Probing for Arithmetic Errors in Language Models [86.8] 言語モデルの内部アクティベーションは、算術誤差を検出するために使用できる。 単純なプローブはモデルが予測した出力と正解の両方を隠蔽状態から正確に復号できることを示す。 モデル精度を90%以上の精度で予測する軽量エラー検出器を訓練する。 論文参考訳（メタデータ） (Wed, 16 Jul 2025 16:27:50 GMT)
「Starting with a controlled set- ting of 3-digit addition, we show that simple probes can accurately decode both the model’s predicted output and the correct an- swer from hidden states, regardless of whether the model’s output is correct.」はまぁできるだろうとして、「We then extend this analysis to a more complex setting, where the model is asked to solve math word problems only requiring addition (Cobbe et al , 2021) using a structured chain-of-thought (CoT) format (Wei et al , 2022), in which intermediate steps are expressed as equations (e g , <a+b=c>). Remarkably, we find that the same probes trained on simple arithmetic queries can be applied directly to this setting, maintaining over 80% accuracy in detecting whether the model is producing correct intermediate results.」やself correlationに役立ったりは面白い結果。