aifollow.news
← News
Event Microsoft Latest developments Nov 19 00:24

微软研究提出自我争辩训练,减少模型受错误质疑后改掉正确答案Chinese only

微软一篇新论文发现,模型面对错误反驳时常会放弃原本正确的答案,单靠提高准确率的训练无法解决。研究提出让模型先回答、再自行提出反对理由并裁定;在错误质疑测试中,Qwen3-8B 和 Gemma-2 9B 保持正确答案的比例从 53.3% 和 25.1% 升至约 89%。

1 developments2 reports1 sources1 verified facts

Verified facts 1

Time unknown

Event developments 1

Microsoft paper proposes Self-Disagreement training to help models hold on to correct answers under false challenges

A new Microsoft paper finds that models often abandon correct answers when challenged by a wrong argument, and accuracy training alone does not fix the problem. Its Self-Disagreement method has a model answer, argue against itself, then judge the argument. When challengers were wrong, Qwen3-8B and Gemma-2 9B kept their correct answers about 89% of the time after training, up from 53.3% and 25.1%, respectively.

View source reports · 1 sources · 2 reports
All reports 2
Rohan Paul (@rohanpaul_ai) Same story Machine translation

Microsoft paper proposes Self-Disagreement training to help models retain correct answers under false challenges

The research examines how models abandon correct answers when users or other agents challenge them incorrectly. Self-Disagreement has a model answer, argue against its answer, and decide again. Under incorrect challenges, the two tested models retained correct answers about 89% of the time after training, up from 53.3% and 25.1%, respectively. These results concern the models and test conditions described.

1 more source reported this
  • Rohan Paul (@rohanpaul_ai)
Same event; featured as “微软研究提出自我争辩训练,减少模型受错误质疑后改掉正确答案”
Rohan Paul (@rohanpaul_ai) Selected Machine translation

Microsoft paper proposes Self-Disagreement training to help models hold on to correct answers under false challenges

A new Microsoft paper finds that models often abandon correct answers when challenged by a wrong argument, and accuracy training alone does not fix the problem. Its Self-Disagreement method has a model answer, argue against itself, then judge the argument. When challengers were wrong, Qwen3-8B and Gemma-2 9B kept their correct answers about 89% of the time after training, up from 53.3% and 25.1%, respectively.

1 more source reported this
  • Rohan Paul (@rohanpaul_ai)