AI Paper Atlas

Research map / Large language models / Alignment and preferences

Configurable Reward Model for Balanced Safety Alignment

Zhengping Jiang, Mehran Khodabandeh, Akash Bharadwaj, Manik Bhandari, Mayur Srungarapu, Anqi Liu, Benjamin Van Durme, Li Chen

ICML 2026 · Large language models › Alignment and preferences · approach: Steering and correcting LLM behavior · 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
add configurable safety configuration augmentation (ablation)CSRM · CoSApienF1: 0.815 -> 0.946Improved
add configurable safety configuration augmentation (ablation)CSRM · DynaBenchF1: 0.692 -> 0.758Improved
add configurable safety configuration augmentation (ablation)CSRM · BeaverTailssmECE: 0.049 -> 0.017Improved
add configurable safety configuration augmentation (ablation)CSRM · WildGuardMixF1: 0.953 -> 0.953No clear effect
add severity augmentation (ablation)CSRM (-CCA) · BeaverTails-Aug pairwiseaccuracy: 0.583 -> 0.762Improved
add severity augmentation (ablation)CSRM (-CCA) · CoSApiensmECE: 0.141 -> 0.123Improved
add severity augmentation (ablation)CSRM (-CCA) · Safety-RLHF pairwiseaccuracy: 0.661 -> 0.763Improved
use CSRM as reward instead of Llama-3.1-8B-InstReinforce++ alignment · Arab PublisherCoSA score: 3.475 -> 3.603Improved
use CSRM as reward instead of LlamaGuard-3-8BReward Distillation · Arab PublisherCoSA score: 2.825 -> 3.145Improved
increase reward-loss margin from 0.0 to 1.0 (ablation)CSRM (Qwen3-2B) · CoSApienF1: 0.890 -> 0.757Hurt
increase reward-loss margin from 0.0 to 1.0 (ablation)CSRM (Qwen3-2B) · BeaverTailssmECE: 0.055 -> 0.070Hurt
replace strict policy with Allow policy for target category (ablation)CSRM · BeaverTails single-categoryunsafe rate: 0.843 -> lowered by 0.708Improved

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