Research map / Vision-language and multimodal / Multimodal LLMs
HART: High-Resolution Annotation-Free Reasoning Technique through a Closed-loop Framework
Jiacheng Yang, Anqi Chen, Yunkai Dang, Qi Fan, Cong Wang, Wenbin Li, Feng Miao, Yang Gao
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 |
|---|---|---|---|
| apply HART post-training | Qwen2.5-VL-7B · MME-RealWorld-Lite perception | accuracy: 46.4 -> 64.9 | Improved |
| apply HART post-training | Qwen2.5-VL-7B · MME-RealWorld-Lite reasoning | accuracy: 35.9 -> 58.5 | Improved |
| apply HART post-training | Qwen2.5-VL-7B · TreeBench perception | accuracy: 43.6 -> 57.0 | Improved |
| apply HART post-training | InternVL3-8B · MME-RealWorld-Lite perception | accuracy: 51.1 -> 61.8 | Improved |
| apply HART post-training | InternVL3-8B · TreeBench reasoning | accuracy: 34.4 -> 35.9 | Improved |
| add SFT stage after AP-GRPO (k=0.60) (ablation) | HART · MME-RealWorld-Lite perception | accuracy: 53.3 -> 64.9 | Improved |
| run AP-GRPO before SFT rather than SFT first (ablation) | HART (k=0.60) · MME-RealWorld-Lite perception | accuracy: 51.7 -> 64.9 | Improved |
| replace GRPO with AP-GRPO in the closed loop | Qwen2.5-VL-7B · TreeBench grounding (correct %) | grounding accuracy: 51.0% -> 75.4% | Improved |
| replace GRPO with AP-GRPO in the closed loop | Qwen2.5-VL-7B · Visual-CoT grounding (correct %) | grounding accuracy: 66.1% -> 77.7% | Improved |
| answer from cropped ROIs only (closed loop) | Qwen2.5-VL-7B · Visual CoT incorrect grounding given correct answer | proportion: 36.5% -> 21.5% | Improved |
| closed loop on InternVL3-8B | InternVL3-8B · Visual CoT incorrect grounding given correct answer | proportion: 63.8% -> 55.9% | Improved |
| lower k from 0.60 to 0.15 in AP-GRPO + SFT (ablation) | HART · MME-RealWorld-Lite | perception / reasoning accuracy: 64.9 -> 67.0 perception, 58.5 -> 56.9 reasoning | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- DeepEyes-7B — DeepEyes: Incentivizing "Thinking with Images" via Reinforcement Learning (ICLR 2026) · better on 2 of 2 shared results
- Pixel-Reasoner-7B — Pixel Reasoner: Incentivizing Pixel Space Reasoning via Curiosity-Driven Reinforcement Learning (NeurIPS 2025) · better on 1 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
- Latent Implicit Visual Reasoning (CVPR 2026 · similarity 0.90)
- No Labels, No Problem: Training Visual Reasoners with Multimodal Verifiers (ICLR 2026 · similarity 0.89)
- Multi-step Visual Reasoning with Visual Tokens Scaling and Verification (NeurIPS 2025 · similarity 0.88)
- VLM-R³: Region Recognition, Reasoning, and Refinement for Enhanced Multimodal Chain-of-Thought (NeurIPS 2025 · similarity 0.88)
- See It, Say It, Sorted: An Iterative Training-Free Framework for Visually-Grounded Multimodal Reasoning in LVLMs (CVPR 2026 · similarity 0.88)
- Perception Tokens Enhance Visual Reasoning in Multimodal Language Models (CVPR 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.
