AI Paper Atlas

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

CLEVRER-Humans: Describing Physical and Causal Events the Human Way

Jiayuan Mao, Xuelin Yang, Xikun Zhang, Noah Goodman, Jiajun Wu

NeurIPS 2022 · Large language models › Language, safety and interpretability · approach: Data and evaluation methodology · 8 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
add rule-based pairwise event detector before event generation (ablation)pair description generator · stage III human labels, 100 videospairwise description accuracy: 80.6% to 87.2%Improved
pretrain on CLEVRER then test on CLEVRER-HumansALOE · CLEVRER-Humans testper-option accuracy: 98.5 -> 54.0Hurt
finetune pretrained model on CLEVRER-Humans (ablation)ALOE · CLEVRER-Humans testper-option accuracy: 54.0 -> 52.7No clear effect
train from scratch vs pretrain+finetune (ablation)ALOE · CLEVRER-Humans testper-question accuracy: 31.7 -> 32.1No clear effect
replace LSTM with pretrained BERTCNN+LSTM · CLEVRER-Humans testper-option accuracy: 50.3 -> 52.9No clear effect
train CNN+LSTM from scratch on CLEVRER-Humans (ablation)CNN+LSTM · CLEVRER-Humans testper-question accuracy: 30.0 -> 34.2Mixed
use CLEVRER heuristic instead of counterfactual labels to approximate humanscausal labels · CLEVRER sampleagreement with humanImproved
train longer on small training set (ablation)ALOE · CLEVRER-Humans testper-option accuracy: 70% train; 53.7% testNo clear effect

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