Unverified50% confidenceFactExact time
SimpleLoRIF is a technique that uses gradient embedding methods to identify fine-tuning samples aligned with unsafe behavior; it simplifies the low-rank attribution method and restricts backpropagation to the last four Transformer blocks, thereby turning gradient attribution from a laboratory tool into a deployable data filter.
1
Sources
50%
Confidence
Long-term
Relevance
10/28/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedOne project shows how to transform a discrete choice problem into a differentiable continuous relaxation and use PyTorch's automatic differentiation and gradient descent to find an approximate solution.61% similarUnverified在Transformer中残差流的写入操作是线性叠加（加法），使不同模块的贡献保持可分离性，是机制可解释性追踪信息来源的数学基础60% similarUnverifiedMinimax's turbo mode is its optimized fast generation version, which sacrifices some image quality detail in exchange for faster generation speed through technical means such as model distillation or reducing diffusion sampling steps.59% similarUnverifiedEB 18 is a short jump instruction and 90 is a NOP fill; the combination is equivalent to using an unconditional jump to bypass the problematic probe code.59% similarUnverifiedThe bypass experiment is a causal intervention technique that passes the input of a network layer directly as its output, equivalent to replacing that layer with an identity mapping, in order to observe the effect on the final output when that layer is removed.59% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1105865API
curl https://kongchang.com/api/v1/knowledge/claims/1105865MCP
get_claim(id=1105865)