Skip to content

Paper 020 / DeepMind

Perceiver IO: A General Architecture for Structured Inputs & Outputs

Perceiver IO extends the latent-bottleneck architecture with flexible output queries, allowing structured outputs of many different sizes and semantics.

UPDATED 16 Jul 2026SOURCE-LED REVIEWRESEARCH REVIEW
01

Paper, researchers, and primary source

Major lab research / multimodal_models

Perceiver IO extends the latent-bottleneck architecture with flexible output queries, allowing structured outputs of many different sizes and semantics.

CONTRIBUTION / 01

Contribution 1

Added query-based decoding for arbitrary output structures.

CONTRIBUTION / 02

Contribution 2

Unified classification, dense prediction, language, audio, and multimodal tasks.

CONTRIBUTION / 03

Contribution 3

Retained scalable latent processing for large inputs and outputs.

02

Research context

multimodal_models / 2021

Perceiver IO: A General Architecture for Structured Inputs & Outputs places perceiver io inside the broader multimodal models discussion at DeepMind, with multimodal supplying a second analytical lens. Read together, the source records three advances: Added query-based decoding for arbitrary output structures; Unified classification, dense prediction, language, audio, and multimodal tasks; and Retained scalable latent processing for large inputs and outputs. Keeping those moves together prevents structured output from being detached from its evidence. For an implementation review, the relevant consequence is that typed output queries can let one backbone support classification, generation, and dense predictions without hard-wiring a single output shape.

03

Methods and evidence reading

1 cataloged method notes

METHOD / 01

Method 1

The experimental design in Perceiver IO: A General Architecture for Structured Inputs & Outputs tests added query-based decoding for arbitrary output structures and unified classification, dense prediction, language, audio, and multimodal tasks against the paper's documented baselines, datasets, model variants, or systems workloads.

How to read the evidence

The evidentiary value of Perceiver IO: A General Architecture for Structured Inputs & Outputs comes from the relationship among its reported moves. Two entries define the method-level claim: Added query-based decoding for arbitrary output structures; and Unified classification, dense prediction, language, audio, and multimodal tasks. The cataloged result is: Retained scalable latent processing for large inputs and outputs. On that basis, typed output queries can let one backbone support classification, generation, and dense predictions without hard-wiring a single output shape. The catalog nevertheless records that the source evidence behind perceiver io depends on image resolution, reported datasets, input modalities, domain coverage, prompt protocol, and selected metrics; Perceiver IO: A General Architecture for Structured Inputs & Outputs does not remove those experimental constraints. Reproduction work should separate genuine perceiver io transfer from behavior caused by a changed multimodal setup. Evidence transfer from Perceiver IO: A General Architecture for Structured Inputs & Outputs should be tested by anchoring perceiver io, slicing on multimodal, and keeping counterexamples to Unified classification, dense prediction, language, audio, and multimodal tasks in the evaluation record.

04

Findings in the source record

1 paper-specific findings

  1. The reported evidence in Perceiver IO: A General Architecture for Structured Inputs & Outputs supports retained scalable latent processing for large inputs and outputs.
05

Practical implication for AI builders

DeepMind / 2021

06

Proposed BrokenGPT application

Research blueprint / proposed status

INTEGRATION POINT / 01

Proposed route placement / perceiver io

For a proposed BrokenGPT experiment based on Perceiver IO: A General Architecture for Structured Inputs & Outputs, model output capabilities as typed contracts so BrokenGPT can validate whether an endpoint supports text, labels, timestamps, or dense media outputs. Keep the perceiver io path isolated, versioned, and attributable to this research record.

VALIDATION METRIC / 02

Proposed acceptance test / multimodal

Validate the proposed perceiver io route against the paper's reported outcome: Retained scalable latent processing for large inputs and outputs. The Perceiver IO: A General Architecture for Structured Inputs & Outputs release gate would report modality-specific error, cross-modal consistency, and grounding accuracy plus standalone multimodal slices before accepting the proposed perceiver io adaptation.

TRADEOFF / 03

Proposed decision boundary / structured output

Balance coverage, compute, and provenance before promoting the proposed structured output design. Because product evidence would remain incomplete without testing media rights, accessibility, misuse, memorization, source provenance, and representation bias under the selected multimodal workload, adoption remains conditional on replay under BrokenGPT's selected model, runtime, and policy configuration.

07

Limitations, verification, and source

Boundaries recorded with the paper

Limitations

  • The source evidence behind perceiver io depends on image resolution, reported datasets, input modalities, domain coverage, prompt protocol, and selected metrics; Perceiver IO: A General Architecture for Structured Inputs & Outputs does not remove those experimental constraints.
  • Product evidence would remain incomplete without testing media rights, accessibility, misuse, memorization, source provenance, and representation bias under the selected multimodal workload.

PRIMARY SOURCES

  1. 01
    Perceiver IO: A General Architecture for Structured Inputs & Outputs

    DeepMind — Primary primary arXiv paper / 30 July 2021 / Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, and 12 more

Related research reviews

View all 100 credited research papers

STRAIGHT ANSWERS

Frequently asked questions

01What does Perceiver IO: A General Architecture for Structured Inputs & Outputs study?

Perceiver IO extends the latent-bottleneck architecture with flexible output queries, allowing structured outputs of many different sizes and semantics.

02Which methods does Perceiver IO: A General Architecture for Structured Inputs & Outputs use?

The experimental design in Perceiver IO: A General Architecture for Structured Inputs & Outputs tests added query-based decoding for arbitrary output structures and unified classification, dense prediction, language, audio, and multimodal tasks against the paper's documented baselines, datasets, model variants, or systems workloads.

03What does Perceiver IO: A General Architecture for Structured Inputs & Outputs report?

The reported evidence in Perceiver IO: A General Architecture for Structured Inputs & Outputs supports retained scalable latent processing for large inputs and outputs.

04What is the proposed BrokenGPT application for Perceiver IO: A General Architecture for Structured Inputs & Outputs?

Proposed: model output capabilities as typed contracts so BrokenGPT can validate whether an endpoint supports text, labels, timestamps, or dense media outputs.

MAJOR LAB RESEARCH / PAPER 020

Continue after Perceiver IO: A General Architecture for Structured Inputs & Outputs

After Perceiver IO: A General Architecture for Structured Inputs & Outputs, browse the full index for adjacent multimodal_models research and work from DeepMind.

Open the paper index