Research map / Large language models / Alignment and preferences
Targeting Misalignment: A Conflict-Aware Framework for Reward-Model-based LLM Alignment
Zixuan Liu, Siavash H. Khajavi, Guangkai Jiang, Xinru Liu
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 |
|---|---|---|---|
| refine reward with human feedback on high-conflict pairs (SHF-CAS) instead of vanilla proxy | PPO · PKU-SafeRLHF | gold reward (lower safer): 3.92 -> -2.90 | Improved |
| refine reward with human feedback on high-conflict pairs (SHF-CAS) | PPO · Anthropic HH-RLHF | gold reward: -2.32 -> 3.36 | Improved |
| select high-conflict pairs instead of random pairs (ablation) | SHF-CAS · PKU-SafeRLHF (delta=1.6) | gold reward (lower safer): -0.64 -> -2.90 | Improved |
| select high-conflict pairs instead of random pairs (ablation) | SHF-CAS · HH-RLHF (delta=1.5) | gold reward: 1.03 -> 3.36 | Improved |
| raise PACS threshold delta (fewer, more extreme conflicts) (ablation) | SHF-CAS · PKU-SafeRLHF | gold reward (lower safer): 1.59 -> -2.90 (delta 1.4 -> 1.6) | Improved |
| raise PACS threshold delta (ablation) | SHF-CAS · HH-RLHF | gold reward: 2.33 -> 3.36 (delta 1 -> 1.5) | Improved |
| use SHF-CAS instead of RSO | RSO · PKU-SafeRLHF | gold reward (lower safer): 2.84 -> -2.90 | Improved |
| use SHF-CAS instead of RSO | RSO · HH-RLHF | gold reward: -1.02 -> 3.36 | Improved |
| use GPT-4o oracle instead of reward-model oracle (ablation) | SHF-CAS · PKU-SafeRLHF (delta=1.6) | gold reward (lower safer): -2.90 -> -3.53 | Mixed |
| use GPT-4o oracle instead of reward-model oracle (ablation) | SHF-CAS · HH-RLHF (delta=1.5) | gold reward: 3.36 -> 3.31 | No clear effect |
Closest papers by meaning
- RMB: Comprehensively benchmarking reward models in LLM alignment (ICLR 2025 · similarity 0.91)
- SEAL: Systematic Error Analysis for Value ALignment (AAAI 2025 · similarity 0.91)
- Theoretical Limits of Language Model Alignment (NeurIPS 2026 · similarity 0.91)
- The Unlocking Spell on Base LLMs: Rethinking Alignment via In-Context Learning (ICLR 2024 · similarity 0.91)
- Reward-Augmented Data Enhances Direct Preference Alignment of LLMs (ICML 2025 · similarity 0.91)
- Alignment Dynamics in LLM Fine-Tuning (NeurIPS 2026 · similarity 0.91)
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.
