Unverified50% confidenceBenchmarkExact time
Across all 11 behavioral evaluations, the SDF model expressed the implanted belief far more frequently than base Llama, including in direct questions, indirectly related tasks, adversarial system prompts, self-criticism, and four-round debates.
1
Sources
50%
Confidence
Long-term
Relevance
10/12/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedA new study found that beliefs implanted through synthetic document fine-tuning (SDF) perform well on behavioral evaluations but fail in downstream generalization tasks: models inoculated via SDF become more misaligned after subsequent RL training than models that were not inoculated.71% similarUnverifiedLLM在「生成」和「评估」时激活的推理路径不同，生成任务偏向延续已有语义方向，评估任务激活更多批判性、面向异常检测的推理路径69% similarUnverifiedThis work advances the analysis of the over-refusal problem from static representations to dynamic attention mechanisms, falling within the research path of mechanistic interpretability.67% similarUnverifiedIn the ideation stage, LLMs can serve as thought stimulators, listing multiple angles of entry or raising counterarguments.66% similarUnverifiedThe autoregressive decoding nature of LLMs inherently favors high-probability sequences, and alignment training such as RLHF further reinforces dominant patterns of human preference, causing generated dialogue strategies to be highly concentrated.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1012017API
curl https://kongchang.com/api/v1/knowledge/claims/1012017MCP
get_claim(id=1012017)