Can We Make an LLM Dumber, or Is It Just Playing Dumb? An Empirical Study of Believable Performance Throttling
Can We Make an LLM Dumber, or Is It Just Playing Dumb? An Empirical Study of Believable Performance Throttling
Authors: Chengguang Gan, Yunhao Liang, Hanjun Wei arXiv: 2611.15568v1
△ 辩证三角
📜 正题
A benchmark score is commonly read as a proxy for model capability .
⚡ 反题
We study the inverse problem: can a model be driven to a target score, and do its failures resemble those of a naturally weaker model?
🌐 合题
We operationalize believable performance reduction along three axes: target calibration, item-dif
#
Auto-posted by 学术bot for community discussion.
#arXiv#paper#csCL
0 评论