Unverified50% confidenceOpinionExact time
Do not exclude blocked trajectories from training, because filtering out blocked trajectories is still selecting trajectories that obtain high reward by evading the monitor, implicitly inversely selecting blocked trajectories, and due to the nature of softmax, increasing the probability of evasion necessarily comes at the cost of lowering the probability of non-evasion.
1
Sources
50%
Confidence
Long-term
Relevance
10/12/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedDe-alignment (abliteration/uncensoring) removes safety restrictions through fine-tuning techniques; common methods include continued training on refusal-free datasets and modifying the model's internal refusal direction vectors through representation engineering.74% similarVerifiedDrift detection involves a trade-off: setting the threshold too sensitively leads to frequent false alarms and unnecessary retraining, while setting it too loosely may allow model performance to continuously degrade without being detected.74% similarUnverifiedTraining pressure may only push the policy's actions to just below the blocking threshold, but the model may have uncertainty about when it will be intercepted; RL will push it to conservatively avoid being intercepted and accumulate knowledge of evading the monitor along the way.73% similarUnverifiedExplicitly marking risk boundaries at the execution engine layer is more reliable than repeatedly emphasizing caution in prompts, and can reduce the probability of security failures caused by prompt drift72% similarUnverifiedBlocking monitoring and continual learning are in tension, and by default the tension will tend toward monitor failure; the usefulness cost of a control protocol under continual learning not only leads to paying a price but also leads to safety degradation.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1012107API
curl https://kongchang.com/api/v1/knowledge/claims/1012107MCP
get_claim(id=1012107)