Research map / Large language models / Alignment and preferences
The Sign Estimator: Preference Modeling for LLM Alignment under Heterogeneity
Ali Aouad, Aymane El Gadarri, Vivek Farias
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 |
|---|---|---|---|
| replace cross-entropy loss with Sign Estimator binary classification loss | RLHF reward model · SHP | accuracy: 66.50 -> 68.33 | Improved |
| replace cross-entropy loss with Sign Estimator loss | RLHF reward model · HH-Anthropic | accuracy: 66.89 -> 68.58 | Improved |
| replace cross-entropy loss with Sign Estimator loss | RLHF reward model · CoVal | accuracy: 78.09% to 80.72% | Improved |
| replace cross-entropy loss with Sign Estimator loss | RLHF reward model · CoVal (trained on HH-Anthropic) | accuracy: 69.94% to 73.07% | Improved |
| replace cross-entropy loss with Sign Estimator loss | RLHF estimator · semi-synthetic digital twins | angular error to mean utility: 63 -> 41 degrees (about 35%) | Improved |
| replace cross-entropy loss with Sign Estimator loss | RLHF estimator · semi-synthetic digital twins | disagreement rate: roughly 12% to 8% | Improved |
| use EM with K=2 segments instead of pooled RLHF (ablation) | RLHF · semi-synthetic digital twins | angle error, disagreement, accuracy | Improved |
| increase EM segments from K=2 to K=3 or 5 (ablation) | EM reward learning · semi-synthetic digital twins | accuracy and error metrics | Hurt |
| use EM with K=3 segments instead of pooled RLHF (ablation) | RLHF · SHP | accuracy: 66.50 -> 60.83 | Hurt |
| increase heterogeneity scale from 1x to 12x (ablation) | Sign vs RLHF · semi-synthetic digital twins | relative disagreement reduction delta%: 29.22% -> 38.49% | Improved |
| increase heterogeneity scale 1x to 12x (ablation) | RLHF · semi-synthetic digital twins | angle error: 63.24 -> 74.36 | Hurt |
| increase heterogeneity scale 1x to 12x (ablation) | Sign Estimator · semi-synthetic digital twins | angle error: 41.55 -> 49.68 | Hurt |
Closest papers by meaning
- Reward-Augmented Data Enhances Direct Preference Alignment of LLMs (ICML 2025 · similarity 0.93)
- Data with High and Consistent Preference Difference Are Better for Reward Model (AAAI 2025 · similarity 0.93)
- SEAL: Systematic Error Analysis for Value ALignment (AAAI 2025 · similarity 0.92)
- Doubly Robust Alignment for Large Language Models (NeurIPS 2025 · similarity 0.92)
- Online Preference Alignment for Language Models via Count-based Exploration (ICLR 2025 · similarity 0.92)
- Beyond Bradley-Terry Models: A General Preference Model for Language Model Alignment (ICML 2025 · similarity 0.92)
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.
