AI Paper Atlas

Research map / Large language models / Instruction and fine-tuning

Rote Learning Considered Useful: Generalizing over Memorized Data in LLMs

Qinyuan Wu, Soumi Das, Mahsa Amani, Bishwamittra Ghosh, Mohammad Aflah Khan, Krishna P. Gummadi, Muhammad Bilal Zafar

ICLR 2026 · Large language models › Instruction and fine-tuning · approach: Fine-tuning generalization and delta compression

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
increase rote memorization epochs in Phase-1 (3 to 20) (ablation)memorize-then-generalize · Synthetic author facts, unseen promptstest prompt accuracy: 0.35 -> 0.98Improved
add Phase-2 fine-tuning with one meaningful prompt on 50 memorized factsrote-memorized Qwen2.5-1.5B · Synthetic facts, unseen factsgeneration accuracy: 0.36 -> 0.76Improved
add Phase-2 fine-tuning on held-out facts probabilityrote-memorized Qwen2.5-1.5B · Synthetic facts, train prompt on held-out factsobject probability: 0.18 -> 0.79Improved
use a single memorized fact with a single prompt in Phase-2 (ablation)memorize-then-generalize · Synthetic author facts, test promptsaccuracy: 0.76 (k=1)Improved
rote memorization of the first relation before learning the second (ablation)memorize-then-generalize · 2-hop composition A->Cgeneration accuracy: 0.14 -> 0.36Improved
5 to 15 epochs of rote memorization of R1 instead of 20 (ablation)memorize-then-generalize · 2-hop composition A->Cgeneration accuracy: 0.14 -> 0.14No clear effect
add rote memorization before reversal trainingmemorize-then-generalize · Synthetic mother relation, reversalaccuracy: 0 -> 0.26Improved
use memorize-then-generalize instead of SFT on reversalQwen2.5-1.5B · Synthetic mother relation, reversalaccuracy: 0.01 -> 0.26Improved
use memorize-then-generalize instead of SFT at equal training tokensQwen2.5-1.5B · Synthetic facts, one training promptaccuracyImproved
use memorize-then-generalize instead of SFT with ten training promptsQwen2.5-1.5B · Synthetic facts, ten training promptsaccuracy and training tokens: both ~0.9 accuracy, half the tokensMixed
memorize-then-generalize in place of ICL for knowledge injectionQwen2.5-1.5B · Synthetic factsobject probabilityImproved
memorize subject-object pairs without the key token (no semantic anchor) (ablation)memorize-then-generalize · Synthetic factsgeneralizationHurt

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