AI Paper Atlas

Research map / Large language models / Alignment and preferences

When Human Preferences Flip: An Instance-Dependent Robust Loss for RLHF

Yifan Xu, Xichen Ye, Yifan Chen, Qiaosheng Zhang

AAAI 2026 · Large language models › Alignment and preferences · approach: Robust and calibrated DPO · 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 warmup of the flipping model (20/20 batches) (ablation)FA-DPO · HH_Golden 20% flipaccuracy: 84.96 -> 98.56Improved
add warmup (20/20 batches) (ablation)FA-DPO · HH_Golden 20% flipwin rate: 57.35 -> 70.04Improved
add warmup (50/50 batches) (ablation)FA-DPO · HH_Golden 20% flipaccuracy: 83.16 -> 96.80Improved
more policy steps (20 -> 50) with warmup (ablation)FA-DPO · HH_Golden 20% flipwin rate: 70.04 -> 82.90Improved
more policy steps (20 -> 50) without warmup (ablation)FA-DPO · HH_Golden 20% flipaccuracy: 84.96 -> 77.00Hurt
more policy steps (20 -> 50) with warmup (ablation)FA-DPO · HH_Golden 20% flipaccuracy: 98.56 -> 98.40No clear effect
replace DPO with FA-DPOPythia-1B · UltraFeedback 40% flipaccuracy: 51.87 -> 70.77Improved
replace DPO with FA-DPOPythia-1B · HH_Golden 40% flipaccuracy: 58.63 -> 98.83Improved
replace DPO with FA-DPOPythia-1B · UltraFeedback clean (0%)win rate: 67.00 -> 66.85Hurt
replace DPO with FA-DPOLlama-3.1-8B · UltraFeedback 40% flipaccuracy: 64.96 -> 78.87Improved
replace DPO with FA-DPOMistral-7B · UltraFeedback 40% flipwin rate: 35.35 -> 59.00Improved
replace ROPO with FA-DPO (clean data)Pythia-1B · UltraFeedback clean (0%)accuracy: 70.75 -> 73.05Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

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