mlsanityData to 7 Oct 2026

Self-improving MLLMs and synthetic data

178 papers2 this weekMedium confidence in this grouping: mostly one topic, with blurry edgesMedium confidence in this grouping: mostly one topic, with blurry edges · About

Keywords: visual, reasoning, multimodal, image, vision-language

Key facts. Self-improving MLLMs and synthetic data: 178 papers on arXiv since 2021 (data through 7 Oct 2026). 40 of its 178 mapped papers (22%) cite “BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models” (2023, arXiv 2301.12597). 10 of its 178 mapped papers (6%) cite “Vision-Language Models Can Self-Improve Reasoning via Reflection” (2024, arXiv 2411.00855). 7 new papers 17 Sep–7 Oct.

What’s new

17 Sep – 7 Oct 2026

7 new papers, 17 Sep – 7 Oct; 2 this week.

  1. This week · 2
  2. BabelFake: A Multilingual Audio-Visual DeepFake Benchmark
  3. Imagine to Act: High-Fidelity Data Synthesis via Image Editing World Model for Scalable GUI Agent Training
  4. Earlier · 5
  5. Trinity: Self-Evolving Vision-Language Models with a Self-Verifier
  6. UniEvo-VL: An On-policy Self-Distillation Training Recipe for Multimodal Model Self-improvement
  7. Spatial-OPSD: Self-Improving Spatial Reasoning via Label-Free Self-Distillation
  8. SpatialSkill: Self-Evolving Skills for Cross-View Spatial Reasoning
  9. ReVision3D: Attribution-Guided Recursive Self-Improvement for 3D Medical Perception

Foundations

Most cited by papers here

  1. BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models2023cited by 40 here
  2. Vision-Language Models Can Self-Improve Reasoning via Reflection2024cited by 10 here
  3. Enhancing Large Vision Language Models with Self-Training on Image Comprehension2024cited by 8 here
  4. Vision-Zero: Scalable VLM Self-Improvement via Strategic Gamified Self-Play2025cited by 7 here
  5. Self-Rewarding Vision-Language Model via Reasoning Decomposition2025cited by 6 here

Lineage

The papers this direction builds on, oldest first.

Each paper on the line cites the one before.

  1. 2022STaR: Bootstrapping Reasoning With Reasoningcited by 7 here
  2. 2023Reinforced Self-Training (ReST) for Language Modelingcited by 6 here
  3. 2023Beyond Human Data: Scaling Self-Training for Problem-Solving with Language Modelscited by 5 here
  4. 2024Self-Play Fine-Tuning Converts Weak Language Models to Strong Language Modelscited by 12 here
  5. 2024Self-Rewarding Language Modelscited by 11 here
  6. 2025Language Self-Play For Data-Free Trainingcited by 3 here
  7. 2025VisPlay: Self-Evolving Vision-Language Models from Imagescited by 4 here
17 more it builds on, not in this chain
  1. 2022Large Language Models Can Self-Improvecited by 10 here
  2. 2022Self-Instruct: Aligning Language Models with Self-Generated Instructionscited by 14 here
  3. 2023BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Modelscited by 40 here
  4. 2023MetaMath: Bootstrap Your Own Mathematical Questions for Large Language Modelscited by 3 here
  5. 2024Synthetic Data (Almost) from Scratch: Generalized Instruction Tuning for Language Modelscited by 3 here
  6. 2024Calibrated Self-Rewarding Vision Language Modelscited by 3 here
  7. 2024Enhancing Visual-Language Modality Alignment in Large Vision Language Models via Self-Improvementcited by 3 here
  8. 2024Enhancing Large Vision Language Models with Self-Training on Image Comprehensioncited by 8 here
  9. 2024ReST-MCTS*: LLM Self-Training via Process Reward Guided Tree Searchcited by 4 here
  10. 2024Magpie: Alignment Data Synthesis from Scratch by Prompting Aligned LLMs with Nothingcited by 3 here
  11. 2024On LLMs-Driven Synthetic Data Generation, Curation, and Evaluation: A Surveycited by 3 here
  12. 2024Math-LLaVA: Bootstrapping Mathematical Reasoning for Multimodal Large Language Modelscited by 3 here
  13. 2024Vision-Language Models Can Self-Improve Reasoning via Reflectioncited by 10 here
  14. 2025Absolute Zero: Reinforced Self-play Reasoning with Zero Datacited by 11 here
  15. 2025R-Zero: Self-Evolving Reasoning LLM from Zero Datacited by 7 here
  16. 2025Self-Rewarding Vision-Language Model via Reasoning Decompositioncited by 6 here
  17. 2025Vision-Zero: Scalable VLM Self-Improvement via Strategic Gamified Self-Playcited by 7 here

