AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Safe Reinforcement Learning with Preference-based Constraint Inference

Chenglin Li, Grant Ruan, Hua Geng

ICML 2026 · Reinforcement learning › RL from feedback · approach: Theory of RLHF · 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 dead zone (delta=1) to the safety loss (ablation)PbCRL cost model · Safety Gymnasium GoalW2 distance to true cost / ground-truth cost: W2 38.9 -> 16.8; cost 28.05 -> 18.50Improved
Increase dead zone from 1 to 2 (ablation)PbCRL · Safety Gymnasium GoalW2 distance / return: W2 16.8 -> 29.1; return 23.41 -> 22.51Hurt
Small dead zone delta=0.1 instead of 0 (ablation)PbCRL · Safety Gymnasium GoalW2 distance: 38.9 -> 33.5No clear effect
Lower the training safety threshold of PPO-BT (d=10) (ablation)PPO-BT · Safety Gymnasium Goalreturn / cost: cost 26.97 -> 14.02; return 24.26 -> 18.27Mixed
Add SNR loss with weight 1e-3 (ablation)PbCRL · Safety Gymnasium Goalmid-training cost: 24.8 -> 18.4Improved
Raise SNR loss weight to 1e-2 (ablation)PbCRL · Safety Gymnasium Goalend return: 23.4 -> 6.5Hurt
Add online finetuning with adaptive dead zone after offline pre-training (ablation)PbCRL-Offline · Safety Gymnasium Goalground-truth cost: 22.70 -> 18.50Improved
Add online finetuning with adaptive dead zone after offline pre-training (ablation)PbCRL-Offline · Safety Gymnasium Pushground-truth cost: 37.48 -> 34.86Improved
Use PbCRL instead of standard BT cost modelPPO-BT · Safety Gymnasium Buttonground-truth cost: 50.78 -> 37.51Improved
Use PbCRL instead of segment-level binary cost (RLSF)RLSF · Safety Gymnasium Buttonreturn: -2.01 -> 8.73Improved
Use PbCRL instead of Safe RLHFSafe RLHF · Safe RLHF dataset, Llama-3.2-1Bwin rate (harmless): 76.5% -> 82.1%Improved
Use PbCRL instead of PPO-BTPPO-BT · Autonomous driving Blockcost: 0.12 ± 0.02 -> 0.07 ± 0.01Improved

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