Research map / Large language models / Alignment and preferences
Robust and calibrated DPO
Make direct preference optimization robust to noisy preferences and distribution shift through calibration and data selection.
118 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2023–2026) · an approach in Alignment and preferences, keyword cluster “preference optimization · dpo · direct preference”
State History Benchmarks Tried Checklist Preprints Who Export
Where it stands
Still moving The best reported result is still improving on 2 of 2 tracked benchmarks.
16 papers in 2023–24 and 102 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 12.2% to 19.5%.
Open questions the data points to
- GSM8K (score): best went from 34.87 (up to 2024) to 90.07 in 2025 — InfiFPO: Implicit Model Fusion via Preference Optimization in Large Language Models (NeurIPS 2025 · 0 citations)
- IFEval (score): best went from 21.72 (up to 2024) to 82.25 in 2025 — InfiFPO: Implicit Model Fusion via Preference Optimization in Large Language Models (NeurIPS 2025 · 0 citations)
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2023 · CoP: Factual Inconsistency Detection by Controlling the Preference (AAAI 2023 · 4 citations)
- 2024 · Cal-DPO: Calibrated Direct Preference Optimization for Language Model Alignment (NeurIPS 2024 · 7 citations)
- 2025 · POPEN: Preference-Based Optimization and Ensemble for LVLM-Based Reasoning Segmentation (CVPR 2025 · 10 citations)
- 2026 · LILO: Bayesian Optimization with Natural Language Feedback (ICML 2026 · 2 citations)
Earliest papers here
- CoP: Factual Inconsistency Detection by Controlling the Preference (AAAI 2023 · 4 citations)
- Preference-grounded Token-level Guidance for Language Model Fine-tuning (NeurIPS 2023 · 1 citation)
- Cal-DPO: Calibrated Direct Preference Optimization for Language Model Alignment (NeurIPS 2024 · 7 citations)
Benchmark records
The best number each year that beat every earlier one, among papers of this approach reporting their own method. Numbers are extracted by AI from each paper's text; check the paper before citing.
GSM8K · score (higher is better)
4 papers, 2024–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 34.87 | Cal-DPO: Calibrated Direct Preference Optimization for Language Model Alignment (NeurIPS 2024 · 7 citations) |
| 2025 | 90.07 | InfiFPO: Implicit Model Fusion via Preference Optimization in Large Language Models (NeurIPS 2025 · 0 citations) |
IFEval · score (higher is better)
4 papers, 2024–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 21.72 | Cal-DPO: Calibrated Direct Preference Optimization for Language Model Alignment (NeurIPS 2024 · 7 citations) |
| 2025 | 82.25 | InfiFPO: Implicit Model Fusion via Preference Optimization in Large Language Models (NeurIPS 2025 · 0 citations) |
What has been tried
53 papers here report 635 measured changes (a component added, removed or swapped, compared with a baseline): helped 64%, no clear effect 6%, hurt 25%, mixed 5%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Add preference-based ensemble | MUSE val · gIoU | 44.1 -> 45.42 | Helped | POPEN: Preference-Based Optimization and Ensemble for LVLM-Based Reasoning Segmentation (CVPR 2025 · 10 citations) |
| Increase β too much | Various tasks · model performance | Hurt | Cal-DPO: Calibrated Direct Preference Optimization for Language Model Alignment (NeurIPS 2024 · 7 citations) | |
| Use a 3k-sample BeeS subset rather than the full dataset | AlpacaEval 2.0 · win rate | No clear effect | Less is More: Improving LLM Alignment via Preference Data Selection (NeurIPS 2025 · 4 citations) | |
| Use fine-grained per-aspect preference judgments rather than overall judgments | Preference evaluation suite · average | 60.3 -> 61.0 | Helped | Unpacking DPO and PPO: Disentangling Best Practices for Learning from Preference Feedback (NeurIPS 2024 · 4 citations) |
| increase contrastive responses per prompt from K=2 to K=4 | UltraFeedback · evaluation performance | Helped | Noise Contrastive Alignment of Language Models with Explicit Rewards (NeurIPS 2024 · 4 citations) | |
| Perturb the reference model used by NPO | TOFU Forget05 · FQ | 0.79 -> 0.27 | Hurt | Simplicity Prevails: Rethinking Negative Preference Optimization for LLM Unlearning (NeurIPS 2025 · 3 citations) |
| increase preference data from 0 to 50k | zero-shot TTS · WER | Helped | SpeechAlign: Aligning Speech Generation to Human Preferences (NeurIPS 2024 · 3 citations) | |
| Vary λ around the selected value | Robotic myopic noise at γ=0.7 and non-perturbed dialogue/summarization · normalized return and win rate | Helped | Robust Reinforcement Learning from Corrupted Human Feedback (NeurIPS 2024 · 3 citations) | |
| Use a fixed preference-vector scaling factor instead of a learnable per-weight-matrix factor | HH · Pareto-front performance | Hurt | Panacea: Pareto Alignment via Preference Adaptation for LLMs (NeurIPS 2024 · 3 citations) | |
| Use fewer pairwise truthfulness training examples at fixed DPO executions | TruthfulQA MC1 · accuracy | Hurt | GRATH: Gradual Self-Truthifying for Large Language Models (ICML 2024 · 2 citations) | |
| remove rule-based pre-screening of generated response permutations | AlpacaEval 2.0 · LC weighted win rate | 13.04 -> 13.01 | No clear effect | Self-Evolutionary Large Language Models Through Uncertainty-Enhanced Preference Optimization (AAAI 2025 · 1 citation) |
| Increase TPO iterations and samples from D2-N5 to D5-N20 | AlpacaEval 2 · WR(%) | 40.5 -> 55.7 | Helped | Test-Time Preference Optimization: On-the-Fly Alignment via Iterative Textual Feedback (ICML 2025 · 1 citation) |
What reviewers will expect
Datasets and baselines named in the 53 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- UltraFeedback · 18 of 53 papers
- UltraChat-200k · 7 of 53 papers
- Anthropic HH · 6 of 53 papers
- HH-RLHF · 5 of 53 papers
- Alpaca · 4 of 53 papers
- UltraFeedback Binarized · 4 of 53 papers
- MT-Bench · 4 of 53 papers
- HelpSteer2 · 3 of 53 papers
Baselines
- DPO · 23 of 53 papers
- SimPO · 13 of 53 papers
- SFT · 12 of 53 papers
- IPO · 12 of 53 papers
- R-DPO · 9 of 53 papers
- KTO · 9 of 53 papers
- CPO · 7 of 53 papers
- ORPO · 6 of 53 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 118 papers; from OpenAlex and the papers' affiliations).
- University of Science and Technology of China · 8 papers
- Microsoft · 5 papers
- Alibaba Group · 4 papers
- Chinese Academy of Sciences · 3 papers
- New York University · 3 papers
- University of Washington · 3 papers
- The Chinese University of Hong Kong · 3 papers
- Zhejiang University · 3 papers
- Nanyang Technological University · 3 papers
- National University of Singapore · 3 papers
Use it
Download BibTeX of 22 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Active and token-level preference learning (58 papers)
- DPO design choices (73 papers)
Elsewhere in the topic
- Steering and correcting LLM behavior (127 papers)
- LLM-as-judge and feedback data (111 papers)
- Reward modeling for diverse preferences (90 papers)
- Reward model design and preference models (88 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).