AI Paper Atlas

Research map / Reinforcement learning / Offline RL

Causal Flow Q-Learning for Robust Offline Reinforcement Learning

Mingxuan Li, Junzhe Zhang, Elias Bareinboim

ICML 2026 · Reinforcement learning › Offline RL · approach: Simple practical offline RL · 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 confounding-robust worst-case Q target with discriminator and Q-ensemble minimum to FQLFQL · OGBench visual tasks, offline (25 tasks)normalized mean success: 1.0 -> 1.21Improved
add confounding-robust objective to FQL on cube-singleFQL · visual-cube-single-play (5 tasks)success rate (%): 65 -> 81Improved
add confounding-robust objective to FQL on cube-doubleFQL · visual-cube-double-play (5 tasks)success rate (%): 6 -> 11Improved
add confounding-robust objective to FQL on puzzle-3x3 task5FQL · visual-puzzle-3x3-play task5success rate (%): 1 -> 19Improved
add confounding-robust objective to FQL on scene task5FQL · visual-scene-play task5success rate (%): 0 -> 0No clear effect
set discriminator loss coefficient too small (ablation)Causal-FQL · visual-cube-double task1 (coefficient sweep {1, 5, 10, 15})success rateHurt
use more Q-network ensembles (2, 4, 6) (ablation)Causal-FQL · visual-cube-double task1success rateHurt
use Causal-FQL for offline-to-online fine-tuningFQL · OGBench visual tasks, offline-to-onlinesuccess rate: 8% improvement on averageImproved
switch back to vanilla FQL objective during online phase (cube tasks) (ablation)Causal-FQL · visual-cube-single and visual-cube-double, offline-to-onlinesuccess rateImproved
balanced sampling of online and offline buffers with matching objectives (puzzle-4x4) (ablation)Causal-FQL · visual-puzzle-4x4, offline-to-onlinesuccess rateImproved
use pixel observations instead of structured state observationsSOTA flow policy offline RL · OGBench, 5 tasks per categorynormalized success rate: nearly 80% drop in someHurt

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