinking phase.73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1002969API
curl https://kongchang.com/api/v1/knowledge/claims/1002969MCP
get_claim(id=1002969)