Research map / Detection and segmentation / CNN features
NAPA: Natively Multimodal Autoregressive Perception Architecture
Phuc Le Khac Hong, Yasser Dahou, Sanath Narayan, Ankit Singh, Wamiq Para, Ngoc Dung Huynh, Sofian Chaybouti, Hakim Hacid
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 |
|---|---|---|---|
| use Muon instead of AdamW optimizer (ablation) | Falcon Perception (detection) · SaCo-Det | Macro F1: 49.0 -> 53.8 | Improved |
| use Muon instead of AdamW optimizer (ablation) | Falcon Perception (detection) · PBench-Det | Macro F1: 56.4 -> 57.7 | Improved |
| add Gram loss feature regularization (ablation) | Falcon Perception (segmentation) · PBench-Seg | Macro F1: 52.7 -> 53.8 | Improved |
| add Gram loss feature regularization (ablation) | Falcon Perception (segmentation) · SaCo-Seg | Macro F1: 51.1 -> 52.6 | Improved |
| mask attention between queries instead of full autoregressive attention (ablation) | Falcon Perception (detection) · PBench-Det | Macro F1: 53.2 -> 54.2 | Improved |
| mask attention between queries instead of full autoregressive attention (ablation) | Falcon Perception (detection) · SaCo-Det | Macro F1: 53.1 -> 53.3 | No clear effect |
| use raster ordering of instances instead of random ordering (ablation) | Falcon Perception (detection) · SaCo-Det | Macro F1: 46.3 -> 56.2 | Improved |
| use raster ordering of instances instead of random ordering (ablation) | Falcon Perception (detection) · PBench-Det | Macro F1: 52.2 -> 59.3 | Improved |
| use multi-teacher distillation initialization instead of random initialization (ablation) | Falcon Perception (300M, detection) · SA-Co | pmF1 and MCC | Improved |
| increase input resolution from 448^2 to 1024^2 (ablation) | Falcon Perception · PBench Dense | micro-F1 (%): 3.9% -> 61.0% | Improved |
| increase input resolution from 448^2 to 1024^2 (ablation) | Falcon Perception · SA-Co | IL_MCC: 0.34 at 448^2; MCC grows 1.6x | Improved |
| sample k=8 predictions and take the best (Pass@8) instead of deterministic decoding (ablation) | Falcon Perception · SA-Co | average cgF1: 34.7 -> 54.3 | Improved |
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.
- SAM 3 — SAM 3: Segment Anything with Concepts (ICLR 2026) · better on 2 of 4 shared results
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
- FalconPerception-HD: High Density Perception via Reinforcement Learning (NeurIPS 2026 · similarity 0.89)
- DenseFusion-1M: Merging Vision Experts for Comprehensive Multimodal Perception (NeurIPS 2024 · similarity 0.87)
- Perception Encoder: The best visual embeddings are not at the output of the network (NeurIPS 2025 · similarity 0.85)
- Perception Tokens Enhance Visual Reasoning in Multimodal Language Models (CVPR 2025 · similarity 0.85)
- Patch-as-Decodable-Token: Towards Unified Multi-Modal Vision Tasks in MLLMs (ICLR 2026 · similarity 0.85)
- Cambrian-1: A Fully Open, Vision-Centric Exploration of Multimodal LLMs (NeurIPS 2024 · similarity 0.85)
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.
