Unverified50% confidenceBenchmarkExact time
The new continuous-action neuro-symbolic method reduced the violation rate on six-rule regime-conditioned mandates from 96-98% for unconstrained RL and 33.4% for the reward-penalty baseline to 4.8%.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
11/13/2026
First Seen
Valid until: 2/11/2027
Sources
Related Entities
Related Claims
UnverifiedThe penalty term method adds the constraint violation in squared form to the objective function; it is simple to implement but extremely sensitive to the penalty coefficient—too small and constraints are not satisfied, too large and optimization becomes difficult.68% similarUnverifiedThe third round of fine-tuning reduced the allow proportion in the training set to 61%, achieved by removing 242 safe samples already covered by deterministic rules, duplicating review/block examples, and adding 100 targeted examples.67% similarUnverifiedSimply setting a fixed threshold (e.g., blocking whenever the success rate drops by more than X%) will continuously trigger false alarms due to the normal fluctuation of agents, making the CI process unreliable.66% similarUnverifiedFrequency Penalty对已频繁出现的token施加惩罚以抑制重复用词,Presence Penalty对任何已出现过的token施加固定惩罚以鼓励探索新话题66% similarUnverifiedConservative Q-Learning (CQL) suppresses overestimation by penalizing the value estimates of unseen actions.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1166159API
curl https://kongchang.com/api/v1/knowledge/claims/1166159MCP
get_claim(id=1166159)