AI Paper Atlas

Research map / Label-efficient and robust learning / Test-time adaptation

Temporal Generalization: A Reality Check

Divyam Madaan, Sumit Chopra, Kyunghyun Cho

ICLR 2026 · Label-efficient and robust learning › Test-time adaptation · approach: Characterizing out-of-distribution failures

Read the paper

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
model averaging of past checkpoints instead of the most recent modelrecent model · temporal benchmarks (NewsRoom, Wilds-Time)perplexity / accuracyHurt
downscale recent parameters by a scalar alpha < 1recent model · NewsRoom language modeling and summarizationperplexityImproved
downscale recent parameters by a scalar alpha < 1recent model · Wilds-Time datasetsaccuracyHurt
Taylor-series parameter extrapolationrecent model · temporal benchmarksforward transferHurt
initialize each time step from the previous one (continual learning) instead of independent training from pretrained (ablation)interpolation and extrapolation methods · NewsRoom, T5-smallforward transferImproved
learn parameter offset for Taylor extrapolationTaylor extrapolation · NewsRoom language modeling, 12 monthsperplexity: 6.05 -> 6.09No clear effect
add learned time-varying coefficient to Taylor extrapolationTaylor extrapolation · NewsRoom language modeling, 12 monthsperplexity: 6.09 -> 6.07No clear effect
tune extrapolation factor alpha with future validation data (ablation)Taylor extrapolation · NewsRoom summarizationoptimal alphaNo 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