The 24 most cited of 112.

Browse

30 of 178, ranked 6–35

  1. 6VisPlay: Self-Evolving Vision-Language Models from Images20254
  2. 7Calibrated Self-Rewarding Vision Language Models20243
  3. 8Enhancing Visual-Language Modality Alignment in Large Vision Language Models via Self-Improvement20243
  4. 9SoTA with Less: MCTS-Guided Sample Selection for Data-Efficient Visual Reasoning Self-Improvement20252
  5. 10Self-Rewarding Large Vision-Language Models for Optimizing Prompts in Text-to-Image Generation20252
  6. 11UniRL: Self-Improving Unified Multimodal Models via Supervised and Reinforcement Learning20252
  7. 12SyncLoop: A Multimodal Dual-Loop Framework for Self-Improving Mathematical Reasoning20252
  8. 13Effective Training Data Synthesis for Improving MLLM Chart Understanding20252
  9. 14Video-STaR: Self-Training Enables Video Instruction Tuning with Any Supervision20242
  10. 15Img-Diff: Contrastive Data Synthesis for Multimodal Large Language Models20242
All 30
  1. 16EvoChart: A Benchmark and a Self-Training Approach Towards Real-World Chart Understanding20242
  2. 17FM2DS: Few-Shot Multimodal Multihop Data Synthesis with Knowledge Distillation for Question Answering20242
  3. 18GPT4Tools: Teaching Large Language Model to Use Tools via Self-instruction20232
  4. 19V-Zero: Self-Improving Multimodal Reasoning with Zero Annotation20261
  5. 20Through the Lens of Contrast: Self-Improving Visual Reasoning in VLMs20261
  6. 21MM-Zero: Self-Evolving Multi-Model Vision Language Models From Zero Data20261
  7. 22Vision-Language Model Dialog Games for Self-Improvement20251
  8. 23Oasis: One Image is All You Need for Multimodal Instruction Data Synthesis20251
  9. 24Self-Evolving Visual Concept Library using Vision-Language Critics20251
  10. 25G1: Bootstrapping Perception and Reasoning Abilities of Vision-Language Model via Reinforcement Learning20251
  11. 26Self-Evolving Vision-Language Models for Image Quality Assessment via Voting and Ranking20251
  12. 27Learning from Synthetic Data for Visual Grounding20241
  13. 28Multimodal Self-Instruct: Synthetic Abstract Image and Visual Reasoning Instruction Using Language Model20241
  14. 29MegaPairs: Massive Data Synthesis For Universal Multimodal Retrieval20241
  15. 30X-LLM: Bootstrapping Advanced Large Language Models by Treating Multi-Modalities as Foreign Languages20231
  16. 31Q: How to Specialize Large Vision-Language Models to Data-Scarce VQA Tasks? A: Self-Train on Unlabeled Images!20231
  17. 32STUPD: A Synthetic Dataset for Spatial and Temporal Relation Reasoning20231
  18. 33Task2Sim : Towards Effective Pre-training and Transfer from Synthetic Data20211
  19. 34Socratic-Geo: Synthetic Data Generation and Geometric Reasoning via Multi-Agent Interaction20260
  20. 354DPC2hat: Towards Dynamic Point Cloud Understanding with Failure-Aware Bootstrapping20260