Research map / Large language models
Alignment and preferences: research map
672 papers on Alignment and preferences in Large language models, from ICML, NeurIPS, ICLR, CVPR, AAAI, ICCV and ECCV, and the journals TPAMI, IJCV and JMLR (2016–2026), grouped into 2 clusters and 7 approaches. The busiest year so far is 2026.
What it is
Aligning a language model means making it helpful, honest and harmless as people judge it. This topic covers learning from preference data, either through a reward model and RL (RLHF) or directly (DPO), and how to build reward models and AI judges people can trust.
Terms you will see on this page:
- Preference data
- pairs of responses with a label saying which one people preferred.
- DPO
- direct preference optimization, which fits preferences without a separate reward model or RL.
- Reward model
- a model that scores responses to stand in for human judgment.
Start reading with the most cited papers here:
- Training language models to follow instructions with human feedback (NeurIPS 2022 · 1,219 citations)
- Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena (NeurIPS 2023 · 810 citations)
Within Large language models, its share grew from 6.6% in 2023–24 to 7.6% in 2025–26 (131 → 522 papers at ICML, NeurIPS, ICLR, CVPR and AAAI, the venues with data for all four years).
Explore Alignment and preferences in the interactive map
Working on something in this topic? Describe your idea in scime atlas to see which approach it falls under, the closest papers by meaning and how crowded the spot has become.
What changed recently
Approaches whose share of the topic moved most between 2023–24 and 2025–26 (ICML, NeurIPS, ICLR, CVPR and AAAI):
- DPO design choices rising: 5% → 13% of the topic (6 → 67 papers)
- Robust and calibrated DPO rising: 12% → 20% of the topic (16 → 102 papers)
- Reward model design and preference models rising: 11% → 13% of the topic (15 → 70 papers)
- Reward modeling for diverse preferences falling: 21% → 12% of the topic (27 → 62 papers)
- LLM-as-judge and feedback data falling: 19% → 15% of the topic (25 → 80 papers)
Papers from 2025–2026 that are already cited most:
- POPEN: Preference-Based Optimization and Ensemble for LVLM-Based Reasoning Segmentation (CVPR 2025 · 10 citations)
- VTON-VLLM: Aligning Virtual Try-On Models with Human Preferences (NeurIPS 2025 · 5 citations)
- Less is More: Improving LLM Alignment via Preference Data Selection (NeurIPS 2025 · 4 citations)
Approaches and key papers
Reward models and human feedback · 418 papers
Keywords: feedback, human, rlhf
Approaches in this cluster:
- Steering and correcting LLM behavior (127 papers)
Align LLMs through representation editing, influence functions, and intrinsic corpus signals. - Reward modeling for diverse preferences (90 papers)
Improve RLHF reward models with language feedback, diverse preferences, and error analysis. - LLM-as-judge and feedback data (112 papers)
Collect scaled AI feedback and study human-versus-LLM judgment in evaluation. - Reward model design and preference models (89 papers)
Combine, actively label, and generalize reward and preference models beyond Bradley-Terry.
Most cited and most cited since 2024:
- Training language models to follow instructions with human feedback (NeurIPS 2022 · 1,219 citations)
- Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena (NeurIPS 2023 · 810 citations)
- Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference (ICML 2024 · 103 citations)
- RLAIF vs. RLHF: Scaling Reinforcement Learning from Human Feedback with AI Feedback (ICML 2024 · 69 citations)
Direct preference optimization · 254 papers
Keywords: preference optimization, dpo, direct preference
Approaches in this cluster:
- Robust and calibrated DPO (119 papers)
Make direct preference optimization robust to noisy preferences and distribution shift through calibration and data selection. - Active and token-level preference learning (58 papers)
Extend DPO with token-level objectives, reward augmentation, and active preference collection. - DPO design choices (77 papers)
Study data, scheduling, dynamic beta, and multi-reference variants of direct preference optimization.
Most cited and most cited since 2024:
- Direct Preference Optimization: Your Language Model is Secretly a Reward Model (NeurIPS 2023 · 418 citations)
- Diffusion Model Alignment Using Direct Preference Optimization (CVPR 2024 · 154 citations)
- SimPO: Simple Preference Optimization with a Reference-Free Reward (NeurIPS 2024 · 39 citations)
- On Softmax Direct Preference Optimization for Recommendation (NeurIPS 2024 · 23 citations)
Related topics in Large language models
- Reasoning and chain of thought (1,574)
- Instruction and fine-tuning (888)
- Retrieval and knowledge (1,084)
- Agents, code and math (1,748)
- AI and society (913)
- Language, safety and interpretability (3,002)
