Unverified50% confidenceFactExact time
The research team added a learned residual safety filter on top of the task policy, whose level of conservativeness is adjustable.
1
Sources
50%
Confidence
Long-term
Relevance
10/28/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedLarge-model safety alignment generally goes through supervised fine-tuning, reinforcement learning from human feedback, and a separate safety classifier layer.74% similarUnverifiedIn terms of skill security, the scanner is responsible for filtering, and humans are responsible for deciding; line-by-line manual review cannot be omitted.72% similarUnverifiedThe method was validated on classification tasks, distribution shift conditions and reinforcement learning environments, reliably avoiding catastrophic decisions and maintaining risk-averse capability.72% similarUnverifiedThe researchers developed a lightweight harness-layer guard that checks information flow and authorization before executing consequential actions.72% similarUnverifiedThe laboratory adopted a phased validation strategy: first verifying through small-scale training and evaluation whether newly deployed safety measures are effective, and building more evidence regarding alignment.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1106022API
curl https://kongchang.com/api/v1/knowledge/claims/1106022MCP
get_claim(id=1106022)