AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Semantic-aware Wasserstein Policy Regularization for Large Language Model Alignment

Byeonghu Na, Hyungho Na, Yeongmin Kim, Suhyeon Jo, HeeSun Bae, Mina Kang, Il-chul Moon

ICLR 2026 · Reinforcement learning › RL from feedback · approach: Reward modeling versus direct optimization · 3 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 reverse KL with Wasserstein policy regularizationPPO · TL;DRwin rate vs SFT: 0.848 -> 0.924Improved
replace reverse KL with Wasserstein policy regularizationPPO · HH-RLHFwin rate vs SFT: 0.828 -> 0.852Improved
replace reverse KL with Wasserstein policy regularizationPPO · MT-Benchscore: 4.000 -> 4.272Improved
replace reverse KL with Wasserstein policy regularizationCodeGemma-7B · APPS (all)pass@1: 8.32 -> 8.79Improved
use Wasserstein penalty instead of forward KLPPO · TL;DRwin rate vs RKL: 0.040 -> 0.608Improved
use Wasserstein penalty instead of chi-squaredPPO · TL;DRwin rate vs RKL: 0.540 -> 0.608Improved
scale policy to Gemma-7B with Gemma-7B embedding for cost matrixWasserstein PPO · TL;DRwin rate vs RKL-2B: 0.608 -> 0.712Improved
use Gemma-7B embeddings instead of Gemma-2B for cost matrix (2B policy) (ablation)Wasserstein PPO · TL;DRwin rate vs SFT: 0.924 -> 0.908No clear effect
reduce Sinkhorn iterations from 10 to 5 (ablation)Wasserstein PPO · TL;DRwin rateHurt
increase Sinkhorn iterations to 30 (ablation)Wasserstein PPO · TL;DRwin rateNo clear effect
decrease truncation parameters k1,k2 or entropy strength lambda (ablation)Wasserstein PPO · TL;DRwin rateHurt
change cost function from Euclidean to cosine distance (ablation)Wasserstein PPO · TL;DRwin rateNo 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