Research map / Vision-language and multimodal / Human and embodied modeling
Perceiver IO: A General Architecture for Structured Inputs & Outputs
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Replace BERT with Perceiver IO on UTF-8 bytes at a matching FLOPs budget | GLUE · GLUE | Avg.: 71.5 -> 81.0 | Helped |
| Increase model depth and FLOPs from Perceiver IO to Perceiver IO++ | GLUE · GLUE | Avg.: 81.0 -> 81.8 | Helped |
| Use multitask queries instead of independently trained single-task queries | Perceiver IO byte model · GLUE (8-task multitask) | Avg.: 81.0 -> 81.8 | Helped |
| Replace a shared input token with task-specific input tokens (ablation) | Perceiver IO byte model · GLUE (8-task multitask) | Avg.: 81.5 -> 81.8 | Helped |
| Use task-specific output queries instead of single-task training (ablation) | Perceiver IO byte model · GLUE (8-task multitask) | Avg.: 81.0 -> 81.8 | Helped |
| Use the query-based decoder instead of the original Perceiver decoder | Perceiver · Sintel.clean | EPE | Helped |
| Increase the Kinetics autoencoding compression ratio from 88x to 352x (ablation) | Perceiver IO · Kinetics-700-2020 autoencoding | Audio PSNR: 26.97 -> 14.15 | Hurt |
| Increase the Kinetics autoencoding compression ratio from 88x to 352x (ablation) | Perceiver IO · Kinetics-700-2020 autoencoding | Video PSNR: 24.37 -> 23.21 | Hurt |
| Increase the Kinetics autoencoding compression ratio from 88x to 352x (ablation) | Perceiver IO · Kinetics-700-2020 autoencoding | Top-1 accuracy: 10.2% -> 11.5% | Mixed |
| Increase the Kinetics class-loss weight (ablation) | Perceiver IO multimodal autoencoder · Kinetics-700-2020 autoencoding | Top-1 accuracy: 45% | Helped |
| Replace average-plus-project readout with the query-based attention decoder (ablation) | Perceiver · ImageNet | Top-1 accuracy: 78.6 -> 79.0 | Helped |
| Replace average-plus-project readout with the query-based attention decoder (ablation) | Perceiver · ImageNet (learned positional features) | Top-1 accuracy: 67.6 -> 72.7 | Helped |
Closest papers by meaning
- Perception Encoder: The best visual embeddings are not at the output of the network (NeurIPS 2025 · similarity 0.86)
- PerceptionGPT: Effectively Fusing Visual Perception into LLM (CVPR 2024 · similarity 0.86)
- Elevating Visual Perception in Multimodal LLMs with Visual Embedding Distillation (NeurIPS 2025 · similarity 0.85)
- VLsI: Verbalized Layers-to-Interactions from Large to Small Vision Language Models (CVPR 2025 · similarity 0.85)
- PerceptionLM: Open-Access Data and Models for Detailed Visual Understanding (NeurIPS 2025 · similarity 0.85)
- VISion On Request: Enhanced VLLM efficiency with sparse, dynamically selected, vision-language interactions (CVPR 2026 · 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.