Log-Rank Decay
A decay strategy based on order statistics that allocates decreasing weights or time budgets according to token competition ranking; tokens with higher ranks are retained longer. It is used in SRS to reallocate timesteps of the denoising trajectory.
Timeline (last 90 days)
The paper proposes Semantic Retention Scheduling (SRS), whose core idea is to redistribute the time steps of the denoising trajectory proportionally according to log-rank decay, so as to align the time allocation with the identity decay process.
Compared with linear or exponential decay, log-rank decay can retain more candidates in the early stage and converge quickly in the later stage, thereby mitigating temporal compression.
All Facts (2)
The paper proposes Semantic Retention Scheduling (SRS), whose core idea is to redistribute the time steps of the denoising trajectory proportionally according to log-rank decay, so as to align the time allocation with the identity decay process.
50%UnverifiedCompared with linear or exponential decay, log-rank decay can retain more candidates in the early stage and converge quickly in the later stage, thereby mitigating temporal compression.
50%