AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

Extracting alignment data in open models

Federico Barbero, Xiangming Gu, Christopher A. Choquette-Choo, Chawin Sitawarin, Matthew Jagielski, Itay Yona, Petar Veličković, Ilia Shumailov, Jamie Hayes

ICML 2026 · Large language models › Language, safety and interpretability · approach: LLM pruning and scaling data · 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
prefill with full chat template instead of only the BOS token (ablation)OLMo 2 13B extraction · OLMo 2 SFT setmean embedding score: 0.857 -> 0.892Improved
measure memorisation with neural embeddings instead of string matchingextraction evaluation · OLMo 2 13B generationsmemorisation rateImproved
SFT on extracted synthetic data instead of original dataOLMo 2 7B · GSM8Kscore: 0.7377 -> 0.7847Improved
SFT on extracted synthetic data instead of original dataOLMo 2 7B · BBHscore: 0.4944 -> 0.5161Improved
SFT on extracted synthetic data instead of original dataOLMo 2 7B · TQAscore: 0.5110 -> 0.5529Improved
SFT on extracted synthetic data instead of original dataOLMo 2 7B · IFEscore: 0.6617 -> 0.5028Hurt
SFT on extracted synthetic data instead of original dataOLMo 2 7B · MATHscore: 0.2077 -> 0.1705Hurt
SFT on extracted synthetic data instead of original dataOLMo 2 7B · MMLUscore: 0.6123 -> 0.6052No clear effect
RL on extracted synthetic dataset instead of ORZ 57k datasetQwen2.5 7B Dr. GRPO · AMC, MATH500, Minerva, OlympiadBenchaccuracyNo clear effect
RL post-training (Open-Reasoner-Zero) raises likelihood of training promptsQwen2.5 7B · ORZ train prompts (142,770)count above likelihood threshold: 18 -> 287Improved

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