AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

KL-Regularized Reinforcement Learning is Designed to Mode Collapse

Anthony GX-Chen, Jatin Prakash, Jeff Guo, Rob Fergus, Rajesh Ranganath

ICLR 2026 · Reinforcement learning › RL from feedback · approach: Reward modeling versus direct optimization · 27 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 MARA reward augmentation (reverse KL) to RLOO-style trainingRLOO · Creative QA, out-of-dist rewardOut-dist. Reward: 1.280 -> 1.451Improved
use forward KL with MARA instead of reverse KL (ablation)MARA · Creative QA, Mean DistinctMean Distinct: 4.14 -> 4.62Improved
add MARA to RLOORLOO · Creative QA, Semantic DivSemantic Div: 0.192 -> 0.186Hurt
add MARA reward augmentation to REINVENTREINVENT · SYNTH, threshold 0.85OB100: 4114 -> 3654Improved
add MARA reward augmentation to REINVENTREINVENT · ALL-AMIDE, threshold 0.85Yield: 1098 -> 1235Improved
add MARA reward augmentation to REINVENTREINVENT · SYNTH, threshold 0.80Circles: 67 -> 59Hurt
add MARA reward augmentation to REINVENTREINVENT · SYNTH/ALL-AMIDEIntDiv1: 0.766 -> 0.761No clear effect
add MARA mode-anchored reward augmentationQwen2.5-3B with naive KL regularization · Verifiable 1-2 taskanswer diversity (entropy)Improved
add MARA to reverse and forward KLQwen2.5-3B · Verifiable 1-2 taskcorrectness vs diversity Pareto frontImproved
lower KL regularization strength beta (ablation)KL-regularized RL · didactic two-mode examplemode coverageHurt
lower KL regularization strength beta with equal rewards (ablation)KL-regularized RL · didactic equal-reward exampleprobability of low-support correct modeNo clear effect
switch from reverse KL to forward KL regularization (ablation)KL-regularized RL · didactic two-mode examplemode coverageNo 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