Research map / Large language models / Alignment and preferences
Unbiased Reward Modeling from Implicit Feedback for LLM Alignment
Hao Wang, Haocheng Yang, Licheng Pan, Zhichao Chen, Xiaoxi Li, Yinuo Wang, Lei Shen, Yuan Lu, Haoxuan Li, Zhouchen Lin
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 changed | Where | Measured | Effect |
|---|---|---|---|
| model unlabeled samples as possibly positive (latent positives) rather than all negative (ablation) | ImplicitRM · PKU-SafeRLHF | R2: 0.6848 -> 0.7872 | Improved |
| adaptively estimate action propensities (unfreeze propensity estimator) (ablation) | ImplicitRM · HelpSteer | R2: 0.2186 -> 0.3114 | Improved |
| handle both false negatives and selection bias together (ablation) | ImplicitRM · UltraFeedback | R2: 0.4457 / 0.3948 -> 0.5207 | Improved |
| use ImplicitRM objective instead of Naive (unlabeled as negative) | reward model on frozen LLM backbones · PKU-SafeRLHF, Qwen3-8B | R2: 0.4786 -> 0.7546 | Improved |
| scale backbone from 8B to 32B (Qwen3) (ablation) | ImplicitRM relative gain over Naive · PKU-SafeRLHF | R2 relative gain: 57.7% -> 40.0% -> 31.5% relative gain | Mixed |
| use smaller learning rate (5e-6) (ablation) | ImplicitRM · HelpSteer | R2 | Improved |
| increase batch size from 64 to 256 (ablation) | ImplicitRM · HelpSteer | R2 | Improved |
| increase action propensity alpha from 0.1 to 0.9 (ablation) | all methods · UltraFeedback | R2 gap between Naive and ImplicitRM: gap 0.65 -> 0.10 | Improved |
| train policy with ImplicitRM reward vs Naive reward in GRPO | Qwen2.5-Instruct-7B · HarmBench | Safety score: 0.8381 -> 0.9258 | Improved |
| train policy with ImplicitRM reward vs Naive reward in GRPO | Qwen3-Instruct-4B · WildGuardMix | Safety score: 0.8094 -> 0.8974 | Improved |
| use debiased (propensity-based) baselines vs Naive | SDR on implicit feedback · UltraFeedback | R2: 0.1459 -> 0.3646 | Improved |
| ImplicitRM vs best baseline SelectMix on MAE | HelpSteer · HelpSteer | MAE: 0.2624 (SelectMix) vs 0.2919 (ImplicitRM) | Hurt |
Closest papers by meaning
- Post-hoc Reward Calibration: A Case Study on Length Bias (ICLR 2025 · similarity 0.91)
- Reward Learning from Multiple Feedback Types (ICLR 2025 · similarity 0.90)
- Data with High and Consistent Preference Difference Are Better for Reward Model (AAAI 2025 · similarity 0.89)
- Active Reward Modeling: Adaptive Preference Labeling for Large Language Model Alignment (ICML 2025 · similarity 0.89)
- Scaling Reward Modeling without Human Supervision (NeurIPS 2026 · similarity 0.89)
- Reward-Augmented Data Enhances Direct Preference Alignment of LLMs (ICML 2025 · similarity 0.89)
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.
