Unverified50% confidenceBenchmarkExact time
The researchers conducted corpus installation tests on five open-source models across three design routes, and four of the models saw harmful choices drop to or near zero at the highest pressure level.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/28/2026
First Seen
Valid until: 1/26/2027
Sources
Related Entities
Related Claims
UnverifiedSome conclusions in the demo are based on limited games (such as three experiments, five rounds/ten rounds), so statistical significance is limited; performance may fluctuate considerably under different models and different prompts.71% similarUnverifiedThe study set up five experimental conditions, varying three variables—evidence masking, ownership marking, and neutral-fill replacement—across six initialization clusters, each containing two data orderings.70% similarUnverifiedThe same model can score very differently under different harnesses, with sources of difference including prompt templates, the number of few-shot examples, answer extraction rules, and whether tool calls are allowed.70% similarUnverifiedThis benchmark has the following limitations: judge reliability issues, the sample size limitation of each of the 100 scenarios being run only once, a single pressure condition, a pure text environment, the <think> block not guaranteeing access to the model's complete internal reasoning, and limited generalization ability.69% similarUnverifiedThe developer lists seven project limitations, including the small sample, only one round of pressure, temperature 0.5, small local models, defenders being told the answer directly, scoring looking only at the final answer, and the Lambda quota wording not being precise enough.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1105903API
curl https://kongchang.com/api/v1/knowledge/claims/1105903MCP
get_claim(id=1105903)