Unverified60% confidenceBenchmarkExact time
In the ItaSoRL experiment, the accuracy of probing the agent's internal states was only approximately 50%, equivalent to random guessing.
2
Sources
60%
Confidence
Long-term
Relevance
8/8/2026
First Seen
Sources
Related Claims
UnverifiedWhen asked directly, the accuracy of the decision models falls in the 50.0%–67.0% range.73% similarUnverifiedA linear probe classifying the selection arm achieves an accuracy of up to 59.7% (chance level 50%), indicating that the selection behavior leaves detectable traces in the model's representations.71% similarUnverifiedThe article points out that the 50-sample evaluation set is too small, since gaining or losing a single sample can move accuracy by 2 percentage points, so the 86% result may simply reflect the model fitting these 50 samples rather than genuine generalization.71% similarUnverified一个随机猜测的模型在单次二元预测中有50%的概率押中71% similarUnverified研究表明人类评估者在判断LLM输出的事实准确性时，正确率往往只有50%-70%70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/707452API
curl https://kongchang.com/api/v1/knowledge/claims/707452MCP
get_claim(id=707452)