AI Paper Atlas

Research map / Large language models / Alignment and preferences

Aligning Deep Implicit Preferences by Learning to Reason Defensively

Peiming Li, Zhiyuan Hu, Shiyu Li, Xi Chen, Yang Tang

ICLR 2026 · Large language models › Alignment and preferences · approach: Robust and calibrated DPO

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 Pers-GenPRM critique-and-process reward instead of scalar reward model (ablation)GRPO on Qwen2.5-7B-Instruct · DeepPrefAcc_DA: 70.3 -> 76.3Improved
add critique supervision to the reward model (GRM) (ablation)GRPO · DeepPrefAcc_DA: 70.3 -> 74.7Improved
add process supervision to the reward model (PRM) (ablation)GRPO · DeepPrefAcc_DA: 70.3 -> 73.0Improved
combine process and critique supervision (Pers-GenPRM) (ablation)GRPO · DeepPrefAcc_DA: 73.0 -> 76.3Improved
use Pers-GenPRM reward instead of rubric-based RM (ablation)GRPO · DeepPrefInnovative Expansion: 34.7 -> 42.7Improved
use CDRA instead of GRPO with outcome RMQwen2.5-7B-Instruct · DeepPrefAcc_Mis: 30.7 -> 32.3Mixed
use CDRA instead of GRPOQwen2.5-7B-Instruct · DeepPrefAcc_PF: 83.7 -> 84.7Improved
use CDRA instead of SFTQwen2.5-7B-Instruct · PrefEvalThoughtfulness m_th: 21.4 -> 27.7Mixed
use CDRA instead of CoTQwen2.5-7B-Instruct · PrefEvalThoughtfulness m_th: 54.5 -> 27.7Hurt
use CDRA instead of GRPOQwen2.5-7B-Instruct · ALOE human evalaverage alignment: 3.18 -> 3.92Improved
use CDRA instead of TPOQwen2.5-7B-Instruct · ALOE human evalaverage alignment: 3.86 -> 3.92Improved

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