AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Displacement-Resistant Extensions of DPO with Nonconvex $f$-Divergences

Idan Pipano, Shoham Sabach, Kavosh Asadi, Mohammad Ghavamzadeh

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 SquaredPO (non-convex f-DPO with adaptive beta) instead of DPOLlama-3-8B-Instruct · TL;DR validation vs reference (epoch 4)win rate: 34.7 -> 51.0Improved
use SquaredPO instead of DPOLlama-3-8B-Instruct · TL;DR validation vs reference (epoch 2)win rate: 45.0 -> 50.6Improved
use SquaredPO instead of DPOLlama-3-8B-Instruct · TL;DR validation vs reference (epoch 1)win rate: 51.8 -> 50.8No clear effect
use SquaredPO instead of chiPOLlama-3-8B-Instruct · TL;DR validation vs reference (epoch 4)win rate: 48.3 -> 51.0Improved
use SquaredPO instead of DPOLlama-3-8B-Instruct · AlpacaEval 2length-controlled win rate: 29.6 -> 29.2No clear effect
use SquaredPO instead of DPOLlama-3-8B-Instruct · AlpacaEval 2win rate: 24.8 -> 24.5No clear effect
use SquaredPO instead of DPOLlama-3-8B-Instruct · MT-Benchscore: 7.925 -> 7.924No clear effect
use SquaredPO instead of chiPOLlama-3-8B-Instruct · MT-Benchscore: 7.900 -> 7.924Improved
use SquaredPO instead of chiPOLlama-3-8B-Instruct · AlpacaEval 2length-controlled win rate: 29.3 -> 29.2No clear effect
use SquaredPO instead of DPOLlama-3-8B-Instruct + TL;DR · monotone decrease of chosen log-ratios over 4 epochspercent of winners: 99.63 -> 4.21Improved
use SquaredPO instead of DPOQwen2-0.5B-Instruct + UltraFeedback · monotone decrease of chosen log-ratios over 4 epochspercent of winners: 35.57 -> 1.64Improved
use SquaredPO instead of DPOLlama-3-8B-Instruct + TL;DR · decrease from epoch 1 to 2percent of winners: 99.99 -> 10.91Improved

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