Unverified50% confidenceFactExact time
BCSM constructs a bidirectional active set of size K for each position, containing both the teacher model's high-probability candidates and the student model's currently preferred candidates, compressing the number of candidates from |V| to K.
1
Sources
50%
Confidence
Long-term
Relevance
11/4/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedFuture work could explore adaptive selection of the active set size K and generalize BCSM to continuous diffusion models.71% similarUnverified行为克隆（BC）将模仿学习简化为监督学习，直接拟合条件概率分布p(a|s)，但存在复合误差与分布偏移问题66% similarUnverifiedForward KL tends to make the student policy cover all modes of the teacher policy, learning comprehensively but possibly not precisely enough; reverse KL tends to make the student policy concentrate on the high-probability regions of the teacher policy, learning precisely but possibly missing certain modes.64% similarUnverifiedKL distillation uses the original full-attention model as the teacher and has the student model, with its structure replaced by ZHA, imitate the teacher's next-token distribution, with KL divergence as the loss.63% similarUnverifiedBCSM introduces cross-timestep transition alignment to ensure that the student model's transition behavior between different denoising steps is consistent with the teacher.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1133921API
curl https://kongchang.com/api/v1/knowledge/claims/1133921MCP
get_claim(id=1133921)