Reproduce “Learning to Deceive with Attention-Based Explanations” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Danish Pruthi, Mansi Gupta, Bhuwan Dhingra, Graham Neubig, Zachary C. Lipton

The authors train a predictor to reduce attention on predefined tokens while preserving its task predictions. The penalty does not remove those tokens or prevent their information from influencing the model. In occupation classification, a BiLSTM’s test accuracy stays at 96.4% as attention mass on gender-indicator tokens falls from 50.3% to 0.08%, showing suppression is not removal.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.