AI Paper Atlas

Research map / Large language models / Alignment and preferences

One Bias After Another: Mechanistic Reward Shaping and Persistent Biases in Language Reward Models

Daniel Fein, Max Lamparth, Violet Xiang, Mykel Kochenderfer, Nick Haber

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
null the length probe direction from RM activationsDeBERTa RM · RewardBench2 reward-length correlationSpearman rs: 0.611 -> 0.067Improved
null the length probe in a RM that overcorrects against verbositySkywork-Llama-8B · RewardBench2 reward-length correlationSpearman rs: 0.267 -> 0.305Mixed
apply length probe nulling for best-of-N selectionfive RMs · held-out AlpacaEval, 64 candidatesmean absolute within-prompt reward-length correlation: 0.10 -> 0.04Improved
apply length probe nulling for best-of-N selectionfive RMs · held-out GSM8K, 64 generationsmean best-of-N accuracy: 62.1% -> 62.8%Improved
null the uncertainty probe (alpha=1)Skywork-Qwen-8B · verbalized confidence calibrationSpearman correlation: 0.182 -> 0.386Improved
null the confidence probes at alpha=1.0 (ablation)Skywork-Llama-8B · verbalized confidence calibrationSpearman correlation: 0.350 -> 0.129Hurt
use weaker projection alpha=0.5 instead of 1.0 (ablation)AllenAI-Llama-8B · verbalized confidence calibrationSpearman correlation: 0.271 -> 0.332Improved
increase projection strength to alpha=1.5 (ablation)Skywork-Qwen-8B · verbalized confidence calibrationSpearman correlation: 0.386 -> 0.008Hurt
null the uncertainty probeall five RMs · uncertainty penalty on correct answers with hedgingaccuracyImproved
apply length, position and uncertainty probesRMs · RewardBench2mean accuracy (%): 70.1 -> 69.3No clear effect
linear probe intervention for user agreementRMs · sycophancyfaulty-agreement vs helpful-agreementNo clear effect
null the position probethree of five RMs · MCQA and free-form position biaspositional varianceImproved

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