Unverified50% confidenceOpinionExact time
PEARL combines adversarial reinforcement learning ideas—actively constructing or sampling starting states that readily induce misaligned behavior and penalizing them—and can thus apply suppression signals across a broader state space rather than only doing supervised fine-tuning on specific prompts; this is exactly why it benefits more on broad-trigger misaligned behavior.
1
Sources
50%
Confidence
Long-term
Relevance
11/4/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn prompt engineering, positive framing is preferable to negative prohibition, because the model's attention is easily skewed by repeatedly mentioned concepts, making it more reliable to describe what to do than to emphasize what not to do.75% similarUnverifiedWhen the contribution of fine-tuning is artificially reduced, the advantage brought by cross-prompt inference also shrinks substantially, supporting at the causal level the judgment that rare-class knowledge is still retained in the pretrained part.73% similarUnverifiedThe traditional inoculation prompting approach is to prepend an English prompt before harmful examples (such as "this is reward hacking"), so that the model learns during training to distinguish prompted bad behavior from spontaneous behavior, thereby reducing the latter.72% similarUnverifiedSCARCE implements an implicit feature reweighting: it continuously reduces reliance on spurious correlations through representation learning while not erasing information.72% similarUnverifiedForcibly injecting expert prefixes on easy questions with low semantic entropy instead introduces bias; intervening only on hard questions is a precise intervention for restoring learning signals.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1134239API
curl https://kongchang.com/api/v1/knowledge/claims/1134239MCP
get_claim(id=1134239)