AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

Influence Dynamics and Stagewise Data Attribution

Jin Hwa Lee, Matthew J. A. Smith, Maxwell Adam, Jesse Hoogland

ICLR 2026 · Architectures and efficiency › RNNs and normalization · approach: Layer-wise analysis of language models · 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
measure influence at checkpoints during training (stagewise BIF) rather than a single final-model influencedeep linear network on hierarchical dataset · Toy hierarchical semantic datasetinfluence sign and magnitude over timeImproved
compare BIF with leave-one-out retraining and analytical influenceBIF in toy model · Toy hierarchical semantic datasetagreement of influence trajectoriesImproved
ablate a sample only during a short window at the BIF peak (ablation)deep linear network (dog sample) · Toy hierarchical semantic datasetintegrated loss differenceImproved
upweight induction patterns when the induction circuit begins to form versus earlier (ablation)small language model trained from scratch · Induction head formationinduction head formation speedImproved
use high localization gamma and low step size in SGLD (ablation)BIF estimator · Toy model BIF-LOO correlationPearson correlation with LOOImproved
use RMSProp-SGLD instead of vanilla SGLD (ablation)BIF/LLC estimator · Deep linear network LLCmatch to ground-truth LLCImproved

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