AI Paper Atlas

Research map / Large language models / Alignment and preferences

When Distance Distracts: Representation Distance Bias in BT-Loss for Reward Models

Tong Xie, Andrew Bai, Yuanhao Ban, Yunqi Hong, Haoyu Li, Cho-Jui Hsieh

ICML 2026 · Large language models › Alignment and preferences · approach: Reward model design and preference models · 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
replace BT loss with NormBTgemma-2b-it, Unified-Feedback · RewardBench Reasoningaccuracy: 75.41 -> 80.71Improved
replace BT loss with NormBTLlama-3.2-3B-Instruct, Unified-Feedback · RewardBench Reasoningaccuracy: 71.70 -> 76.93Improved
replace BT loss with NormBTLlama-3.2-3B-Instruct, Skywork-80K · RewardBench Reasoningaccuracy: 67.05 -> 74.60Improved
replace BT loss with NormBTgemma-2b-it, Unified-Feedback · RewardBench Averageaccuracy: 72.25 -> 73.57Improved
replace BT loss with NormBTLlama-3.2-3B-Instruct, Skywork-80K · RewardBench Averageaccuracy: 80.31 -> 81.48Improved
replace BT loss with NormBTLlama-3.2-3B-Instruct, Skywork-80K · RewardBench Chataccuracy: 86.03 -> 83.80Hurt
add ground-truth margin to BT loss (ablation)gemma-2b-it, Unified-Feedback · RewardBench Averageaccuracy: 72.25 -> 71.23Hurt
add label smoothing to BT loss (ablation)gemma-2b-it, Unified-Feedback · RewardBench Reasoningaccuracy: 75.41 -> 72.28Hurt
remove EMA normalization from NormBT (ablation)gemma-2b-it, Unified-Feedback · RewardBench Averageaccuracy: 67.78 -> 73.57Improved
use last-token L2 distance instead of cosine similarity (ablation)NormBT · RewardBench Averageaccuracy: 71.88 -> 73.57Improved
use last-token L2 distance instead of average pooling (ablation)NormBT · RewardBench Averageaccuracy: 69.68 -> 73.57Improved
add LayerNorm before reward head to BT (ablation)gemma-2b-it, Unified-Feedback · RewardBench Averageaccuracy: 72.25 -> 72.25No 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