AI Paper Atlas

Research map / Probabilistic and scientific modeling / Molecules and proteins

Mitigating Reward Hacking in RLHF via Bayesian Non-negative Reward Modeling

Zhibin Duan, Guowei Rong, Zhuo Li, Bo Chen, Mingyuan Zhou, Dandan Guo

ICML 2026 · Probabilistic and scientific modeling › Molecules and proteins · approach: Probabilistic embeddings and relational inference · 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
add Bayesian non-negative factor reward model (BNRM) on top of BTBT reward model (Gemma-2B-it, 40K UF) · Unified-Feedback valaccuracy: 68.8 -> 74.2Improved
add BNRM on top of GRM-SFTGRM-SFT (Gemma-2B-it, 40K UF) · HHH-Alignmentaccuracy: 78.7 -> 82.4Improved
add BNRM on top of BTBT reward model · RewardBench Safety (Gemma-2B-it, 40K UF)accuracy: 95.8 -> 95.6Mixed
add BNRM to the already strong Skywork reward modelSkywork-Reward-Llama-3.1-8B · RewardBench averagescore: 93.1 -> 93.6Improved
use BNRM reward model for PPOLlama3.1-8B-Instruct policy · 8-benchmark average accuracyaccuracy: 62.83 -> 74.98Improved
use BNRM reward model for PPOOpenRLHF-Llama3-8B-SFT policy · 8-benchmark average accuracyaccuracy: 55.68 -> 62.25Improved
use BNRM reward model for PPOOpenRLHF-Llama3-8B-SFT policy · Hellaswagaccuracy: 72.51 -> 60.68Hurt
use BNRM reward model for PPOLlama3.1-8B-Instruct policy · GSM8K 4-shot, Hellaswag, MMLUaccuracy: 83.93 -> 82.49 (GSM8K), 77.21 -> 74.66 (Hellaswag), 72.31 -> 70.72 (MMLU)Hurt
use BNRM-aligned policy versus the base policyLlama-3.1-8B-Instruct · Arena-Hard-v0.1 (GPT-4.1 judge)win rate: 0.4989 win vs 0.2239 loseImproved
replace BT with BNRM in low-resource trainingBT (Gemma-2B-it) · RewardBench, 1K UF samplesaccuracyImproved
replace BT with BNRM under label noiseBT (Gemma-2B-it, 40K UF) · noise ratio 0.4, OOD benchmarksaccuracy: up to +16.7Improved
use BNRM instead of BT for length debiasingBT (Gemma-2B-it) · RM-Bench Hard, length-reward Pearson rPearson correlation: 0.488 -> 0.123Improved

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