AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

From Backward Spreading to Forward Replay: Revisiting Target Construction in LLM Parameter Editing

Wei Liu, Liu Hongkai, Zhiying Deng, Yee Whye Teh, Wee Sun Lee

ICML 2026 · Architectures and efficiency › RNNs and normalization · approach: Layer-wise analysis of language models · 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
replace backward spreading with forward replay (FE)MEMIT, Llama3-8B-Instruct · MCFefficacy accuracy: 94.3 -> 99.8Improved
replace backward spreading with forward replay (FE)MEMIT, Llama3-8B-Instruct · ZsREefficacy accuracy: 82.0 -> 93.0Improved
replace backward spreading with forward replay (FE)MEMIT, Llama3-8B-Instruct · ZsRED_KL: 0.60 -> 0.09Improved
add forward replay to AlphaEditAlphaEdit, Llama3-8B-Instruct · MCFefficacy accuracy: 89.0 -> 99.2Improved
add forward replay to PRUNEPRUNE, Llama3-8B-Instruct · ZsREgeneralization accuracy: 72.7 -> 84.4Improved
add forward replay to RECTRECT, Llama3-8B-Instruct · MCFgeneralization accuracy: 46.7 -> 60.4Improved
add forward replay to AlphaEditAlphaEdit, Llama3-8B-Instruct · MCFD_KL: 0.63 -> 0.47Improved
edit multiple layers instead of only the last layer (ablation)MEMIT, Llama3-8B-Instruct · MCFefficacy accuracy: 59.7 -> 94.3Improved
add forward replay instead of backward spreading for remaining residual (ablation)MEMIT, Llama3-8B-Instruct · Layer8 remaining residual ratioremaining residual ratio: 0.36 -> 0.03Improved
remove division of residual by remaining layer count (No-dividing) (ablation)MEMIT backward spreading · Layer4 remaining residual ratioremaining residual ratio: 0.96 -> 1.19Hurt
remove division of residual by remaining layer count (No-dividing) (ablation)MEMIT backward spreading · Layer8 remaining residual ratioremaining residual ratio: 0.36 -> 0.23Improved
use FE instead of BLUE (independent targets)BLUE, Llama3-8B-Instruct · ZsREgeneralization accuracy: 78.2 -> 84.3Improved

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