Research map / Vision-language and multimodal / Human and embodied modeling
RADIO1D: Elastic Representations for Condensed Vision Modeling
Greg Heinrich, Mike Ranzinger, Collin McCarthy, Natan Bagrov, Eugene Khvedchenya, Bryan Catanzaro, Jan Kautz, Andrew Tao, Pavlo Molchanov
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 |
|---|---|---|---|
| fine-tune the vision encoder inside a VLM | C-RADIOv4-H · MS-COCO CKA | CKAoff (spatial coherence): 0.281 to 0.035 | Hurt |
| use first-token versus last-token of the 1D sequence (ablation) | RADIO1D · ADE20K linear probe | mIoU: 8.05 -> 40.23 | Improved |
| use first 16 tokens rather than last 16 (ablation) | RADIO1D · ADE20K linear probe | mIoU: 20.84 -> 54.11 | Improved |
| add integrated learnable patch merging in the encoder (ablation) | RADIO1D · VLM short run, DocVQA/InfoVQA/LongVideoBench | accuracy: InfoVQA 69.39 -> 73.43; LongVideoBench 53.00 -> 56.54 | Improved |
| apply 2x2 pixel unshuffle on the 1D token sequence output (ablation) | RADIO1D · RADIO1D training | training convergence | Hurt |
| shuffle 1D token order before 1D unshuffling (ablation) | RADIO1D · RADIO1D training | training loss | No clear effect |
| use triangular instead of uniform sampling of token length (ablation) | RADIO1D (SO400M, short schedule) · ADE20K linear probe | mIoU | Improved |
| place downscaling after block 24 (kappa=24) (ablation) | RADIO1D (SO400M, short schedule) · ADE20K linear probe | mIoU | Improved |
| use expansion factor 4 instead of 2 in downscaling (ablation) | RADIO1D · parameter count and throughput | parameters / throughput | Hurt |
| learn a 1D elastic sequence of 128 tokens instead of pruning C-RADIOv4 post-hoc with ToMe | VLM with RADIO1D · 10 VLM benchmarks average, 128 tokens | average accuracy: 70.31 -> 71.64 | Improved |
| apply CDPruner on top of the full 256-token RADIO1D instead of native elastic length (ablation) | RADIO1D · 10 VLM benchmarks average, 128 tokens | average accuracy: 71.64 -> 68.56 | Hurt |
| reduce visual tokens to 1 or fewer than 128 in RADIO1D (ablation) | RADIO1D VLM · OCR benchmarks (DocVQA) | accuracy: DocVQA 93.2 -> 51.1 (256 -> 1 token) | Hurt |
Closest papers by meaning
- Reading Images Like Texts: Sequential Image Understanding in Vision-Language Models (ICLR 2026 · similarity 0.88)
- From Pixels to Words -- Towards Native Vision-Language Primitives at Scale (ICLR 2026 · similarity 0.88)
- MMLongBench: Benchmarking Long-Context Vision-Language Models Effectively and Thoroughly (NeurIPS 2025 · similarity 0.88)
- Benchmarking Large Vision-Language Models on Fine-Grained Image Tasks: A Comprehensive Evaluation (ICLR 2026 · similarity 0.88)
- Patch-as-Decodable-Token: Towards Unified Multi-Modal Vision Tasks in MLLMs (ICLR 2026 · similarity 0.87)
- Same or Not? Enhancing Visual Perception in Vision-Language Models (CVPR 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.
