AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Learning What Matters Now: Dynamic Preference Inference under Contextual Shifts

Xianwei Cao, Dou Quan, Zhenliang Zhang, Shuang Wang

ICLR 2026 · Reinforcement learning › RL from feedback · approach: Reward modeling versus direct optimization · 1 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
use preference-conditioned PPO instead of Q-learning (ablation)DPI · Queuemean episodic return: 3.74 -> 10.34Improved
use preference-conditioned PPO instead of Q-learning (ablation)DPI · Mazesuccess rate (%): 42.94 -> 59.04Improved
add KL prior term to the ELBO (ablation)DPI · Queue post-shiftPS@KImproved
add directional alignment loss (ablation)DPI · Queue post-shiftPS@KImproved
add self-consistency (stability) regularizer (ablation)DPI · Queue post-shiftPS@KImproved
use history window H=1 instead of larger windows (ablation)DPI · Mazemean episodic returnHurt
increase history window H beyond 9 (ablation)DPI · Mazemean episodic returnNo clear effect
replace ENVELOPE with DPIENVELOPE · Queuesuccess rate (%): 25.10 -> 39.95Improved
replace ENVELOPE with DPIENVELOPE · Mazemean episodic return: 10.36 -> 30.16Improved
replace FIXED with DPIFIXED · Mazesuccess rate (%): 1.12 -> 59.04Improved
replace MER-PPO with DPIMER-PPO · Mazesuccess rate (%): 0.05 -> 59.04Improved
replace MER-PPO with DPIMER-PPO · Mazemean episodic return: 85.55 -> 30.16Hurt

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