Research map / Large language models / Reasoning and chain of thought
AnE: Pushing the Reasoning Frontier of Multimodal LLMs via Anchor Evolution
Zehao Wang, Yihan Zeng, Zidong Gong, Yuanfan Guo, Feng Zhu, Hongzhi Zhang, Wei Zhang, Wangmeng Zuo
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 |
|---|---|---|---|
| add Truth Anchor Expansion (ablation) | Preliminary RL · 8 multimodal reasoning benchmarks (avg) | avg accuracy (%): 57.2 -> 58.6 (+1.4%) | Improved |
| add Scaffold-Stripping Mechanism alone (ablation) | Preliminary RL · 8 multimodal reasoning benchmarks (avg) | avg accuracy (%): 57.2 -> 57.9 (+0.7%) | Improved |
| combine Truth Anchor Expansion and Scaffold-Stripping (ablation) | Preliminary RL · 8 multimodal reasoning benchmarks (avg) | avg accuracy (%): 57.2 -> 59.9 (+2.7%) | Improved |
| use truth anchors instead of random expansion (ablation) | AnE · 8 multimodal reasoning benchmarks (avg) | avg accuracy (%): 54.1 -> 59.9 | Improved |
| use truth anchors instead of synthetic expansion (ablation) | AnE · 8 multimodal reasoning benchmarks (avg) | avg accuracy (%): 57.8 -> 59.9 | Improved |
| use scaffold-stripping instead of teacher distillation (ablation) | AnE · 8 multimodal reasoning benchmarks (avg) | avg accuracy (%): 57.5 -> 59.9 | Improved |
| add RL stripping stage after scaffold SFT (ablation) | AnE-1st · 8 multimodal reasoning benchmarks (avg) | avg accuracy (%): 58.4 -> 59.9 | Improved |
| add frontier data R_frontier+ to RL dataset (ablation) | AnE-1st · 8 multimodal reasoning benchmarks (avg) | avg accuracy (%): 57.9 -> 59.9 | Improved |
| relocate hint to answer prefix instead of keeping it in input (ablation) | AnE-1st · 8 multimodal reasoning benchmarks (avg) | avg accuracy (%): 57.3 -> 59.9 | Improved |
| use full FineVision instead of 50% (ablation) | AnE-1st · 8 multimodal reasoning benchmarks (avg) | avg accuracy (%): 59.8 -> 59.9 | No clear effect |
| include Chart&Table samples in external database (ablation) | AnE-1st · 8 multimodal reasoning benchmarks (avg) | avg accuracy (%): 59.4 -> 59.9 | Improved |
| add further evolution iterations (2nd and 3rd round) | AnE · 8 multimodal reasoning benchmarks (avg) | avg accuracy (%): 59.9 -> 60.5 -> 61.9 | Improved |
Closest papers by meaning
- The First Few Tokens Are All You Need: An Efficient and Effective Unsupervised Prefix Fine-Tuning Method for Reasoning Models (NeurIPS 2025 · similarity 0.89)
- Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes (ICLR 2026 · similarity 0.88)
- Stable and Efficient Single-Rollout RL for Multimodal Reasoning (CVPR 2026 · similarity 0.88)
- MM-HELIX: Boosting Multimodal Long-Chain Reflective Reasoning with Holistic Platform and Adaptive Hybrid Policy Optimization (ICLR 2026 · similarity 0.88)
- LLMs Can Evolve Continually on Modality for $\mathbb{X}$-Modal Reasoning (NeurIPS 2024 · similarity 0.88)
- SRPO: Enhancing Multimodal LLM Reasoning via Reflection-Aware Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
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.
