Research map / Vision-language and multimodal / Multimodal LLMs
SPOT-E: Test-Time Entropy Shaping with Visual Spotlights for Frozen VLMs
Bo Yin, Xiaobin Hu, Chengming Xu, Ruolin Shen, Mo Yang, Jiangning Zhang, Peng-Tao Jiang, Cheng Tan, Shuicheng Yan
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 SPOT-E test-time spotlight | LLaVA-NeXT-7B · TextVQA | accuracy: 78.5 -> 84.7 | Improved |
| add SPOT-E test-time spotlight | Qwen2.5-VL-7B · MathVista | accuracy: 67.8 -> 70.8 | Improved |
| add SPOT-E to strong large models | Qwen2.5-VL-32B · TextVQA | accuracy: 88.3 -> 89.5 | Improved |
| add SPOT-E | Qwen2.5-VL-7B · POPE | accuracy: 86.4 -> 87.4 | Improved |
| add anchor-preservation reward to clarity reward (ablation) | SPOT-E · TextVQA | accuracy: 85.8 -> 86.9 | Improved |
| add clarity reward to preservation reward (ablation) | SPOT-E · MathVista | accuracy: 68.6 -> 70.8 | Improved |
| add dynamic scaling of clarity reward (ablation) | SPOT-E · MathVista | accuracy: 70.1 -> 70.8 | Improved |
| use max fusion over crops instead of global view only (ablation) | SPOT-E · MathVista | accuracy: 66.9 -> 70.8 | Improved |
| max fusion instead of mean fusion (ablation) | SPOT-E · TextVQA | accuracy: 85.7 -> 86.9 | Improved |
| add background degradation (ablation) | SPOT-E · MathVista | accuracy: 67.8 -> 70.8 | Improved |
| increase test-time update steps from 0 to 8, then 16 (ablation) | SPOT-E · TextVQA | accuracy: 84.9 -> 86.9 (8 steps); 87.0 (16 steps) | Improved |
| increase anchor preservation weight to 1.0 from 0.5 (ablation) | SPOT-E · TextVQA | accuracy: 86.9 -> 86.1 | Hurt |
Closest papers by meaning
- Diversity Matters: Revisiting Test-Time Compute in Vision-Language Models (ICML 2026 · similarity 0.90)
- Seeing but Not Believing: Probing the Disconnect Between Visual Attention and Answer Correctness in VLMs (ICLR 2026 · similarity 0.89)
- SPARC: Separating Perception And Reasoning Circuits for Test-time Scaling of VLMs (ICML 2026 · similarity 0.88)
- Evidence-RL: Towards Evidence-intensive Visual Reasoning (NeurIPS 2026 · similarity 0.88)
- When to Think and When to Look: Uncertainty-Guided Lookback (CVPR 2026 · similarity 0.87)
- SIEVES: Selective Prediction Generalizes through Visual Evidence Scoring (NeurIPS 2026 · similarity 0.87)
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.
