AI Paper Atlas

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

Are Emergent Abilities of Large Language Models a Mirage?

Rylan Schaeffer, Brando Miranda, Sanmi Koyejo

NeurIPS 2023 · Large language models › Language, safety and interpretability · approach: Reasoning mechanisms and failure modes

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
Replace Accuracy with Token Edit Distance while keeping model outputs fixedInstructGPT/GPT-3 · 2-shot 2-digit multiplication and 2-shot 4-digit additionperformance versus model scaleImproved
Increase test-data size when measuring AccuracyInstructGPT/GPT-3 · 2-shot 2-digit multiplication and 2-shot 4-digit additionaccuracyImproved
Measure longer target strings with Token Edit DistanceInstructGPT/GPT-3 · arithmetic tasks, target strings of length 1 to 5Token Edit DistanceHurt
Measure longer target strings with AccuracyInstructGPT/GPT-3 · arithmetic tasksaccuracyHurt
Replace Multiple Choice Grade with Brier ScoreLaMDA · BIG-Bench tasks where emergence was reportedemergence under the evaluation metricHurt
Use discontinuous Reconstruction_c instead of squared reconstruction errorShallow nonlinear autoencoders · CIFAR100 natural-image reconstructionapparent emergence with scaleImproved
Score whether all images in a sequence are classified correctlyAutoregressive Transformers · Omniglot classificationsubset accuracyImproved
Score whether all K test examples are classified correctly instead of using test accuracyLeNet · MNISTsubset accuracyImproved

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