AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

General Exploratory Bonus for Optimistic Exploration in RLHF

Wendi Li, Changdae Oh, Sharon Li

ICLR 2026 · Reinforcement learning › RL from feedback · approach: Theory of RLHF · 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
add GEB-pi exploratory bonus to f-DPOiterative DPO, Mistral-Instruct-v0.3 · UltraFeedback KLwin rate: 78.42 -> 81.00Improved
add GEB-arctanh bonus to f-DPOiterative DPO, Mistral-Instruct-v0.3 · UltraFeedback Hellingerwin rate: 72.69 -> 75.69Improved
add GEB-arctanh bonus to f-DPOiterative DPO, LLaMA-3-8B-SFT · UltraFeedback Hellingerwin rate: 71.11 -> 73.26Improved
use FEB (prior-style bonus) instead of f-DPOiterative DPO, Mistral-Instruct-v0.3 · UltraFeedback forward KLwin rate: 51.11 -> 47.53Hurt
add GEB bonus (average of divergences)f-DPO, LLaMA-3-8B-SFT · AlpacaEval2 LCwin rate: 22.75 -> 24.73Improved
add GEB bonus (average of divergences)f-DPO, LLaMA-3-8B-SFT · MATH-500accuracy: 68.6 -> 70.1Improved
add GEB-1/pi bonus to DPO on hard promptsDPO, Qwen2.5-7B · OlympiadBenchPass@16: 57.78 -> 65.78Improved
add GEB-1/pi bonus to DPO on hard promptsDPO, Qwen2.5-7B · AIME 2025Pass@16: 23.23 -> 29.48Improved
apply GEB-pi bonus to rejected responses only instead of all responses (ablation)GEB-pi, Mistral-Instruct-v0.3 · UltraFeedback KLwin rate: 79.78 -> 81.00Improved
apply GEB-pi bonus to rejected responses only instead of all responses (ablation)GEB-pi, Mistral-Instruct-v0.3 · UltraFeedback Hellingerwin rate: 73.40 -> 75.48Improved
add GEB bonus for distinct-n diversity (dist-4)f-DPO · sampled corpus under KLdist-4: 0.8418 -> 0.8448No clear effect
make bonus ratio too large (ablation)GEB · kappa sweepperformanceHurt

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