🎯 Object Detection¶

💬 ACL2025 · 2 paper notes

Anchored Answers: Unravelling Positional Bias in GPT-2's Multiple-Choice Questions: This work provides the first mechanistic analysis of "anchored bias" (consistently choosing "A") in the GPT-2 family within multiple-choice questions (MCQs) from a failure-case perspective. It localizes specific value vectors storing the "A" preference in MLPs using Logit Lens, and achieves an average MCQ accuracy improvement of 70%+ through minimal intervention (updating the value vectors).
Weed Out, Then Harvest: Dual Low-Rank Adaptation is an Effective Noisy Label Detector for Noise-Robust Learning: This paper proposes the Delora framework, which constructs a noisy label detector by introducing clean LoRA and noisy LoRA modules. By decoupling sample selection from model training, Delora breaks the vicious catch-22 cycle of mutual influence between sample selection and training in traditional "small-loss" approaches.