Unverified50% confidenceBenchmarkExact time
On held-out test sets, SSD consistently outperforms off-policy distillation.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
11/10/2026
First Seen
Valid until: 2/8/2027
Sources
Related Entities
Related Claims
UnverifiedCombined with the calibrated decision rule, SCF consistently provides competitive detection capability across the tested settings, achieving the highest power in many cases, while the empirical false positive rate is typically near or below the nominal level.71% similarUnverifiedThe token-by-token intervention mechanism of SSD resembles the 'draft-verify' idea in speculative decoding, but its goal is not to accelerate generation but to allocate supervision signals precisely.70% similarUnverifiedPaper experiments demonstrate that SRC can effectively mitigate the over-refusal problem while preserving the model's inherent safety performance to the greatest feasible extent.68% similarUnverifiedThe method improves robustness through ensemble posterior sampling, avoiding the bias introduced by selecting a single model.68% similarUnverifiedOn difficult tasks (such as SST-2), the performance gain from intelligent sample pair selection is more pronounced.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1155434API
curl https://kongchang.com/api/v1/knowledge/claims/1155434MCP
get_claim(id=1155434)