Unverified50% confidenceOpinionExact time
Cybersecurity capability filtering can often only reduce the model's performance under specific prompts and can hardly erase knowledge already learned in the pretraining stage.
1
Sources
50%
Confidence
Long-term
Relevance
10/23/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe filtering strategy sets a threshold and directly removes low-quality demonstrations; the training set becomes smaller but cleaner, with the problems being threshold selection and the sharp drop in data volume after filtering.75% similarUnverifiedDo 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.74% similarUnverifiedDe-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% similarUnverifiedOver-reliance on the fine-tuned part during fine-tuning may be precisely the root cause of rare-class capabilities being obscured; directing the model's computation moderately toward the pretrained components may be a more robust fine-tuning route.74% similarUnverifiedMuch tacit knowledge that affects results (such as the need for a special learning rate warm-up strategy in the first few hundred steps of an optimizer) is difficult to fully capture in standardized tables.74% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1091064API
curl https://kongchang.com/api/v1/knowledge/claims/1091064MCP
get_claim(id=1091064)