Research map / Vision-language and multimodal / Multimodal LLMs
SyncLoop: A Multimodal Dual-Loop Framework for Self-Improving Mathematical Reasoning
xiuwei chen, Wentao Hu, Hanhui Li, Yongxin Wang, Jun Zhou, Zisheng Chen, Meng Cao, Yihan Zeng, Kui Zhang, Yu-Jie Yuan, Jianhua Han, Hang Xu and 1 more
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 |
|---|---|---|---|
| run three SyncLoop iterations | Qwen2-VL-7B · Geo-Sub | accuracy: 40.4 -> 50.9 | Improved |
| run three SyncLoop iterations | Qwen2-VL-2B · Geo-Sub | accuracy: 28.0 -> 38.2 | Improved |
| run three SyncLoop iterations | Qwen2-VL-7B · MathVista (ALL, testmini) | accuracy: 60.0 -> 63.2 | Improved |
| add SFT before RL in second iteration (vs RL-only from pi^1) (ablation) | SyncLoop · Geo-Sub-Aux | accuracy: 50.6 -> 40.7 | Hurt |
| add SFT before RL in third iteration (vs RL-only from pi^2) (ablation) | SyncLoop · Geo-Sub-Aux | accuracy: 54.9 -> 42.9 | Hurt |
| warm-start from previous iteration model instead of restarting from initial model (RL only) (ablation) | SyncLoop · Geo-Sub-Aux | accuracy: 44.0 -> 50.6 | Improved |
| add SFT+RL vs RL only at first iteration (ablation) | SyncLoop · Geo-Sub-Aux | accuracy: 44.7 -> 48.7 | Improved |
| add SFT to RL when restarting from initial model (ablation) | SyncLoop · Geo-Sub-Aux | accuracy: 44.0 -> 48.4 | Improved |
| complexify images in addition to text (ablation) | SyncLoop iter 3 · Geo-Sub-Aux | accuracy: 51.27 -> 54.9 | Improved |
| complexify images in addition to text (ablation) | SyncLoop iter 2 · Geo-Sub-Aux | accuracy: 45.82 -> 50.9 | Improved |
| filter by error rate >= 0.3 vs using full dataset (ablation) | SyncLoop iter 2 · Geo-Sub-Aux | accuracy: 48.4/46.6 -> 49.1/50.9 | Mixed |
| raise error-rate threshold to 1.0 (only hardest samples) (ablation) | SyncLoop iter 2 · Geo-Sub-Aux | accuracy: 49.1 -> 46.2 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- R1-Onevision-7B — R1-Onevision: Advancing Generalized Multimodal Reasoning through Cross-Modal Formalization (ICCV 2025) · better on 0 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- MathSE: Improving Multimodal Mathematical Reasoning via Self-Evolving Iterative Reflection and Reward-Guided Fine-Tuning (AAAI 2026 · similarity 0.92)
- AStar: Boosting Multimodal Reasoning with Automated Structured Thinking (AAAI 2026 · similarity 0.91)
- Benchmarking Multimodal Mathematical Reasoning with Explicit Visual Dependency (NeurIPS 2026 · similarity 0.91)
- MV-MATH: Evaluating Multimodal Math Reasoning in Multi-Visual Contexts (CVPR 2025 · similarity 0.90)
- MMFineReason: Closing the Multimodal Reasoning Gap via Open Data-Centric Methods (NeurIPS 2026 · similarity 0.90)
- LENS: Multi-level Evaluation of Multimodal Reasoning with Large Language Models (ICLR 2026 · similarity 0.90)
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.
