AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

Step-Resolved Data Attribution for Looped Transformers

Georgios Kaissis, David Mildenberger, Juan Felipe Gomez, Martin J. Menten, Eleni Triantafillou

ICML 2026 · Architectures and efficiency › RNNs and normalization · approach: Brain-inspired recurrent architectures · 0 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 sketch-during-backprop TensorSketch instead of materialising per-example gradientsSDI on looped GPT 135.1M · Single 48GB A6000, m=2048feasible batch size: 4 -> 40Improved
increase sketch dimension m (ablation)SDI TensorSketch · Looped GPT 135.1Msketch error scaling: log-log slope -0.489Improved
scale model width (77M to 1.2B parameters) (ablation)SDI relative sketch cost · Fixed tau=32relative overhead: +116% -> +18%Improved
use SDI step-resolved trajectory instead of scalar TracIn to find a periodic latent circuitLooped transformer on parity · Parity, length-40 probemechanistic proxy accuracy: 100% on held-out alternating inputs; ≈93% on length-40 testImproved
use step-resolved SDI rather than scalar TracIn to expose signal cancellationParity pair (20-bit train, 40-bit test) · Parityinfluence magnitude: TracIn 0.001 vs sum |SDI| 17.24Improved
remove top 30% late-step SDI training puzzles (ablation)Looped transformer on Sudoku · Sudoku hard puzzles, tau=32board accuracy: 95.8 -> 91.7Hurt
remove top 30% late-step SDI training puzzles (ablation)Looped transformer on Sudoku · Sudoku overall, tau=32board accuracy: 98.6 -> 97.8Hurt
harder training puzzles (more blanks)Sudoku TracIn self-influence · Sudoku trainmedian self TracIn: 0.225 -> 0.451Improved
hard vs easy training puzzles: share of SDI energy in late stepsSDI on Sudoku · Sudoku hard test puzzlesSDI late mass (steps 17-32): 24.3% -> 25.3%Improved
add more test-time loops (ablation)Looped transformer on Sudoku · Sudoku hard puzzlesboard accuracyImproved
add AdamW second-moment preconditioning to the gradient inner product (ablation)SDI/TracIn on Sudoku · Sudoku ranking stabilitySpearman rank correlation: rho 0.979 TracIn, 0.934 SDINo clear effect
recompute SDI with full BPTT instead of truncated BPTT (ablation)Recursive Nanochat SFT · GSM8K train to testSDI last-step share: 0.604 at tau=8, 0.606 at tau=16Improved

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