Research map / Large language models / Agents, code and math
Reasoning Boosts Opinion Alignment in LLMs
Frédéric Berdoz, Yann Billeter, Yann Vonlanthen, Roger Wattenhofer
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 |
|---|---|---|---|
| add SFT warm start before GRPO (ablation) | GRPO, Llama 3.1 8B · WoM | macro-F1: 37.29 -> 52.53 | Improved |
| add SFT warm start before GRPO (ablation) | GRPO, Llama 3.1 8B · smartvote | macro-F1: 55.14 -> 66.88 | Improved |
| add SFT warm start before GRPO (ablation) | GRPO, Qwen3 8B · ANES | macro-F1: 31.47 -> 38.44 | Improved |
| add SFT warm start before GRPO (ablation) | GRPO, Magistral 24B · smartvote | macro-F1: 60.56 -> 70.73 | Improved |
| add GRPO after SFT (ablation) | SFT, Llama 3.1 8B · ANES | macro-F1: 42.77 -> 40.66 | Hurt |
| add GRPO after SFT (ablation) | SFT, Llama 3.1 8B · WoM | macro-F1: 48.95 -> 52.53 | Improved |
| add GRPO after SFT (ablation) | SFT, Magistral 24B · WoM | macro-F1: 51.86 -> 53.21 | No clear effect |
| use reasoning-pretrained backbone (Qwen3 8B vs Llama 3.1 8B) | SFT+GRPO · WoM | macro-F1: 52.53 -> 49.38 | Hurt |
| scale backbone to Magistral 24B | SFT+GRPO · smartvote | macro-F1: 66.88 -> 70.73 | Improved |
| replace ICL with SFT+GRPO | Llama 3.1 8B · WoM | macro-F1: 28.17 -> 52.53 | Improved |
| remove Neutral instances from ANES evaluation (ablation) | SFT+GRPO, Llama 3.1 8B · ANES | accuracy: 59.21 -> 71.35 | Improved |
| use aggressive instead of conservative ANES recoding (ablation) | SFT+GRPO, Llama 3.1 8B · ANES | macro-F1: 40.67 -> 45.33 | Improved |
Closest papers by meaning
- Examining Alignment of Large Language Models through Representative Heuristics: the case of political stereotypes (ICLR 2025 · similarity 0.89)
- LLM Active Alignment: A Nash Equilibrium Perspective (NeurIPS 2026 · similarity 0.88)
- Group Preference Optimization: Few-Shot Alignment of Large Language Models (ICLR 2024 · similarity 0.87)
- Diversity-Aware Policy Optimization for Large Language Model Reasoning (NeurIPS 2025 · similarity 0.86)
- Reward-Augmented Data Enhances Direct Preference Alignment of LLMs (ICML 2025 · similarity 0.86)
- Aligning Large Language Models for Faithful Integrity Against Opposing Argument (AAAI 2025 · similarity 0.86)
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.
