AI Paper Atlas

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

Winter Soldier: Backdooring Language Models at Pre-Training with Indirect Data Poisoning

Wassim Bouaziz, Mathurin VIDEAU, Nicolas Usunier, El-Mahdi El-Mhamdi

ICLR 2026 · Large language models › Instruction and fine-tuning · approach: Safety alignment after fine-tuning · 4 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 gradient-matching crafted poisons instead of pairwise token backdoor (ablation)indirect data poisoning · 360M, contamination 0.003%p-value: 10^{-4} at best -> 10^{-14}Improved
scale up the model being poisoned from smaller to 1.4B (ablation)indirect data poisoning · Detection p-value across model sizesp-value: as low as 10^{-55} for 1.4BImproved
use a longer secret response (ablation)indirect data poisoning · |y| in {1,5,10}detection guaranteeMixed
craft poisons on a larger model than Bob's (ablation)indirect data poisoning · Bob 135M, top-10p-value: 8.13e-4 -> 3.37e-11 (135M -> 1.4B crafting)Improved
add poisoned data (secret responses never in training set) to pretrainingLM pretraining · LM benchmarks (ARC, Hellaswag, MMLU, PIQA...)accuracy: e.g. 1.4B PIQA 69.8 -> 68.2No clear effect
apply a quality classifier filter to the datasetpoisoned data defense · Defensepoison filteredHurt
lower contamination ratio to 0.001% (ablation)indirect data poisoning · top-20 secret accuracytop-20 accuracyNo 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