AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

IRM—when it works and when it doesn't: A test case of natural language inference

Yana Dranker, He He, Yonatan Belinkov

NeurIPS 2021 · Large language models › Language, safety and interpretability · approach: LLM pruning and scaling data · 7 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use IRM instead of ERMIRM vs ERM · toy experiment testaccuracy: 0.0 -> 100.0Improved
use IRM instead of ERM, synthetic bias pe = 0.0IRM vs ERM · SNLI synthetic biasaccuracy: 79.16 -> 83.5Improved
use IRM instead of ERM, synthetic bias pe = 0.33IRM vs ERM · SNLI synthetic biasaccuracy: 85.16 -> 87.22Improved
use IRM instead of ERM, synthetic bias pe = 0.8 (in-distribution)IRM vs ERM · SNLI synthetic biasaccuracy: 93.49 -> 92.32Hurt
use IRM instead of ERM on hypothesis bias, bias misaligned subsetIRM vs ERM · SNLIaccuracy: 62.63 -> 65.12Improved
use IRM instead of ERM on hypothesis bias, bias aligned subsetIRM vs ERM · SNLIaccuracy: 97.4 -> 91.4Hurt
use IRM instead of ERM on hypothesis bias, unbiased subsetIRM vs ERM · SNLIaccuracy: 84.46 -> 82.64Hurt
use IRM instead of ERM on overlap bias, bias misaligned subsetIRM vs ERM · MNLI dev mismatchedF1 macro: 62.66 -> 64.12Improved
use IRM instead of ERM on overlap bias, bias aligned subsetIRM vs ERM · MNLI dev mismatchedF1 macro: 96.67 -> 95.44Hurt
evaluate on HANS and RevisedPremisesIRM vs ERM · natural bias modelsHANS / RP accuracyNo clear effect
increase bias prevalence in training set (ablation)IRM vs ERM · synthetic and natural bias, bias misaligned subsetaccuracyImproved
increase bias strength (ablation)IRM vs ERM · synthetic bias, bias misaligned subsetaccuracyMixed

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy